Liquidity Stress Testing in Asset Management -- Part 1. Modeling the Liability Liquidity Risk
Thierry Roncalli, Fatma Karray-Meziou, François Pan, Margaux Regnault
LLiquidity Stress Testing in Asset ManagementPart 1. Modeling the Liability Liquidity Risk ∗ Thierry RoncalliQuantitative ResearchAmundi Asset Management, Paris [email protected]
Fatma Karray-MeziouRisk ManagementAmundi Asset Management, Paris [email protected]
Fran¸cois PanRisk ManagementAmundi Asset Management, Paris [email protected]
Margaux RegnaultQuantitative ResearchAmundi Asset Management, Paris [email protected]
December 2020
Abstract
This article is part of a comprehensive research project on liquidity risk in assetmanagement, which can be divided into three dimensions. The first dimension coversliability liquidity risk (or funding liquidity) modeling, the second dimension focuses onasset liquidity risk (or market liquidity) modeling, and the third dimension considersasset-liability liquidity risk management (or asset-liability matching). The purpose ofthis research is to propose a methodological and practical framework in order to performliquidity stress testing programs, which comply with regulatory guidelines (ESMA,2019) and are useful for fund managers. The review of the academic literature andprofessional research studies shows that there is a lack of standardized and analyticalmodels. The aim of this research project is then to fill the gap with the goal to developmathematical and statistical approaches, and provide appropriate answers.In this first part that focuses on liability liquidity risk modeling, we propose severalstatistical models for estimating redemption shocks. The historical approach mustbe complemented by an analytical approach based on zero-inflated models if we wantto understand the true parameters that influence the redemption shocks. Moreover,we must also distinguish aggregate population models and individual-based modelsif we want to develop behavioral approaches. Once these different statistical modelsare calibrated, the second big issue is the risk measure to assess normal and stressedredemption shocks. Finally, the last issue is to develop a factor model that can translatestress scenarios on market risk factors into stress scenarios on fund liabilities.
Keywords: liquidity, stress testing, liability, redemption rate, redemption frequency, re-demption severity, zero-inflated beta model, copula.
JEL classification:
C02, G32. ∗ We are grateful to Pascal Blanqu´e, Bernard de Wit, Vincent Mortier and Eric Vandamme for theirhelpful comments, and Th´eo Roncalli for his research assistance on zero-inflated models. This research hasalso benefited from the support of Amundi Asset Management, which has provided the data. However, theopinions expressed in this article are those of the authors and are not meant to represent the opinions orofficial positions of Amundi Asset Management. a r X i v : . [ q -f i n . R M ] J a n iquidity Stress Testing in Asset Management Liquidity stress testing in asset management is a complex topic because it is related to threedimensions — liquidity risk, asset management and stress testing, whose linkages have beenlittle studied and are hard to capture. First, liquidity is certainly the risk that is the mostdifficult to model with the systemic risk. If we consider market, credit, operational andcounterparty risks, there is a huge amount of academic literature on these topics in terms ofmodels, statistical inference and analysis. In terms of liquidity risk, the number of practicalstudies and applied approaches is limited. Even though a great deal of research has beencompleted on this subject, much of it is overly focused on descriptive analyses of liquidity, orits impact on systemic risk, or policy rules for financial regulation. Moreover, this researchgenerally focuses on the banking sector (Grillet-Aubert, 2018). For instance, the liquiditycoverage ratio (LCR) and the net stable funding ratio (NSFR) of the Basel III regulatoryframework are of no help when measuring the liquidity risk in asset management. In fact,the concept of liquidity risk in asset management is not well defined. More generally, it isa recent subject and we must admit that the tools and models used in asset managementare very much lagging those developed in the banking sector. This is why the culture ofasset-liability management (ALM) is poor in investment management, both on the side ofasset managers and asset owners. Therefore, if we add the third dimension, stress testing,we obtain an unknown and obscure topic, because the combination of liquidity risk andstress testing applied to asset management is a new and difficult task.This is not the first time that the regulatory environment has sped up the developmentof a risk management framework. Previous occurrences include the case of market risk withthe Amendment of the first Basel Accord (BCBS, 1996), credit risk with the second con-sultative paper on Basel II (BCBS, 2001), credit valuation adjustment with the publicationof the Basel III Accord (BCBS, 2010), interest rate risk in the banking book with the IR-RBB guidelines (BCBS, 2016), etc. However, the measurement of these risks had alreadybenefited from the existence of analytical models developed by academics and professionals.One exception was operational risk, since banks started from a blank page when asked tomeasure it (BCBS, 1999). Asset managers now face a similar situation at this moment.Between 2015 and 2018, the US Securities and Exchange Commission established severalrules governing liquidity management (SEC, 2015, 2016, 2018a,b). In particular, Rule 22e-4requires investment funds to classify their positions in one of four liquidity buckets (highlyliquid investments, moderately liquid investments, less liquid investments and illiquid in-vestments), establish a liquid investment minimum, and develop policies and procedures onredemptions in kind. From September 2020, European asset managers must also complywith new guidelines on liquidity stress testing (LST) published by the European Securi-ties and Markets Authority (ESMA, 2019). These different regulations are rooted in theagenda proposed by the Financial Stability Board to monitor and manage systemic risk ofnon-bank non-insurer systemically important financial institutions (FSB, 2010). Even if theoriginal works of the FSB were biased, the idea that the asset management industry cancontribute to systemic risk has gained ground and is now widely accepted. Indeed, FSB(2015) confused systemic risk and systematic market risk (Roncalli and Weisang, 2015a).However, Roncalli and Weisang (2015b) showed that “ the liquidation channel is the maincomponent of systemic risk to which the asset management industry contributes ”. In thiscontext, liquidity is the major risk posed by the asset management industry that regulatorsmust control. But liquidity risk is not only a concern for regulators. It must also be apriority for asset managers. The crisis of money market funds in the fourth quarter of 2008demonstrated the fragility of some fund managers (Schmidt et al. , 2016). Market liquiditydeteriorated in March and April 2020, triggering a liquidity shock on some investment funds2iquidity Stress Testing in Asset Managementand strategies. However, aside from the 2008 Global Financial Crisis and 2020 coronaviruspandemic, which have put all asset managers under pressure, the last ten years have demon-strated that liquidity is also an individual risk for fund managers. It was especially trueduring episodes of flash crash , where fund managers reacted differently. In a similar way,idiosyncratic liquidity events may affect asset managers at the individual level (Thompson,2019). Following some high-profile fund suspensions in mid-2019, asset managers receivedrequests from asset owners to describe their liquidity policies and conduct a liquidity reviewof their portfolios. Therefore, we notice that liquidity is increasingly becoming a priority forasset managers for three main reasons, because it is a reputational risk, they are challengedby asset owners and it can be a vulnerability factor for financial performance.However, even though liquidity stress testing in asset management has become one ofthe hot topics in finance, it has attracted few academics and professionals, implying that theresearch on this subject is not as dynamic as one might expect. In fact, it is at the same stageas operational risk was in the early 2000s, when there was no academic research on this topic.And it is also at the stage of ALM banking risk, where the most significant contributionshave come from professionals. Since liquidity stress testing in asset management is an asset-liability management exercise, modeling progress mainly comes from professionals, becausethe subject is so specific, requires business expertise and must be underpinned by industry-level data. This is obviously an enormous hurdle for academics, and this explains the lack ofmodeling and scientific approach that asset managers encounter when they want to developa liquidity stress testing framework. Therefore, the objective of this research is twofold.First, the idea is to provide a mathematical and statistical formalization to professionals inorder to go beyond expert qualitative judgement. Second, the aim is to assist academicsin understanding this topic. This is important, because academic research generally booststhe development of analytical models, which are essential for implementing liquidity stresstesting programs in asset management.Liquidity stress testing in asset management involves so many dimensions that we havedecided to split this research into three parts:1. liability liquidity risk modeling;2. asset liquidity risk modeling;3. asset-liability liquidity risk management.Indeed, managing liquidity risk consists of three steps. First, we have to model the liabilityliquidity of the investment fund, especially the redemption shocks. By construction, thisstep must incorporate client behavior. Second, we have to develop a liquidity model forassets. For that, we must specify a transaction cost model beyond the traditional bid-askspread measure. In particular, the model must incorporate two dimensions: price impactand trading limits. These first two steps make the distinction between funding liquidityand market liquidity. As noticed by Brunnermeier and Pedersen (2009), these two types ofliquidity may be correlated. However, we suppose that they are independent at this levelof analysis. While the first step gives the liquidity amount of the investment fund that canbe required by the investors, the second step gives the liquidity amount of the investmentfund that can be available in the market. Therefore, the third step corresponds to the asset-liability management in terms of liquidity, that is the matching process between requiredliquidity and available liquidity. This implies defining the part of the redemption shockthat can be managed by asset liquidation and the associated liquidity costs. It also implies For instance, during the US stock market flash crash (May 6, 2010), the US Treasury flash crash (October15, 2014), the US ETF flash crash (August 24, 2015), etc.
Asset-liability management(liquidity matching)Asset (or market) liquidityLiability (or funding) liquidity
The three-stage process has many advantages in terms of modeling. First, it splits acomplex question into three independent and more manageable problems. This is partic-ularly the case of liability and asset modeling. Second, managing liquidity risk becomesa sequential process, where the starting point is clearly identified. As shown in Figure 1,we should begin with the liability risk. Indeed, if we observe no inflows or outflows, theprocess stops here. As such, the first stage determines the amount to sell in the market andit is measured with respect to the investor behavior. The liquidity risk has its roots in theseverity of the redemption shock. Market liquidity is part of the second phase. Dependingon the redemption size and the liquidity of the market, the fund manager will decide thebest solution to adopt. And the sequential process will conclude with the action of the fundmanager . Finally, the third advantage concerns the feasibility of stress testing programs.In this approach, stress testing concerns the two independent dimensions. We can stress theliquidity on the liability side, or we can stress the liquidity on the asset side, or both, butthe rule is simple.In the sequential approach, the liability of the investment fund is the central node of theliquidity risk, and the vertex of the liquidity network. However, it is not so simple, becausethe three nodes can be interconnected (Figure 2). If market liquidity deteriorates sharply,investors may be incited to redeem in order to avoid a liquidity trap. In this case, fundingliquidity is impacted by market liquidity, reinforcing the feedback loop between fundingand market liquidity, which is described by Brunnermeier and Pedersen (2009). But this isnot the only loop. For instance, the choice of an ALM decision may also influence fundingliquidity. If one asset manager decides to suspend redemptions, it may be a signal forthe investors of the other asset managers if they continue to meet redemptions. Again,we may observe a feedback loop between funding liquidity and asset-liability management. Of course, the fund manager’s action is not uniquely determined, because it depends on several pa-rameters. This means that two fund managers can take two different decisions even if they face the samesituation in terms of redemption and market liquidity.
Asset-liabilitymanagementMarketliquidityrisk Fundingliquidityrisk
Liquidity is a long-standing issue and also an elusive concept. This is particularly truein asset management, where liquidity covers several interpretations. For example, someasset classes are considered as highly liquid whereas other asset classes are illiquid. In thefirst category, we generally find government bonds and large cap stocks. The last categoryincludes real estate and private equities. However, categorizing liquidity of a security is noteasy and there is no consensus. Let us consider for example Rule 22e-4(b) that is applied inthe US. The proposed rule was based on the ability to convert the security to cash withina given period and distinguished six buckets: (a) convertible to cash within 1 business day,(b) convertible to cash within 2-3 business days, (c) convertible to cash within 4-7 calendardays, (d) convertible to cash within 8-15 calendar days, (e) convertible to cash within 16-30calendar days (f) convertible to cash in more than one month. Finally, the adopted rule isthe following:1. highly liquid investments (convertible to cash within three business days);2. moderately liquid investments (convertible to cash within four to seven calendar days);3. less liquid investments (expected to be sold in more than seven calendar days);4. illiquid investments (cannot be sold in seven calendar days or less without significantprice impact).Classifying a security into a bucket may be different from one fund manager to another.Moreover, the previous categories depend on the market conditions. Nevertheless, evenif the current market liquidity is abundant, securities that can be categorized in the firstbucket must also face episodes of liquidity shortage (Blanqu´e and Mortier, 2019a). A typicalexample concerns government bonds facing idiosyncratic risks. Blanqu´e and Mortier (2019a)5iquidity Stress Testing in Asset Managementgave the case of Italian bonds in 2018 during the discussion on the budget deficit. However,most of the time, when we consider the liquidity of an asset class, we assume that it is static.Certainly, this way of thinking reflects the practice of portfolio management. Indeed, it iscommon to include a constant illiquidity premium when estimating the expected returnsof illiquid assets. But investors should stick to their investments without rebalancing andtrading if they want to capture this illiquidity premium. The split between liquid and illiquidinvestments does not help, because it is related to the absolute level of asset illiquidity, andnot liquidity dynamics. However, the issue is more complex:“ [...] there is also broad belief among users of financial liquidity – traders, in-vestors and central bankers – that the principal challenge is not the average levelof financial liquidity... but its variability and uncertainty ” (Persaud, 2003).This observation is important because it is related to the liquidity question from a regulatorypoint of view. The liquidity risk of private equities or real assets is not a big concern forregulators, because one knows that these asset classes are illiquid. Even if they become moreilliquid at some point, this should not dramatically influence investors (asset managers andowners). Regulators and investors are more concerned by securities that are liquid undersome market conditions and illiquid under other market conditions. At first sight, it istherefore a paradox that liquidity stress testing programs must mainly focus on highly ormoderately liquid instruments than on illiquid instruments. In fact, liquidity does not likesurprises and changes. This is why the liquidity issue is related to the cross-section of theexpected illiquidity premium for illiquid assets, but to the time-series illiquidity variance forliquid assets.This is all the more important that the liquidity risk must be measured and managed ina stress testing framework, which adds another layer of complexity. Indeed, stress scenariosare always difficult to interpret, and calibrating them is a balancing act, because they mustcorrespond to extreme but also plausible events (Roncalli, 2020). This is why the historicalmethod is the most used approach when performing stress testing. However, it is very poorand not flexible in terms of risk management. Parametric approaches must be preferred sincestress periods are very heterogenous and outcomes are uncertain. Therefore, it makes moresense to estimate and use stressed liquidity parameters than directly estimate a stressedliquidity outcome. In this approach, the normal model is the baseline model on which wecould apply scenario analysis on the different parameters that define the liquidity model.This is certainly the best way to proceed if we want to develop a factor-based liquidity stresstesting program, which is an important issue for fund management. Otherwise, liquiditystress testing would be likely to remain a regulatory constraint or a pure exercise of riskmeasurement, but certainly not a risk management process supporting investment policiesand fund management.This paper is organized as follows. Section Two introduces the concept of redemptionrates and defines the historical approach of liquidity stress testing. In Section Three, weconsider parametric models that can be used to estimate redemption shocks. This impliesmaking the distinction between the redemption event and the redemption amount. From astatistical point of view, this is equivalent to modeling the redemption frequency and theredemption severity. After having developed an aggregate population model, we consideran individual-based model. It can be considered as a first attempt to develop a behavioralmodel, which is the central theme of Section Four. We analyze the simple case where re-demptions between investors are independent and then extend the model where redemptionsare correlated to take into account spillover effects and contagion risk. Then, we developfactor-based models of liquidity stress testing in Section Five. Finally, Section Six offerssome concluding remarks. 6iquidity Stress Testing in Asset Management
In order to assess the liquidity risk of an investment fund, we must model its ‘ funding ’liquidity. Therefore, managing the liquidity in asset management looks like a banking asset-liability management process (Roncalli, 2020). However, there is a major difference sincebanking ALM concerns both balance sheet and income statement. This is not the case of aninvestment fund, because we only focus on its balance sheet and the objective is to modelthe redemption flows.
In order to define the liability risk, we first have to understand the balance sheet of acollective investment fund. A simplified illustration is given in Figure 3 for a mutual fund.The total (gross) assets A ( t ) correspond to the market value of the investment portfolio.They include stocks, bonds and all financial instruments that are invested. On the liabilityside, we have two main balance sheet items. The first one corresponds to the debits D ( t ),which are also called current or accrued liabilities. They are all the expenses incurred bythe mutual fund. For instance, the current liabilities include money owed to lending banks,fees owed to the fund manager and the custodian, etc. The second liability item is the unitcapital C ( t ), which is owned by the investors. Each investor owns a number of units (orshares) and is referred to as a ‘ unitholder ’. This unit capital is equivalent to the equityconcept of a financial institution. A unitholder is then also called a shareholder in referenceto capital markets. Figure 3: Balance sheet of mutual funds A ( t ) C ( t ) D ( t ) Unit capitalDebitsTotalAssets Assets Liabilities2.1.1 Definition of net asset value
The total net assets (TNA) equal the total value of assets less the current or accrued liabil-ities: TNA ( t ) = A ( t ) − D ( t )The net asset value (NAV) represents the share price or the unit price. It is equal to:NAV ( t ) = TNA ( t ) N ( t ) (1)where the total number N ( t ) of shares or units in issue is the sum of all units owned byall unitholders. The previous accounting rules show that the capital is exactly equal to the7iquidity Stress Testing in Asset Managementtotal net assets, which is also called the assets under management (AUM). The investmentfund’s capital is therefore an endogenous variable and depends on the performance of thetotal net assets: C ( t ) = N ( t ) · NAV ( t )= TNA ( t )At time t +1, we assume that the portfolio’s return is equal to R ( t + 1). Since D ( t ) (cid:28) A ( t ),it follows that: TNA ( t + 1) = A ( t + 1) − D ( t + 1)= (1 + R ( t + 1)) A ( t ) − D ( t + 1) ≈ (1 + R ( t + 1)) · TNA ( t )meaning that: NAV ( t + 1) ≈ (1 + R ( t + 1)) · NAV ( t )The investment fund’s capital is therefore time-varying. It increases when the performanceof the asset is positive, and it decreases otherwise. Remark 1
In the sequel, we assume that the mutual fund is priced daily, meaning that theNAV of the mutual fund is calculated at the end of the market day. Therefore, the time t represents the current market day, whereas the time t + 1 corresponds to the next marketday. Let us now introduce the impact of subscriptions and redemptions. In this case, new andcurrent investors may purchase new mutual fund units, while existing investors may redeemall or part of their shares. Subscription and redemption orders must be known by the fundmanager before t + 1 in order to be executed. In this case, the number of units becomes: N ( t + 1) = N ( t ) + N + ( t + 1) − N − ( t + 1)where N + ( t + 1) is the number of units to be created and N − ( t + 1) is the number of unitsto be redeemed. At time t + 1, we have:NAV ( t + 1) = TNA ( t + 1) N ( t + 1)= TNA ( t + 1) N ( t ) + N + ( t + 1) − N − ( t + 1)We deduce that:TNA ( t + 1) = N ( t ) · NAV ( t + 1) + F + ( t + 1) − F − ( t + 1) (2)where F + ( t + 1) = N + ( t + 1) · NAV ( t + 1) and F − ( t + 1) = N − ( t + 1) · NAV ( t + 1) arethe investment inflows and outflows. Again, we notice that the investment fund’s capital istime-varying and depends on the fund flows.From Equation (2), we deduce that:TNA ( t + 1) = N ( t ) · NAV ( t + 1) + F + ( t + 1) − F − ( t + 1) ≈ N ( t ) · (1 + R ( t + 1)) · NAV ( t ) + F + ( t + 1) − F − ( t + 1)= (1 + R ( t + 1)) · TNA ( t ) + F + ( t + 1) − F − ( t + 1)8iquidity Stress Testing in Asset ManagementThe current net assets are approximatively equal to the previous net assets plus the perfor-mance value plus the net flow. We retrieve the famous formula of Sirri and Tufano (1998)when we want to estimate the net flow from the NAV and TNA of the fund: F ( t + 1) = F + ( t + 1) − F − ( t + 1)= TNA ( t + 1) − (1 + R ( t + 1)) · TNA ( t )= TNA ( t + 1) − (cid:18) NAV ( t + 1)NAV ( t ) (cid:19) TNA ( t ) (3) Since the capital is a residual, we face three liability risks. The first one deals with the ac-crued liabilities D ( t ). Generally, the debits are a very small part of the liabilities. However,we can potentially face some situations where the debits are larger than the assets, implyingthat the net asset value becomes negative. In particular, this type of situation occurs whenthe fund is highly leveraged. The second risk concerns the inflows. If the investment fundhas a big subscription, it may have some difficulties buying the financial instruments. Forinstance, this type of situation may occur when the fund must buy fixed-income securitiesin a bond bull market and it is difficult to find investors who are looking to sell bonds.The third liability risk is produced by the outflows. In this case, the fund manager mustsell assets, which could be difficult in illiquid and stressed market conditions. The last twosituations are produced when supply and demand dynamics are totally unbalanced (highersupply for buying assets or higher demand for selling assets). In this article, we focus onthe third liability risk, which is also called redemption risk. In order to assess an investment fund’s redemption risk, we need an objective measurementsystem, which is well scaled. For instance, the outflows F − ( t ) are not very interesting,because they depend on the investment fund’s assets under management. In fact, they mustbe scaled in order to be a homogeneous measure that can be used to compare the redemptionbehavior across time, across funds and across investors. The (gross) redemption rate is defined as the ratio between the fund’s redemption flows andtotal net assets: R ( t ) = F − ( t )TNA ( t ) (4)We verify the property that R ( t ) ∈ [0 , R ( t ) = 100 / F − ( t + 1) = R ( t + 1) · TNA ( t ) (5)In this case, R ( t + 1) is a random variable and is not known at the current time t . Byassuming that redemption rates are stationary, the challenge is then to model the associatedprobability distribution F . 9iquidity Stress Testing in Asset Management The guidelines on the liquidity stress testing published by ESMA (2019) refer to both grossand net redemptions:“
LST should be adapted appropriately to each fund, including by adapting:[...] the assumptions regarding investor behaviour (gross and net redemptions) ”(ESMA, 2019, page 36).Following this remark, we can also define the net flow rate by considering both inflows andoutflows: R ± ( t ) = F ( t )TNA ( t ) (6)This quantity is more complex than the previous one, because it cannot be used from anex-ante point of view: ˆ F ( t + 1) (cid:54) = R ± ( t + 1) · TNA ( t )The reason is that the outflows are bounded and cannot exceed the current assets undermanagement. This is not the case for the inflows. For example, we consider a fund witha size of $100 mn. By construction, we have ˆ F − ( t + 1) ≤ F + ( t + 1) > R ± ( t ) may not reflect the liability risk in a stress testing framework. Forexample, let us consider an asset class that has experienced a bull market over the lastthree years. Certainly, we will mainly observe positive net flows and a very small numberof observations with negative net flows. We may think that these data are not relevant forbuilding stress scenarios. More generally, if an asset manager uses net flow rates for stresstesting purposes, only the observations during historical stress periods are relevant, meaningthat the calibration is based on a small fraction of the dataset.In fact, the use of net flows is motivated by other considerations. Indeed, the computationof R ( t ) requires us to know the outflows F − ( t ) exactly. Moreover, as we will see later, R ( t )must be computed for all the investor categories that are present in the fund (retail, privatebanking, institutional, etc.). This implies in-depth knowledge of the fund’s balance sheetliability, meaning that the asset manager must have a database with all the flows of all theinvestors on a daily basis. From an industrial point of view, this is a big challenge in terms ofIT systems between the asset manager and the custodian. This is why many asset managersdon’t have the disaggregated information on the liability flows. An alternative measure isto compute the net redemption rate, which corresponds to the negative part of the net flowrate: R − ( t ) = max (cid:18) , − F ( t )TNA ( t ) (cid:19) It has the good mathematical property that R − ( t ) ∈ [0 , R − ( t ) = max (cid:18) , F − ( t ) − F + ( t )TNA ( t ) (cid:19) (7) In order to simplify the calculus, we do not take into account the daily performance of the fund. F − ( t ) = TNA ( t ) and F + ( t ) = 0. Moreover, wenotice that the net redemption rate is equal to the gross redemption rate when there are noinflows: R − ( t ) = max (cid:18) , F − ( t )TNA ( t ) (cid:19) = R ( t )Otherwise, we have: R − ( t ) < R ( t )From a risk management point of view, it follows that redemption shocks based on net re-demptions may be underestimated compared to redemption shocks based on gross redemp-tions. However, we will see later that the approximation R ( t ) ≈ R − ( t ) may be empiricallyvalid under some conditions. The computation of redemption rates only makes sense if they are homogeneous, coherentand comparable. Let us assume that we compute the redemption rate R ( t ) at the level ofthe asset management company, and we have the historical data for the last ten years. Byassuming that there are 260 market days per year, we have a sample of 2 600 redemptionrates. We can compute the mean, the standard deviation, different quantiles, etc. Does ithelp with building a stress scenario for a mutual fund? Certainly not, because redemptionsdepend on the specific investor behavior at the fund level and not on the overall investorbehavior at the asset manager level. For instance, we can assume that an investor doesnot have the same behavior if he is invested in an equity fund or a money market fund.We can also assume that the redemption behavior is not the same for a central bank, aretail investor, or a pension fund. Therefore, we must build categories that correspondto homogenous behaviors. Otherwise, we will obtain categories, whose behavior is non-stationary. But, without the stationarity property, risk measurement is impossible andstress testing is a hazardous exercise.Therefore, liability categorization is an important step before computing redemptionrates. For instance, ESMA (2019) considers four factors regarding investor behavior: in-vestor category, investor concentration, investor location and investor strategy. Even thoughthe last three factors are significant, the most important factor remains the investor type.For instance, AMF (2017, page 12) gives an example with the following investor types: largeinstitutional (tier one), small institutional (tier two), investment (or mutual) fund, privatebanking network and retail investor. Other categories can be added: central bank, sovereign,corporate, third-party distributor, employee savings plan, wealth management, etc. More-over, it is also important to classify funds into homogeneous buckets such as balanced funds,bond funds, equity funds, etc. An example of an investor/fund categorization matrix isgiven in Table 1. Remark 2
The granularity of the investor/fund classification is an important issue. It isimportant to have a very detailed classification at the level of the database in order to groupcategories together from a computational point of view. In order to calibrate stress scenarios,we must have a sufficient number of observations in each cell of the classification matrix.Let us for instance consider the case of central banks. We can suppose that their behavioris very different to the other investors. Therefore, it is important for an asset manager tobe aware of the liabilities with respect to central banks. Nevertheless, there are few centralbanks in the world, meaning we may not have enough observations for calibrating some cells(e.g. central bank/equity or central bank/real asset), and we have to merge some cells (acrossinvestor and fund categories). A b s o l u t e r e t u r n B a l a n ce d B o nd C o m m o d i t y E nh a n ce d t r e a s u r y E q u i t y M o n e y m a r k e t R e a l a ss e t S t r u c t u r e d Central bankCorporateInstitutionalInsuranceInternalPension fundRetailSovereignThird-party distributorWealth management
We consider a fund. We note TNA i ( t ) the assets under management of the investor i forthis fund. By definition, we have:TNA i ( t ) = N i ( t ) · NAV ( t )where NAV ( t ) is the net asset value of the fund and N i ( t ) is the number of units held bythe investor i for the fund. The fund’s assets under management are equal to:TNA ( t ) = (cid:88) k TNA ( k ) ( t )where TNA ( k ) ( t ) = (cid:80) i ∈IC ( k ) TNA i ( t ), and IC ( k ) is the k th investor category. It followsthat: TNA ( t ) = (cid:88) k (cid:88) i ∈IC ( k ) TNA i ( t )= (cid:88) k (cid:88) i ∈IC ( k ) N i ( t ) · NAV ( t )= N ( t ) · NAV ( t )where N ( t ) = (cid:80) k (cid:80) i ∈IC ( k ) N i ( t ) is the total number of units in issue. We retrieve thedefinition of the assets under management (or total net assets) at the fund level. We canobtain a similar breakdown for the outflows : F − ( t ) = (cid:88) k (cid:88) i ∈IC ( k ) F − i ( t ) = (cid:88) k F − ( k ) ( t )The redemption rate for the investor category IC ( k ) is then equal to: R ( k ) ( t ) = F − ( k ) ( t )TNA ( k ) ( t ) (8) We have F − k ( t ) = (cid:80) i ∈IC ( k ) F − i ( t ). R ( t ) = F − ( t )TNA ( t )= (cid:80) k F − ( k ) ( t )TNA ( t )= (cid:80) k TNA ( k ) ( t ) · R ( k ) ( t )TNA ( t )= (cid:88) k ω ( k ) ( t ) · R ( k ) ( t ) (9)where ω ( k ) ( t ) represents the weights of the investor category IC ( k ) in the fund: ω ( k ) ( t ) = TNA ( k ) ( t )TNA ( t )Equation (9) is very important, because it shows that the redemption rate at the fund levelis a weighted-average of the redemption rates of the different investor categories.Let us now consider different funds. We note R ( f,k ) ( t ) as the redemption rate of theinvestor category IC ( k ) for the fund f at time t . By relating the fund f to its fund category FC ( j ) , we obtain a database of redemption rates by investor category IC ( k ) and fund category FC ( j ) : DB ( j,k ) ( T ) = (cid:8) R ( f,k ) ( t ) : f ∈ FC ( j ) , t ∈ T (cid:9) DB ( j,k ) ( T ) is then the sample of all redemption rates of the investor category IC ( k ) for all thefunds that fall into the fund category FC ( j ) during the observation period T . We notice that DB ( j,k ) ( t ) does not have a unique element for a given date t because we generally observeseveral redemptions at the same date for different funds and the same investor category. The key parameter for computing the redemption flows is the redemption rate, which isdefined for an investor category and a fund category. It is not calibrated at the fund level,because past redemption data for a given specific fund are generally not enough to obtaina robust estimation. This is why we have pooled redemption data as described in theprevious paragraph. Using these data, we can estimate the probability distribution F of theredemption rate and define several statistics that can help to build stress scenarios. In what follows, we consider the liability data provided by Amundi Asset Management fromJanuary, 1 st th Amundi Cube Database ’ andcontains 1 617 403 observations if we filter based on funds with assets under managementgreater than e is given in Table 40 on page 92.The number of observations is 464 399 for retail investors, 310 452 for third-party distribu-tors, 267 600 for institutionals, etc. The investor category which is the smallest is centralbanks with 15 523 observations. In terms of fund categories, bond, equity and balancedfunds dominate with respectively 452 942, 436 401 and 361 488 observations. The smallest The Amundi database contains 13 investor and 13 fund categories.
Remark 3
In what follows, we apply a filter that consists in removing observations thatcorresponds to dedicated mutual funds (FCP and SICAV) and mandates (see Table 41 onpage 93). The motivation is to focus on mutual funds with several investors, and this issuewill be extensively discussed in Section 4.1.3 on page 44.
Figure 4: Retail investor
20 40 60 80 100-100-50050100 0 20 40 60 80 10002040608010020 40 60 80 100-100-50050100 0 20 40 60 80 100020406080100
We first begin by comparing the gross redemption rate R , the net flow rate R ± and thenet redemption rate R − . Some results are given in Figures 4 and 5 for retail and insuranceinvestors and bond and equity funds. In the case of insurance companies, we notice that theapproximation R ≈ R − ≈ − R ± is valid when the redemption rate is greater than 20%. Thisis not the case for retail investors, where we observe that some large redemptions may beoffset by large subscriptions . The difference between retail and insurance categories lies inthe investor concentration. When an investor category is concentrated, there is a low prob-ability that this offsetting effect will be observed. This is not the case when the granularityof the investor category is high. We also observe that the approximation R ≈ R − ≈ − R ± depends on the fund category. For instance, it is not valid for money market funds. Thereason is that we generally observe subscriptions in a bull market and redemptions in a bearmarket when the investment decision mainly depends on the performance of the asset class.This is why large redemptions and subscriptions tend to be mutually exclusive (in the math-ematical sense) in equity or bond funds. The mutual exclusivity property is more difficult We observe the same phenomenon when we consider the data of third-party distributors (see Figure 37on page 94).
20 40 60 80 100-100-50050100 0 20 40 60 80 10002040608010020 40 60 80 100-100-50050100 0 20 40 60 80 100020406080100
Figure 6: Money market fund
20 40 60 80 100-100-50050100 0 20 40 60 80 10002040608010020 40 60 80 100-100-50050100 0 20 40 60 80 100020406080100
For a given investor/fund category, we note F as the probability distribution of the redemp-tion rate. We can define several risk measures (Roncalli, 2020, pages 62-63): • the mean: M = (cid:90) x d F ( x ) • the standard deviation-based risk measure: SD ( c ) = M + c (cid:90) (cid:0) x − M (cid:1) d F ( x ) • the quantile (or the value-at-risk) at the confidence level α : Q ( α ) = F − ( α ) • the average beyond the quantile (or the conditional value-at-risk): C ( α ) = E (cid:2) R | R ≥ F − ( α ) (cid:3) The choice of a risk measure depends on its use. For instance, M can be used by the fundmanager daily, because it is the expected value of the daily redemption rate. If the fundmanager prefers to have a more conservative measure, he can use SD (1). M and SD ( c ) makesense in normal periods from a portfolio management perspective, but they are less relevantin a stress period. This is why it is better to use Q ( α ) and C ( α ) from a risk managementpoint of view. In the asset management industry, the consensus is to set α = 99%.In Table 2, we have reported the values of M , Q (99%) and C (99%) by considering theempirical distribution of gross redemption rates by client category. We do not consider the SD -measure because we will see later that there is an issue when it is directly computedfrom a sample of historical redemption rates. On average, the expected redemption rate isroughly equal to 20 bps. It differs from one client category to another, since the lowest valueof M is observed for central banks whereas the highest value of M is observed for corporates.The 99% value-at-risk is equal to 3 . .
5% every 100 days, that is every five months. Again, there are some big differencesbetween the client categories. The riskiest category is corporate followed by sovereign andauto-consumption. If we focus on the conditional value-at-risk, we are surprised by the highvalues taken by C (99%). If we consider all investor categories, C (99%) is more than 15%,and the ratio R (99%) between C (99%) and Q (99%) is equal to 4 .
51. This is a very highfigure since this ratio is generally less than 2 for market and credit risks. For example,in the case of a Gaussian distribution N (cid:0) , σ (cid:1) , the ratio R ( α ) between the conditionalvalue-at-risk and the value-at-risk is equal to: R ( α ) = C ( α ) Q ( α ) = φ (cid:0) Φ − ( α ) (cid:1) (1 − α ) Φ − ( α )16iquidity Stress Testing in Asset ManagementThis ratio is respectively equal to 1 .
37 and 1 .
15 when α = 90% and α = 99%. Moreover,Roncalli (2020, page 118) showed that it is a decreasing function of α and:lim α → − R ( α ) = 1We deduce that the ratio is lower than 1 . α .Therefore, the previous figure R (99%) = 4 .
51 indicates that redemption risk is more skewedthan market and credit risks.Table 2: Redemption statistical measures in % by investor categoryClient
M Q (99%) C (99%) R (99%)Auto-consumption 0 .
38 7 .
44 24 .
86 3 . .
04 0 .
00 4 . ∞ Corporate 0 .
54 12 .
71 28 .
21 2 . .
13 0 .
50 13 .
06 26 . .
06 1 .
13 4 .
86 4 . .
27 5 .
11 22 .
79 4 . .
26 5 .
25 21 .
24 4 . .
23 3 .
41 20 .
22 5 . .
15 1 .
92 9 .
18 4 . .
45 8 .
28 39 .
85 4 . .
23 3 .
90 13 .
72 3 . .
22 3 .
50 15 .
79 4 . M Q (99%) C (99%) R (99%)Balanced 0 .
14 1 .
77 8 .
14 4 . .
20 3 .
18 14 .
23 4 . .
40 6 .
30 31 .
15 4 . .
18 2 .
68 12 .
94 4 . .
06 21 .
76 46 .
13 2 . .
11 1 .
19 9 .
32 7 . .
04 0 .
45 3 .
52 7 . .
22 3 .
50 15 .
79 4 . M -statistic in % by investor/fund category(1) (2) (3) (4) (5) (6) (7) (8)Auto-consumption 0 .
27 0 .
36 0 .
65 0 .
30 1 .
58 0 .
18 0 . .
01 0 .
06 0 .
11 0 . .
08 0 .
15 0 .
27 0 .
25 1 .
52 0 .
07 0 . .
17 0 .
05 0 .
10 0 .
10 0 .
55 0 .
00 0 . .
03 0 .
05 0 .
13 0 .
06 0 .
06 0 .
08 0 . .
13 0 .
16 0 .
64 0 .
18 1 .
47 0 .
06 0 . .
17 0 .
15 0 .
12 0 .
16 0 .
90 0 .
08 0 . .
08 0 .
10 0 .
33 0 .
21 0 .
76 0 .
02 0 . .
15 0 .
14 0 .
26 0 .
16 0 .
91 0 .
07 0 .
04 0 . .
01 0 .
01 0 .
16 0 .
19 1 .
91 0 .
06 0 . .
12 0 .
24 0 .
67 0 .
19 0 .
92 0 .
28 0 .
08 0 . .
14 0 .
20 0 .
40 0 .
18 1 .
06 0 .
11 0 .
04 0 . Q -statistic in % by investor/fund category(1) (2) (3) (4) (5) (6) (7) (8)Auto-consumption 2 .
93 7 .
57 12 .
62 5 .
46 25 .
98 3 .
23 7 . .
00 0 .
00 0 .
12 0 . .
30 1 .
58 4 .
90 3 .
88 24 .
14 0 .
00 12 . .
39 0 .
05 1 .
30 0 .
03 13 .
09 0 .
00 0 . .
06 1 .
70 2 .
35 1 .
08 2 .
51 0 .
25 1 . .
84 1 .
94 8 .
68 3 .
10 34 .
82 0 .
00 5 . .
32 0 .
21 3 .
87 0 .
50 18 .
39 0 .
00 5 . .
73 0 .
56 2 .
40 2 .
20 14 .
75 0 .
05 3 . .
01 1 .
50 4 .
72 1 .
65 18 .
36 1 .
17 0 .
45 1 . .
11 0 .
14 7 .
98 0 .
22 66 .
36 0 .
00 8 . .
32 4 .
59 11 .
13 3 .
38 14 .
66 3 .
96 1 .
11 3 . .
77 3 .
18 6 .
30 2 .
68 21 .
76 1 .
19 0 .
45 3 . C -statistic in % by investor/fund category(1) (2) (3) (4) (5) (6) (7) (8)Auto-consumption 21 .
08 23 .
37 40 .
73 21 .
24 54 .
96 15 .
50 24 . .
28 6 .
05 10 .
11 4 . .
31 14 .
98 22 .
80 22 .
48 38 .
37 6 .
52 28 . .
22 5 .
14 9 .
24 9 .
58 32 .
33 0 .
00 13 . .
48 3 .
16 10 .
60 4 .
91 4 .
97 7 .
91 4 . .
99 15 .
40 62 .
30 16 .
27 58 .
10 6 .
26 22 . .
35 14 .
65 10 .
59 15 .
32 37 .
28 7 .
62 21 . .
45 9 .
84 32 .
56 18 .
61 46 .
88 2 .
17 20 . .
02 8 .
34 15 .
99 8 .
95 44 .
38 5 .
03 3 .
03 9 . .
39 1 .
35 15 .
20 17 .
97 86 .
47 5 .
73 39 . .
69 14 .
53 42 .
24 11 .
22 32 .
68 20 .
16 6 .
85 13 . .
14 14 .
23 31 .
15 12 .
94 46 .
13 9 .
32 3 .
52 15 . (1) = balanced, (2) = bond, (3) = enhanced treasury, (4) = equity, (5) = money market, (6) = other, (7)= structured, (8) = total for the classification matrix are given in Tables 4,5 and 6. We notice that the two dimensions are important, since one dimension does notdominate the other. This means that a low-risk (resp. high-risk) investor category tendsto present the lowest (resp. highest) redemption statistics whatever the fund category. Inaddition, the ranking of redemption statistics between fund categories is similar whateverthe investor category. Nevertheless, we observe some exceptions and new stylized facts. Forinstance, we have previously noticed that bond and equity funds have similar redemptionrates on average. This is not the case for the corporate, corporate pension fund and sovereigncategories, for which historical C -statistics are more important for equity funds than bondfunds. For the corporate pension fund category, the risk is also higher for balanced fundsthan for bond funds. According to BCBS (2017, page 60), a historical stress scenario “ aims at replicating thechanges in risk factor shocks that took place in an actual past episode ”. If we apply thisdefinition to the redemption risk, the computation of the historical stress scenario is simple.First, we have to choose a stress period T stress and second, we compute the maximumredemption rate: X (cid:0) T stress (cid:1) = max t ∈ T stress R ( t )For example, if we apply this definition to our study period, we obtain the results givenin Table 7. We recall that the study period runs from January 2019 to August 2020 andincludes the Coronavirus pandemic crisis, which was a redemption stress period. We observethat the X -statistic is generally equal to 100%! This is a big issue, because it is not helpfulto consider that liquidity stress testing of liabilities leads to a figure of 100%. The problem isthat the X -statistic is not adapted to redemption risk. Let us consider an investor category IC ( k ) and a fund category FC ( j ) . The X -statistic is computed by taking the maximum ofall redemption rates for all funds that belong to the fund category: X ( j,k ) (cid:0) T stress (cid:1) = max t ∈ T stress (cid:8) R ( f,k ) ( t ) : f ∈ FC ( j ) (cid:9) If there is one fund with only one investor and if this investor redeems 100%, X ( j,k ) ( T stress )is equal to 100%. However, the asset manager does not really face a liquidity risk in thissituation, because there is no other investor in this fund. So, the other investors are notpenalized. We have excluded this type of fund. However, we face a similar situation in manyother cases: small funds with a large fund holder, funds with a low number of unitholders,etc. Moreover, this type of approach penalizes big asset managers, which have hundreds offunds. Let us consider an example. For a given investor/fund category, the fund manager A has 100 funds of $100 million, whereas the fund manager B has one fund of $10 billion. Froma theoretical point of view, A and B face the same redemption risk, since they both manage$10 billions for the same investor/fund category. However, it is obvious that X A (cid:29) X B ,meaning that the historical stress scenario for the fund manager A will be much higher thanthe historical stress scenario for the fund manager B . This is just a probabilistic countingprinciple as shown in Appendix A.1 on page 78. If we consider the previous example, thehistorical stress scenario for the fund manager A is larger than 99 .
9% when the historicalstress scenario for the fund manager B is larger than 6 .
68% (see Figure 38 on page 95).More generally, the two stress scenarios are related in the following manner: X n = 1 − (1 − X ) n where X is the X -measure for one fund and X n is the X -measure for n funds. They are not calculated if the number of observations is less than 200. X -statistic in % by investor/fund category(1) (2) (3) (4) (5) (6) (7)Auto-consumption 100 .
00 100 .
00 100 .
00 100 .
00 99 .
65 100 . .
17 29 .
60 50 . .
64 83 .
44 100 .
00 94 .
14 97 .
72 100 . .
00 100 .
00 15 .
79 100 .
00 94 .
78 0 . .
79 15 .
35 100 .
00 100 .
00 14 .
71 100 . .
09 100 .
00 100 .
00 100 .
00 100 .
00 100 . .
99 100 .
00 56 .
96 100 .
00 99 .
93 77 . .
00 100 .
00 100 .
00 100 .
00 100 .
00 100 . .
00 100 .
00 100 .
00 100 .
00 100 .
00 100 .
00 100 . .
44 21 .
12 24 .
91 100 .
00 100 .
00 100 . .
00 100 .
00 100 .
00 100 .
00 97 .
04 100 .
00 97 . (1) = balanced, (2) = bond, (3) = enhanced treasury, (4) = equity, (5) = money market, (6) = other, (7)= structured Remark 4
Another approach consists in computing the average redemption rate daily: R ( j,k ) ( t ) = (cid:88) f ∈FC ( j ) TNA ( f ) (cid:80) f ∈FC ( j ) TNA ( f ) R ( f,k ) ( t ) where the weights are proportional to the size of funds f that belong to the j th fund category FC ( j ) . In this case, we have: X ( j,k ) (cid:0) T stress (cid:1) = max t ∈ T stress R ( j,k ) ( t ) This method does not have the previous drawback, but it has other shortcomings such as aninformation loss. However, the biggest disadvantage is that the historical stress scenario isgenerally based on the largest fund, except when the funds have similar size.
Since X -measures can not be used to build redemption shocks, we propose using Q or C -measures. Q (99%) is the daily value-at-risk at the 99% confidence level. This meansthat its return period is 100 days. On average, we must observe that redemption shocks aregreater than Q (99%) two and a half times per year. We can also use the conditional value-at-risk C (99%) if we want more severe redemption shocks. The drawback of C (99%) is thatwe don’t know the return period of such event. However, it does make sense because it is avery popular measure in risk management, and it is well received by regulatory bodies andsupervisors (Roncalli, 2020). Nevertheless, we must be cautious about the computed figuresobtained in Tables 5 and 6 on page 18. For example, we don’t have the same confidencelevel between the matrix cells, because the estimates are not based on the same numberof observations. In the case of retail investors or third-party distributors, we generally usea huge number of observations whereas this is not the case with the other categories. InTable 8, we give an example of confidence level codification. We see that some cells are notwell estimated since the number of observations is less than 10 000. For some of them, thenumber of observations is very low (less than 200), implying that the confidence on theseestimates is very poor.Therefore, the estimated values cannot be directly used as redemption shocks. However,they help risk managers and business experts to build redemption shocks. Starting fromthese figures, they can modify them and build a table of redemption shocks that respectthe risk coherency C investor between investor categories and the risk coherency C fund between20iquidity Stress Testing in Asset ManagementTable 8: Confidence in estimated values with respect to the number of observations(1) (2) (3) (4) (5) (6) (7)Auto-consumption • • • • • • • • • • • • • • • • ◦ ◦ ◦ Central bank • • • ◦ ◦ ◦ • ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
Corporate • • • • • • • • • • • • ◦ ◦ ◦
Corporate pension fund • • • • • • • • • • • ◦ ◦ ◦
Employee savings plan • • • • • • • • • • • ◦ ◦ ◦ • •
Institutional • • • • • • • • • • • • • • • ◦ ◦ ◦
Insurance • • • • • • • • • • • • • • ◦ ◦ ◦
Other • • • • • • • • • • • • • • ◦ ◦ ◦
Retail • • • • • • • • • • • • • • • • • • • •
Sovereign • • • • • • • • • • • ◦ ◦ ◦
Third-party distributor • • • • • • • • • • • • • • • • • • • ◦ ◦ ◦ − ◦ ◦ − ◦ − • − • • −
10 000, • • • +10 000 fund categories . The risk coherency C investor means that if one investor category is assumedto be riskier than another, the global redemption shock of the first category must be greaterthan the global redemption shock of the second category: IC ( k ) (cid:31) IC ( k ) ⇒ S ( k ) ≥ S ( k ) For example, if we consider the Q -measure, we can propose the following risk ordering:1. central bank, corporate pension fund2. employee savings plan, retail3. other, third-party distributor4. institutional, insurance5. auto-consumption, corporate, sovereignIn this case, the redemption shock S ( j,k ) for the ( j, k )-cell depends on the global redemptionshock S ( k ) for the investor category IC ( k ) . For instance, we can set the following rule ofthumb: S ( j,k ) = m ( j ) · S ( k ) (10)where m ( j ) is the multiplicative factor of the fund category FC ( j ) . In a similar way, the riskcoherency C fund means that if one fund category is assumed to be riskier than another, theglobal redemption shock of the first category must be greater than the global redemptionshock of the second category: FC ( j ) (cid:31) FC ( j ) ⇒ S ( j ) ≥ S ( j ) For example, if we consider the Q -measure, we can propose the following risk ordering:1. structured2. balanced, other3. bond, equity For instance, if we consider the sovereign category, it is difficult to explain the big difference of C (99%)between bond and equity funds S ( j,k ) for the ( j, k )-cell depends then on the redemption shock S ( j ) for the fund category IC ( j ) . Again, we can set the following rule of thumb: S ( j,k ) = m ( k ) · S ( j ) (11)where m ( k ) is the multiplicative factor of the investor category IC ( k ) . We can also combinethe two rules of thumb and we obtain the mixed rule: S ( j,k ) = m ( k ) · S ( j ) + m ( j ) · S ( k ) Q -measure. Table 9 gives anexample of S ( j,k ) by considering the risk coherency C investor , whereas Table 10 correspondsto the risk coherency C fund . The mixed rule is reported in Table 11. These figures can thenbe modified by risk managers and business experts by considering the specificity of somematrix cells. For instance, it is perhaps not realistic to have the same redemption shockfor balanced funds between auto-consumption and corporates. Moreover, these redemptionshocks can also be modified by taking into account the C -measure. For instance, the con-ditional value-at-risk for bond funds is much higher for third-party distributors than forsovereigns. Perhaps we can modify the redemption shock of 3 .
3% and have a larger valuefor third-party distributors. It is even more likely that the estimated values of Q and C are based on 75 591 observations for the third-party distributor category, and 2 261 for thesovereign category. Therefore, we can consider that the estimated value of 4 .
59% obtainedin Table 5 on page 18 does make more sense than the proposed value of 3 .
3% obtained inTable 11 for the third-party distributor/bond matrix cell. In a similar way, we can considerthat the estimated value of 0 .
14% does make less sense than the proposed value of 7 .
0% forthe sovereign/bond matrix cell.The previous analysis shows that building redemption shocks in a stress testing frame-work is more of an art than a science. A pure quantitative approach is dangerous becauseit is data-driven and it does not respect some coherency properties. However, historicalstatistics are very important because they provide an anchor point for risk managers andbusiness experts in order to propose stress scenarios that are satisfactory from regulatory,risk management and fund management points of view. Historical data are also importantbecause they help to understand the behavior of clients. It is different from one fund cate-gory to another, it also depends on the granularity of the classification, it may depend onthe time period, etc. In what follows, we complete this pure historical analysis using moretheoretical models. These models are important, because an historical approach is limitedwhen we want to understand contagion effects between investors, correlation patterns be-tween funds, time properties of redemption risk, the impact of the holding period, etc. Theidea is not to substitute one model with another, but to rely on several approaches, becausethere is not just one single solution to the liability stress testing problem. We use the following values: S ( k ) = 0 .
5% for central banks and corporate pension funds, S ( k ) = 2%for employee savings plans and retail, S ( k ) = 3 .
5% for other and third-party distributors, S ( k ) = 5% forinstitutionals and insurance companies, and S ( k ) = 8% for auto-consumption, corporates and sovereigns.For the multiplication factor, we assume that m ( j ) = 0 .
25 for structured, m ( j ) = 0 . m ( j ) = 1 for bond and equity, m ( j ) = 1 .
75 for enhanced treasury, and m ( j ) = 6 for money market. We use the following values: S ( j ) = 0 .
5% for structured, S ( j ) = 1 .
5% for balanced and other, S ( j ) =3% for bond and equity, S ( j ) = 5% for enhanced treasury, and S ( j ) = 20% for money market. For themultiplication factor, we assume that m ( k ) = 0 .
25 for central banks and corporate pension funds, m ( k ) = 0 . m ( k ) = 1 for other and third-party distributors, m ( k ) = 1 . m ( k ) = 2 for auto-consumption, corporates and sovereigns. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . (1) = balanced, (2) = bond, (3) = enhanced treasury, (4) = equity, (5) = money market, (6) = other, (7)= structured, (8) = total The direct computation of value-at-risk, conditional value-at-risk and other statistics fromhistorical redemption rates is particularly problematic. Indeed, we observe a large proportionof zeros in the redemption rate database. On average, we have 68 .
9% of zeros, this proportionreaches 99 .
5% for some investors and it is more than 99 .
9% for some matrix cells. Therefore,the data of redemption rates are “ clumped-at-zero ”, meaning that the redemption rate is asemi-continuous random variable, and not a continuous random variable (Min and Agresti,2002). This discontinuity is a real problem when estimating the probability distribution F .This is why we consider that the redemption rate is not the right redemption risk factor. Weprefer to assume that the redemption risk is driven by two dimensions or two risk factors:1. the redemption frequency, which measures the occurrence E of the redemption;2. the redemption severity R (cid:63) , which measures the amount of the redemption.It is obvious that this modeling approach finds its root in other risk models that deal withextreme events or counting processes, such as operational and insurance risks (Roncalli,2020). In the frequency-severity approach, we distinguish the redemption event E that indicates ifthere is a redemption, and the redemption amount R (cid:63) that measures the redemption ratein case of a redemption. An example is provided in Figure 7. The probability to observea redemption is equal to 5%, and in the case of a redemption, the amount can be 2%, 5%,15% and 50%. It follows that the redemption rate is the convolution of two risk factors.Figure 7: Zero-inflated modeling of the redemption risk E E = 1 R (cid:63) = 50% R (cid:63) = 15% R (cid:63) = 5% R (cid:63) = 2% E = 095% R = 0% ( Pr = 95% ) R = 2% ( Pr = 3% ) R = 5% ( Pr = 1 . ) R = 15% ( Pr = 0 . ) R = 50% ( Pr = 0 . ) We assume that the redemption event E follows a bernoulli distribution B ( p ), whereas theredemption severity R (cid:63) follows a continuous probability distribution G . We have:Pr {E = 1 } = Pr { R > } = p and: Pr { R ≤ x | E = 1 } = G ( x )We deduce that the unconditional probability distribution of the redemption rate is givenby: F ( x ) = Pr { R ≤ x } = { x ≥ } · (1 − p ) + { x > } · p · G ( x )Its density probability function is singular at x = 0: f ( x ) = (cid:26) − p if x = 0 p · g ( x ) otherwisewhere g ( x ) is the density function of G . Some examples are provided in Figure 8 when p = 5%. We observe that the density function is composed of a dirac measure and acontinuous function. In the case of G , the distribution is right-skewed, meaning that theprobability to observe small redemptions is high. In the case of G , we have a bell curve,meaning that the redemption amount is located around the mean if there is a redemption.Finally, the distribution is left-skewed in the case of G , meaning that the probability toobserve high redemptions is high if there is of course a redemption, because we recall thatthe probability to observe a redemption is only equal to 5%. It is defined as the non-zero redemption rate. R = E· R (cid:63) In Appendix A.2.1 on page 78, we show that: E [ R ] = p E [ R (cid:63) ] (13)and: σ ( R ) = pσ ( R (cid:63) ) + p (1 − p ) E [ R (cid:63) ] (14)Moreover, the skewness coefficient is equal to: γ ( R ) = ϑ ( R (cid:63) )( pσ ( R (cid:63) ) + p (1 − p ) E [ R (cid:63) ]) / (15)where: ϑ ( R (cid:63) ) = pγ ( R (cid:63) ) σ ( R (cid:63) ) + 3 p (1 − p ) σ ( R (cid:63) ) E [ R (cid:63) ] + p (1 − p ) (1 − p ) E [ R (cid:63) ]For the excess kurtosis coefficient, we obtain: γ ( R ) = ϑ ( R (cid:63) )( pσ ( R (cid:63) ) + p (1 − p ) E [ R (cid:63) ]) (16)where: ϑ ( R (cid:63) ) = ( pγ ( R (cid:63) ) + 3 p (1 − p )) σ ( R (cid:63) ) + 4 p (1 − p ) γ ( R (cid:63) ) σ ( R (cid:63) ) E [ R (cid:63) ] +6 p (1 − p ) (1 − p ) σ ( R (cid:63) ) E [ R (cid:63) ] + p (1 − p ) (cid:0) − p + 6 p (cid:1) E [ R (cid:63) ]In Figure 9 we have reported the moments of the redemption rate R by considering thefollowing set of parameters: E [ R (cid:63) ] = 40%, σ ( R (cid:63) ) = 20%, γ ( R (cid:63) ) = 0 and γ ( R (cid:63) ) = 0; E [ R (cid:63) ] = 20%, σ ( R (cid:63) ) = 20%, γ ( R (cid:63) ) = 0 and γ ( R (cid:63) ) = 0; E [ R (cid:63) ] = 40%, σ ( R (cid:63) ) = 40%, γ ( R (cid:63) ) = − γ ( R (cid:63) ) = 0; E [ R (cid:63) ] = 40%, σ ( R (cid:63) ) = 20%, γ ( R (cid:63) ) = 0 and γ ( R (cid:63) ) = 1.We notice that the parameter values of R (cid:63) have a major impact on the statistical moments,but the biggest effect comes from the frequency probability p . Indeed, we verify the followingproperties: (cid:26) lim p → + E [ R ] = lim p → + σ ( R ) = 0lim p → + γ ( R ) = lim p → + γ ( R ) = ∞ (17)This means that the redemption risk is very high for small frequency properties. In thiscase, the expected redemption rate and its standard deviation are very low, but skewnessand kurtosis risk are very high! This creates a myopic situation where the asset managermay have the feeling that redemption risk is not a concern because of historical data. Indeed,when p is low, the probability of observing large redemption rates is small, implying that theyare generally not observed in the database. For instance, let us consider two categories thathave the same redemption severity distribution, but differ from their redemption frequencyprobability. One has a probability of 50%, the other has a probability of 1%. It is notobvious that the second category experienced sufficient severe redemption events such thatthe historical data are representative of the severity risk.26iquidity Stress Testing in Asset ManagementFigure 9: Statistical moments of the redemption rate R in zero-inflated models For the M -measure, we have: M = p E [ R (cid:63) ] (18)The formula of the value-at-risk is equal to: Q ( α ) = p ≤ − α G − (cid:18) α + p − p (cid:19) otherwise (19)We notice that computing the quantile α of the unconditional distribution F is equivalentto compute the quantile α G of the severity distribution G : α G = max (cid:18) , α + p − p (cid:19) The relationship between p , α and α G is illustrated in Figure 39 on page 95. Let us focuson the 99% value-at-risk: Q (99%) = p ≤ G − (cid:18) p − p (cid:19) otherwiseIf the redemption frequency probability is greater than 1%, the value-at-risk correspondsto the quantile ( p − /p . The relationship between p and α G = ( p − /p is shownin Figure 10. If p is greater than 20%, α G is greater than 95%. If p is less than 5%, weobserve a high curvature of the relationship, implying that we face a high estimation risk.For instance, if p is equal to 1 . . p becomes 2 . p is low, implying that a small error in the estimated value of p leads toa high impact on the value-at-risk.Figure 10: Relationship between p and α G for the 99% value-at-risk For the conditional value-at-risk, we obtain: C ( α ) = 11 − α (cid:90) α Q ( u ) d u (20)where Q ( u ) is the quantile function of R for the confidence level u . In the case where p > − α , we obtain: C ( α ) = p − α (cid:90) − p − (1 − α ) G − ( u ) d u Another expression of the conditional value-at-risk is: C ( α ) = 11 − α (cid:90) Q ( α ) x d F ( x )In the case where p > − α , we obtain: C ( α ) = p − α (cid:90) Q ( α ) xg ( x ) d x where g ( x ) is the probability density function of G ( x ). All these formulas can be computednumerically thanks to Gauss-Legendre integration.We now introduce a new risk measure which is very popular when considering parametricmodel. Roncalli (2020) defines the distribution-based (or parametric-based) stress scenario28iquidity Stress Testing in Asset Management S ( T ) for a given horizon time T such that the return time of this scenario is exactly equalto T . From a mathematical point of view, we have:1Pr { R ≥ S ( T ) } = T Pr { R ≥ S ( T ) } is the exceedance probability of the stress scenario, implying that the quan-tity Pr { R ≥ S ( T ) } − is the return time of the exceedance event. For example, if we set S ( T ) = Q ( α ), we have Pr { R ≥ S ( T ) } = 1 − α and T = (1 − α ) − . The return timeassociated to a 99% value-at-risk is then equal to 100 days, the return time associated to a99 .
9% value-at-risk is equal to 1 000 days (or approximately 4 years), etc. This parametricapproach of stress testing is popular among professionals, regulators and academics whenthey use the extreme value theory for modeling the risk factors.By combining the two definitions S ( T ) = Q ( α ) and T = (1 − α ) − , we obtain themathematical expression of the parametric stress scenario: S ( T ) = Q (cid:18) − T (cid:19) (21)If we consider the zero-inflated model, we deduce that: S ( T ) = p ≤ T − G − (cid:18) − p T (cid:19) otherwise (22)The magnitude of T is the year, but the unit of T is the day. For example, since one yearcorresponds to 260 market days, the five-year stress scenario is equal to : S (5) = G − (cid:18) − p (cid:19) The choice of the severity distribution is an important issue. Since R (cid:63) is a random variablebetween 0 and 1, it is natural to use the two-parameter beta distribution B ( a, b ). We have: G ( x ) = B ( x ; a, b )where B ( x ; a, b ) is the incomplete beta function. The corresponding probability densityfunction is equal to: g ( x ) = x a − (1 − x ) b − B ( a, b )where B ( a, b ) is the beta function: B ( a, b ) = Γ ( a ) Γ ( b )Γ ( a + b )Concerning the statistical moments, the formulas are given in Appendix A.2.2 on page 81.We report some examples of density function in Figure 11. Instead of providing theparameters a and b , we have indicated the value µ and σ of the mean and the volatility. Thefirst distribution is skewed, because the volatility is high compared to the mean. The otherthree distributions have a mode. Figure 12 shows the corresponding statistical moments of29iquidity Stress Testing in Asset ManagementFigure 11: Density function of the beta distribution Figure 12: Statistical moments of the zero-inflated beta distribution Q -measure highly depends on the redemption frequency p . Again, we observe that thesensitivity of the value-at-risk is particularly important when p is small . The ratio betweenthe 99% conditional value-at-risk and the 99% value-at-risk is given in Figure 14. When theredemption frequency p is high, the ratio is less than 1 . . When the redemption frequency p is small,the ratio may be greater than 2 .
0. These results shows that the sensitivity to redemptionrisk is very high when the observed redemption frequency is low. The stress scenarios S ( T )are given in Figure 15 when the redemption frequency p is equal to 1%. By definition, S ( T )increases with the return time T . From a theoretical point of view, the limit of the stressscenario is 100%: lim T →∞ S ( T ) = lim T →∞ G − (cid:18) − p T (cid:19) = 1However, we observe that stress scenarios reach a plateau at five years, meaning that stressscenarios beyond 5 years have no interest. This is true for small values of p , but it is evenmore the case for larger values of p as shown in Figures 40 and 41 on page 96.Figure 13: Q (99%)-measure in % with respect to the redemption frequency Remark 5
In order to better understand the use of the C -measure as a stress scenario,we compute the implied return time such that the stress scenario is exactly equal to the We assume that the redemption frequency is greater than 1 / .
69 bps. Otherwise, the quantileis equal to zero. Because of the impact of p on the confidence level α G — see Figure 10 on page 28. When p tends to one, the ratio is respectively equal to 1 .
15, 1 .
09, 1 .
06 and 1 .
03 for the four probabilitydistributions of the redemption severity. R (99%) with respect to the redemption frequency Figure 15: Stress scenario S ( T ) in % ( p = 1%) conditional value-at-risk: T C ( α ) = {T : S ( T ) = C ( α ) } Results are given in Table 12. We notice that the value is between . and . . On average,we can consider that the return time of the conditional value-at-risk is about one year.This is . times the return time of the value-at-risk . Table 12: Implied return time T C (99%) in year µ
10% 20% 30% 50% σ
10% 10% 20% 20%1% 1 .
03 0 .
86 0 .
87 0 . .
00 0 .
94 0 .
89 0 . .
99 0 .
95 0 .
90 0 . p
5% 0 .
99 0 .
97 0 .
90 0 . .
99 0 .
98 0 .
90 0 . .
98 0 .
99 0 .
91 0 . .
98 0 .
99 0 .
91 0 . The choice of the beta distribution is natural since the support is [0 , R (cid:63) . For example, the Kumaraswamydistribution is another good candidate, but it is close to the beta distribution. When thesupport of the probability distribution is [0 , ∞ ), we apply the truncation formula : G [0 , ( x ) = G ( x ) G (1)For instance, we can use the gamma or log-logistic distribution. However, our experienceshows that some continuous probability distributions are not adapted such as the log-gammaand log-normal distributions, because the logarithm transform performs a bad scale forrandom variables in [0 , R (cid:63) is a logit transformation of arandom variable X ∈ ( −∞ , ∞ ), meaning that : X = logit ( R (cid:63) ) = ln (cid:18) R (cid:63) − R (cid:63) (cid:19) For instance, in the case of the logit-normal distribution, we have:logit ( R (cid:63) ) ∼ N (cid:0) a, b (cid:1) We recall that the return time of the 99% value-at-risk is equal to 100 market days or 100260 ≈ .
38 years. For the probability density function, we have: g [0 , ( x ) = g ( x ) G (1) We also have: R (cid:63) = logit − ( X ) = 11 + e − X G ( x ) = Pr ( R (cid:63) ≤ x )= Pr ( X ≤ logit ( x ))= Φ (cid:18) logit ( x ) − ab (cid:19) and: g ( x ) = 1 bx (1 − x ) φ (cid:18) logit ( x ) − ab (cid:19) A summary of these alternative approaches is given in Table 13. In the sequel, we continueto use the beta distribution, because it is easy to calibrate and it is the most popular approachwhen modeling a random variable in [0 , G ( x ) g ( x ) SupportBeta B ( a, b ) B ( x ; a, b ) x a − (1 − x ) b − B ( a, b ) [0 , G ( a, b ) γ ( a, bx )Γ ( a ) b a x a − e − bx Γ ( a ) [0 , ∞ )Kumaraswamy K ( a, b ) 1 − (1 − x a ) b abx a − (1 − x a ) b − [0 , LL ( a, b ) x b a b + x b b ( x/a ) b − a (cid:16) x/a ) b (cid:17) [0 , ∞ )Logit-normal LN (cid:0) a, b (cid:1) Φ (cid:18) logit ( x ) − ab (cid:19) bx (1 − x ) φ (cid:18) logit ( x ) − ab (cid:19) [0 , As explained previously, the zero-inflated beta model is appealing for producing stress sce-narios. For that, we proceed in two steps. We first calibrate the parameters of the model,and then we compute the stress scenarios for a given return time.
Let Ω = { R , . . . , R n } be the sample of redemption rates for a given matrix cell. Threeparameters have to be estimated: the redemption frequency p and the parameters a and b that control the shape of the beta distribution. We note n as the number of observationsthat are equal to zero and n = n − n as the number of observations that are strictlypositive . In Appendix A.3 on page 81, we show that the maximum likelihood estimatesare: ˆ p = n n + n γ ( α, x ) is the lower incomplete gamma function. We have n = (cid:80) ni =1 { R i = 0 } = n − n and n = (cid:80) ni =1 { R i > } = (cid:80) ni =1 E i . (cid:110) ˆ a, ˆ b (cid:111) = arg max a,b − n ln B ( a, b ) + (cid:88) R i > ( a −
1) ln R i + (cid:88) R i > ( b −
1) ln (1 − R i )The estimates ˆ a and ˆ b can be found by numerical optimization.This is the traditional approach for estimating a zero-inflated model. However, it is notconvenient since the parameters (cid:16) ˆ p, ˆ a, ˆ b (cid:17) should be modified by risk managers and businessexperts before computing redemption shocks. Indeed, the calibration process of parametricstress scenarios follows the same process when one builds historical stress scenarios, andestimated values (cid:16) ˆ p, ˆ a, ˆ b (cid:17) cannot be directly used because they do not necessarily respectsome risk coherency principles and their robustness varies across matrix cells.A second approach consists in using the method of moments. In this case, the estimatorof p has the same expression: ˆ p = n n + n (23)For the parameters of the beta distribution, we first calculate the empirical mean ˆ µ and thestandard deviation ˆ σ of the positive redemption rates R (cid:63) , and then we use the followingrelationships (Roncalli, 2020, page 193):ˆ a = ˆ µ (1 − ˆ µ )ˆ σ − ˆ µ (24)and: ˆ b = ˆ µ (1 − ˆ µ ) σ − (1 − ˆ µ ) (25)The differences between the two methods are the following: • In the case of the method of maximum likelihood, a and b are explicit parameters.Once the parameters p , a and b are estimated, we can calculate the mean µ andstandard deviation σ for the severity distribution. In this approach, µ and σ areimplicit, because they are deduced from a and b . • In the case of the method of moments, a and b are implicit parameters. Indeed, theyare calculated after having estimated the mean µ and standard deviation σ for theseverity distribution. In this approach, µ and σ are explicit and define the severitydistribution.The first approach is known as the p − a − b parameterization, whereas the second approachcorresponds to the p − µ − σ parameterization. By construction, this last approach is moreconvenient in a liquidity stress testing framework, because the parameters µ and σ areintuitive and self-explanatory measures, which is not the case of a and b . Therefore, theycan be manipulated by risk managers and business experts.We have estimated the parameters p , a , b , µ and σ with the two methods. Table 14 showsthe redemption frequency. On average, ˆ p is equal to 31%, but we observe large differencesbetween the matrix cells. For instance, ˆ p is less than 5% for central banks, corporatepension funds and employee savings plans, whereas the largest values of ˆ p are observed forretail investors and third-party distributors. The values of ˆ µ and ˆ σ are reported in Tables15 and 16. The average redemption severity is 0 . . Remark 6
In Tables 42 and 43 on page 97, we have also reported the implicit values of ˆ a and ˆ b that are deduced from ˆ µ and ˆ σ . Moreover, we have reported the estimated values bythe method of maximum likelihood on pages 98 and 99. p in %(1) (2) (3) (4) (5) (6) (7) (8)Auto-consumption 21 .
63 19 .
41 30 .
00 25 .
46 50 .
60 6 .
39 22 . .
16 0 .
34 1 .
47 0 . .
04 6 .
19 6 .
25 2 .
87 39 .
81 0 .
21 14 . .
11 3 .
38 3 .
98 3 .
37 7 .
57 0 .
00 4 . .
67 2 .
83 2 .
97 2 .
71 2 .
29 2 .
75 2 . .
36 6 .
28 1 .
96 6 .
51 32 .
83 1 .
04 8 . .
19 6 .
72 3 .
45 7 .
22 27 .
92 1 .
04 9 . .
67 3 .
87 3 .
68 19 .
35 21 .
52 2 .
22 8 . .
59 45 .
04 58 .
76 70 .
50 45 .
75 17 .
51 27 .
32 45 . .
30 3 .
18 1 .
05 10 .
07 18 .
23 0 .
06 10 . .
77 37 .
36 45 .
97 45 .
94 65 .
94 32 .
86 6 .
52 40 . .
66 27 .
10 24 .
19 38 .
34 37 .
57 11 .
14 24 .
79 31 . µ in % (method of moments)(1) (2) (3) (4) (5) (6) (7) (8)Auto-consumption 1 .
24 1 .
88 2 .
15 1 .
19 3 .
11 2 .
81 1 . .
55 2 .
50 3 .
82 3 . .
54 2 .
84 7 .
26 3 . .
29 2 .
08 2 . .
67 2 .
62 2 .
80 4 .
46 3 . .
36 2 .
20 2 .
19 3 .
21 2 . .
87 2 .
60 1 .
10 3 .
51 0 .
99 2 . .
34 0 .
31 0 .
44 0 .
23 1 .
98 0 .
43 0 .
15 0 . .
06 1 .
84 10 .
48 4 . .
35 0 .
64 1 .
45 0 .
42 1 .
40 0 .
86 1 .
21 0 . .
40 0 .
73 1 .
64 0 .
48 2 .
82 0 .
98 0 .
18 0 . σ in % (method of moments)(1) (2) (3) (4) (5) (6) (7) (8)Auto-consumption 7 .
38 6 .
86 9 .
73 5 .
98 8 .
80 9 .
10 7 . .
55 9 .
57 7 .
49 8 . .
36 13 .
51 13 .
14 12 . .
26 8 .
40 8 . .
46 9 .
99 9 .
23 11 .
46 10 . .
66 10 .
56 10 .
11 8 .
13 9 . .
61 9 .
36 7 .
27 11 .
88 6 .
70 10 . .
80 2 .
58 3 .
32 2 .
10 7 .
52 3 .
22 2 .
64 2 . .
25 9 .
90 21 .
63 14 . .
68 3 .
48 7 .
63 2 .
58 4 .
71 5 .
84 6 .
98 3 . .
31 4 .
35 8 .
93 3 .
50 8 .
66 6 .
08 3 .
03 4 . (1) = balanced, (2) = bond, (3) = enhanced treasury, (4) = equity, (5) = money market, (6) = other, (7)= structured, (8) = total p − µ − σ parameterization Using the previous estimates (ˆ p, ˆ µ, ˆ σ ), risk managers and business experts can define thetriplet ( p, µ, σ ) for the different matrix cells. For that, they must assess the confidence inestimated values with respect to the number of observations. For the frequency parameter,we use the value of n , which has been already reported in Table 8 on page 21. For theseverity parameters ˆ µ and ˆ σ , we use the value of n , which is much smaller than n . Usingthe data given in Table 41 on page 93, we have built the confidence measure in Table 17. Weconfirm that the confidence measure in ˆ µ and ˆ σ is lower than the confidence measure in ˆ p . Inparticular, there are many matrix cells, where the number n of observations is lower than200. This explains why Tables 15 and 16 contain a lot of missing values. Therefore, exceptfor a few matrix cells, the estimated values ˆ µ and ˆ σ must be challenged by risk managersand business experts. Again, they can use risk coherency principles C investor and C fund tobuild their own figures of p , µ and σ .Table 17: Confidence in estimated values ˆ µ and ˆ σ with respect to the number n of obser-vations (1) (2) (3) (4) (5) (6) (7)Auto-consumption • • • • • • • • • • • • ◦ ◦ ◦ Central bank ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦ ◦
Corporate • • ◦ ◦ • • ◦ ◦ ◦ ◦ ◦ ◦
Corporate pension fund ◦ • ◦ ◦ • • ◦ ◦ ◦ ◦ ◦ ◦
Employee savings plan • ◦ ◦ ◦ • ◦ ◦ ◦ ◦ ◦
Institutional • • • • ◦ • • • • ◦ ◦ ◦ ◦
Insurance • • ◦ • • • • ◦ ◦ ◦ ◦
Other • • ◦ • • • • ◦ ◦ ◦
Retail • • • • • • • • • • • • • • • • • •
Sovereign • ◦ ◦ ◦ ◦ • • ◦ ◦ ◦ ◦ ◦ ◦
Third-party distributor • • • • • • • • • • • • • • • • ◦ ◦ ◦ − ◦ ◦ − ◦ − • − • • −
10 000, • • • +10 000
Once the triplet ( p, µ, σ ) is defined for each matrix cell, we compute stress scenarios usingthe following formula: S ( T ; p, µ, σ ) = B − (cid:32) − p T ; µ (1 − µ ) σ − µ, µ (1 − µ ) σ − (1 − µ ) (cid:33) where B − ( α ; a, b ) is the α -quantile of the beta distribution with parameters a and b . Theparametric stress scenario S ( T ; p, µ, σ ) depends on the return time T and the three parame-ters of the zero-inflated model. An example is provided in Figure 16. For each plot, we indi-cate the triplet ( p, µ, σ ). For instance, the first plot corresponds to the triplet (2% , , σ , which is the key parameter when computing parametric stress scenarios. The reasonis that the parameters p and µ determine the mean E [ R ], whereas the uncertainty aroundthis number is mainly driven by the parameter σ . The redemption volatility controls thenthe shape of the probability distribution of the redemption rate (both the skewness and thekurtosis), implying that σ has a major impact on the stress scenario S ( T ) when T is large. They are defined on page 20. S ( T ; p, µ, σ ) in % In this section, we go beyond the zero-inflated model by considering the behavior of eachinvestor. In particular, we show that the redemption rate depends on the liability structureof the mutual fund. Moreover, the behavior of investors may be correlated, in particular ina stress period. In this case, the modeling of spillover effects is important to define stressscenarios.
The individual-based model and the zero-inflated model are highly connected. Indeed, thezero-inflated model can be seen as a special case of the individual-based model when wesummarize the behavior of all unitholders by the behavior of a single client.
Let TNA ( t ) be the assets under management of an investment fund composed of n clients:TNA ( t ) = n (cid:88) i =1 TNA i ( t )where TNA i ( t ) is the net asset of the individual i . The redemption rate of the fund is equalto the redemption flows divided by the total net assets: R = (cid:80) ni =1 TNA i · R i TNA= n (cid:88) i =1 ω i · R i ω i represents the weight of the client i in the fund: ω i = TNA i TNASince we have R i = E i · R (cid:63)i , we obtain: R = n (cid:88) i =1 ω i · E i · R (cid:63)i Generally, we assume that the clients are homogenous, meaning that E i and R (cid:63)i are iid random variables. If we denote by ˜ p and ˜G the individual redemption probability and theindividual severity distribution. The individual-based model is then characterized by the4-tuple (cid:16) n, ω, ˜ p, ˜G (cid:17) , where n is the number of clients and ω = ( ω , . . . , ω n ) is the vector ofweights. Like the zero-inflated model, we consider a ˜ µ − ˜ σ parameterization of ˜G , meaningthat the model is denoted by IM ( n, ω, ˜ p, ˜ µ, ˜ σ ). Remark 7
When the individual severity distribution ˜G is no specified, we assume that itis a beta distribution B (cid:16) ˜ a, ˜ b (cid:17) , whose parameters ˜ a and ˜ b are calibrated with respect to theseverity mean ˜ µ and the severity volatility ˜ σ using the method of moments. In a similarway, we assume that the vector of weights is equally-weighted when it is not specified. Inthis case, the individual-based model is denoted by IM ( n, ˜ p, ˜ µ, ˜ σ ) . Figure 17: Histogram of the redemption rate in % (˜ p = 50% , ˜ µ = 50% , ˜ σ = 10%)In Figure 17, we report the histogram of the redemption rate R in the case ˜ p = 50%,˜ µ = 50% and ˜ σ = 10%. In the case n = 1, we obtain a singular distribution. Indeed,there is a probability of 50% that there is no redemption. The singularity decreases withrespect to the number n of investors, because the probability to have a redemption increases.39iquidity Stress Testing in Asset ManagementThis is normal since the redemption frequency of a mutual fund depends on the number ofunitholders. This explains that the redemption frequency is larger for a retail fund than foran institutional fund. The singularity of the distribution function F at the point R = 0is entirely explained by the two parameters ˜ p and n as shown in Appendix A.4.1 on page83, because we have: Pr { R = 0 } = (1 − ˜ p ) n The fact that the probability distribution is not continuous has an impact on the skewnessand the kurtosis. In Table 18, we have reported the probability Pr { R = 0 } . If there is afew investors in the fund, the probability to observe no redemption in the fund is high. Forinstance, if ˜ p = 5% and n = 10, we have Pr { R = 0 } = 59 . p = 1% and n = 10,we have Pr { R = 0 } = 90 . p is the individualredemption probability, 1 / ˜ p is the return time of a redemption at the investor level. Forexample, ˜ p = 5% (resp. ˜ p = 1%) means that investors redeem every 20 days (resp. 100days) on average. At the fund level, the return time to observe a redemption is equal to(1 − Pr { R = 0 } ) − . For instance, in the case ˜ p = 5% and n = 10, we observe a redemptiontwo days per week in the fund on average . This analysis may help to distinguish activeand passive investors. In the case of passive investors when the redemption event occursonce a year or less, ˜ p is less than 40 bps. In the case of active investors that redeem once amonth, ˜ p is greater than 5%. Therefore, the skewness effect depends if the fund has activeinvestors or not, and if the fund is granular or not.Table 18: Probability to observe no redemption Pr { R = 0 } in % p Number n of investors(in %) 1 2 5 10 50 100 1000 100000 .
01 99 .
99 99 .
98 99 .
95 99 .
90 99 .
50 99 .
00 90 .
48 36 . .
02 99 .
98 99 .
96 99 .
90 99 .
80 99 .
00 98 .
02 81 .
87 13 . .
05 99 .
95 99 .
90 99 .
75 99 .
50 97 .
53 95 .
12 60 .
65 0 . .
10 99 .
90 99 .
80 99 .
50 99 .
00 95 .
12 90 .
48 36 .
77 0 . .
20 99 .
80 99 .
60 99 .
00 98 .
02 90 .
47 81 .
86 13 .
51 0 . .
50 99 .
50 99 .
00 97 .
52 95 .
11 77 .
83 60 .
58 0 .
67 0 . .
00 99 .
00 98 .
01 95 .
10 90 .
44 60 .
50 36 .
60 0 .
00 0 . .
00 98 .
00 96 .
04 90 .
39 81 .
71 36 .
42 13 .
26 0 .
00 0 . .
00 95 .
00 90 .
25 77 .
38 59 .
87 7 .
69 0 .
59 0 .
00 0 . .
00 90 .
00 81 .
00 59 .
05 34 .
87 0 .
52 0 .
00 0 .
00 0 . .
00 75 .
00 56 .
25 23 .
73 5 .
63 0 .
00 0 .
00 0 .
00 0 . .
00 50 .
00 25 .
00 3 .
13 0 .
10 0 .
00 0 .
00 0 .
00 0 . The mean effect
The mean shape is easy to understand since it is the product of theredemption probability and the mean of the redemption severity: E [ R ] = ˜ p ˜ µ Curiously, it depends neither on the number of investors in the fund, nor on the liabilitystructure (see Figure 18). Since ˜ µ ∈ [0 , E [ R ] ≤ ˜ p , meaning that we must The exact value is equal to 1 / (1 − . . given in Table 4 on page 18. If we consider all investor and fund categories,the mean is equal to 22 bps. The largest value is observed for the sovereign/money marketcategory and is equal to 1 . R in % The volatility effect
By assuming that the liability weights are equal ( ω i = 1 /n ), thevolatility of the redemption rate is equal to: σ ( R ) = ˜ p (cid:0) ˜ σ + (1 − ˜ p ) ˜ µ (cid:1) n Globally, we observe that σ ( R ) is an increasing function of ˜ p , ˜ µ and ˜ σ as shown in Figure19. When the redemption probability increases, we observe a convexity shape because wehave: σ ( R ) = ˜ p (cid:0) ˜ σ + ˜ µ (cid:1) n − ˜ p ˜ µ n However, this is not realistic since ˜ p ≤
20% in practice. Another interesting property isthat σ ( R ) tends to zero when the number of investors in the fund increases (Figure 20).If we compute the volatility of the redemption rate, we obtain the figures given in Table 48on page 100. We observe that σ ( R ) (cid:29) E [ R ], implying that R is a high-skewed randomvariable. This challenges the use of the SD ( c ) measure presented on page 16. Correspondence between zero-inflated and individual-based models
We noticethat the zero-inflated model ZI ( p, µ, σ ) is a special case of the individual-based model by Another way to compute the empirical mean of R is to calculate the product of the aggregate redemptionfrequency p (Table 14 on page 36) and the aggregate severity mean (Table 15 on page 36). R in % ( n = 10) Figure 20: Volatility of the redemption rate R in % ( p = 10%, µ = 50%, σ = 30%) p − µ − σ parameterization is not obvious,because ZI ( p, µ, σ ) is an aggregate population model.In this paragraph, we would like to find the relationships between the parameters of thezero-inflated model and those of the individual-based model, such that the two models arestatistically equivalent: ZI ( p, µ, σ ) ≡ IM ( n, ω, ˜ p, ˜ µ, ˜ σ )There are different approaches. A first one is to minimize the Kolmogorov-Smirnov statisticsbetween ZI ( p, µ, σ ) and IM ( n, ω, ˜ p, ˜ µ, ˜ σ ). Another approach consists in matching theirmoments. We consider the second approach because we obtain analytical formulas, whereasthe solution of the first approach can only be numerical. In Appendix A.5 on page 85, weshow that: p = 1 − (1 − ˜ p ) n and µ = ˜ p − (1 − ˜ p ) n ˜ µ whereas the parameter σ satisfies the following relationship: σ = (cid:18) ˜ p H ( ω )1 − (1 − ˜ p ) n (cid:19) ˜ σ + (cid:32) ˜ p ((1 − ˜ p ) − (1 − ˜ p ) n ) H ( ω ) − ˜ p (1 − ˜ p ) n (1 − H ( ω ))(1 − (1 − ˜ p ) n ) (cid:33) ˜ µ where H ( ω ) = (cid:80) ni =1 ω i is the Herfindahl index. It is interesting to notice that p is a functionof n and ˜ p , µ is a function of n , ˜ p and ˜ µ , but σ does not only depends on the parameters n ,˜ p , ˜ µ and ˜ σ : p = ϕ ( n, ˜ p ) µ = ϕ ( n, ˜ p, ˜ µ ) σ = ϕ ( n, ˜ p, ˜ µ, ˜ σ, H ( ω ))Indeed, the aggregate severity volatility also depends on the Herfindahl index of the fundliability structure. Remark 8
The previous relationships can be inverted in order to define the parameters ofthe individual-based model with respect to the parameters of the zero-inflated model: ˜ p = ϕ (cid:48) ( p ; n )˜ µ = ϕ (cid:48) ( p, µ ; n )˜ σ = ϕ (cid:48) ( p, µ, σ ; n, H ( ω )) However, we notice that there are two degrees of freedom – n and H ( ω ) – that must be fixed. In Tables 19 and 20, we report some examples of calibration when n is equal to 10and ω i is equal to 10%. For instance, if the parameters of the individual-based model are˜ p = 1 . µ = 50% and ˜ σ = 10%, we obtain p = 9 . µ = 5 .
23% and σ = 1 .
48% for thezero-inflated model. If we know the weights of the investors in the investment fund, we cantherefore calibrate the zero-inflated model from the individual-based model (Table 19), butalso the individual-based model from the zero-inflated model (Table 20).43iquidity Stress Testing in Asset ManagementTable 19: Calibration of the zero-inflated model from the individual-based modelParameter IM ( n, ˜ p, ˜ µ, ˜ σ ) ZI ( p, µ, σ )set ˜ p ˜ µ ˜ σ p µ σ .
20% 50 .
00% 10 .
00% 1 .
98% 5 .
05% 1 . .
00% 50 .
00% 10 .
00% 9 .
56% 5 .
23% 1 . .
00% 30 .
00% 20 .
00% 9 .
56% 3 .
14% 2 . ZI ( p, µ, σ ) IM ( n, ˜ p, ˜ µ, ˜ σ )set p µ σ ˜ p ˜ µ ˜ σ .
00% 2 .
00% 5 .
00% 0 .
51% 19 .
55% 49 . .
00% 2 .
00% 5 .
00% 1 .
05% 19 .
08% 48 . .
00% 5 .
00% 10 .
00% 1 .
05% 47 .
71% 97 . We notice that the variance of the redemption rate depends on the Herfindahl index: H ( ω ) = n (cid:88) i =1 ω i This implies that the liability structure ω is an important parameter to understand theprobability distribution of the redemption rate. The arithmetics of the Herfindahl index
We know that the Herfindahl index isbounded: 1 n ≤ H ( ω ) ≤ H ( ω ) is equal to one when one investor holds 100% of the investment fund ( ∃ i : ω i = 1),whereas the lower bound is reached for an equally-weighted liability structure ( ω i = n − ).Therefore, H ( ω ) is a measure of concentration. A related statistic is the “ effective numberof unitholders ”: N ( ω ) = 1 H ( ω ) N ( ω ) indicates how many equivalent investors hold the investment fund. For instance, weconsider two funds with the following liability structures ω (1) = (25% , , , ω (2) = (42% , , , , , , N (cid:0) ω (1) (cid:1) = 4 and N (cid:0) ω (2) (cid:1) = 3 . ω i ∝ q i and 0 < q < : N ( ω ) = 1 − q (1 − q ) Because we have: H ( ω ) = (1 − q ) q ∞ (cid:88) i =1 q i = (1 − q ) q q (1 − q ) = (1 − q ) − q q ≤ .
98, then N ( ω ) < ω i ∝ q i Approximation of the probability distribution ˜F ( x | ω ) We recall that the uncondi-tional probability distribution of the redemption rate is given by F ( x ) = Pr { R ≤ x } . Sincethe redemption rate depends on the liability structure ω in the individual-based model IM ( n, ω, ˜ p, ˜ µ, ˜ σ ), we note ˜F ( x | ω ) the associated probability distribution: ˜F ( x | ω ) = Pr (cid:40) n (cid:88) i =1 ω i · E i · R (cid:63)i ≤ x (cid:41) We now consider the model IM ( N ( ω ) , ˜ p, ˜ µ, ˜ σ ) and define ˜F ( x | H ) as follows: ˜F ( x | H ) = Pr N ( ω ) N ( ω ) (cid:88) i =1 E i · R (cid:63)i ≤ x = Pr H ( ω ) H ( ω ) − (cid:88) i =1 E i · R (cid:63)i ≤ x The issue is to know under which conditions we can approximate ˜F ( x | ω ) by ˜F ( x | H ).Let us consider some Monte Carlo experimentations. We assume that the liability weightsare geometric distributed: ω i ∝ q i . In Figure 22, we compare the two probability distribu-tions ˜F ( x | ω ) and ˜F ( x | H ) for several sets of parameters (˜ p, ˜ µ, ˜ σ ). The weights ω i for We recall that ˜G is the beta distribution by default. q = 0 .
95 are given in Figure 42 on page 100. We notice that the approximation of ˜F ( x | ω )by ˜F ( x | H ) is good and satisfies the Kolmogorov-Smirnov test at the 99% confidence level.This is not the case if we assume that q = 0 .
90 or q = 0 .
50 (see Figures 43 and 44 on page101). Figure 22: Comparison of ˜F ( x | ω ) and ˜F ( x | H ) ( q = 0 .
95 and H ( ω ) − = 38) To better understand these results, we assume that ˜ p = 0 .
3, ˜ µ = 0 . σ = 0 .
4. When q is equal to 0 .
50, the effective number of unitholders is low and is equal to 3. In this case,the probability distribution ˜F ( x | H ) is far from the probability distribution ˜F ( x | ω ) asshown in Figure 23. In fact, this case corresponds to an investment fund which is highlyconcentrated. The risk is then to observe redemptions from the largest unitholders. Inparticular, we notice that ˜F ( x | H ) presents some steps. The reason is that the redemptionrate can be explained by the redemption of one unitholder, the redemption of two unitholdersor the redemption of three unitholders. If we assume that q is equal to 0 .
90, the effectivenumber of unitholders is larger and becomes 38. In this case, the probability distribution ˜F ( x | H ) is close to the probability distribution ˜F ( x | ω ), because the step effects disappear(see Figure 24). To summarize, the approximation of ˜F ( x | ω ) by ˜F ( x | H ) cannot be goodwhen the effective number of unitholders (or H ( ω ) − ) is low. Remark 9
In many cases, we don’t know the comprehensive liability structure ω , but onlythe largest weights. In Appendix A.6 on page 86, we derive an upper bound H + m ( ω ) of theHerfindahl index H ( ω ) from the m largest weights. Therefore, we can deduce a lower boundof the effective number of unitholders: N ( ω ) > N − m ( ω ) = 1 H + m ( ω ) An example is provided in Table 21 when we assume that ω i ∝ q i . When the fund ishighly concentrated, we obtain a good approximation of N ( ω ) with the th or th largest H ( ω ) − = 3 Figure 24: The case H ( ω ) − = 18 unitholders. Otherwise, N ( ω ) is underestimated. However, this is not a real issue becausewe can think that generated stress scenarios will be overestimated. Indeed, using a lowervalue N ( ω ) increases σ ( R ) as shown in Figure 20 on page 42, implying that the redemptionrisk is generally overestimated. Table 21: Lower bound N − m ( ω ) of the effective number of unitholders m q = 0 . q = 0 . q = 0 . q = 0 . q = 0 . q = 0 . ∞ Stress scenarios based on the largest unitholders
The previous results show thatthe main risk in a concentrated fund comes from the behavior of the largest unitholders. Itjustifies the use of stress scenarios based on the order statistics ω k : n :min ω i = ω n ≤ · · · ≤ ω k : n ≤ ω k +1: n ≤ · · · ≤ ω n : n = max ω i Then, we can define the stress scenario that corresponds to the full redemption of the m largest unitholders: S ( m ) = m (cid:88) k =1 ω n − k +1: n An example is given in Table 22 when the liability structure is defined by ω i ∝ q i . Of course,these stress scenarios S ( m ) make sense only if the fund presents some liability concentration.Otherwise, they are not informative.Table 22: Stress scenarios S ( m ) when ω i ∝ q i m q = 0 . q = 0 . q = 0 . q = 0 . q = 0 . q = 0 . .
0% 10 .
0% 5 .
0% 3 .
0% 1 .
0% 0 . .
0% 19 .
0% 9 .
8% 5 .
9% 2 .
0% 1 . .
9% 41 .
0% 22 .
6% 14 .
1% 4 .
9% 2 . .
9% 65 .
1% 40 .
1% 26 .
3% 9 .
6% 4 . The calibration of the individual-based model ismuch more complicated than the calibration of the zero-inflated model. The reason is thatit depends on the liability structure of the funds, which are not necessarily the same for thedifferent funds. Let us consider the case of a single fund f . We can estimate the parameters˜ p , ˜ µ and ˆ σ using the quadratic criterion: { ˜ p (cid:63) , ˜ µ (cid:63) , ˜ σ (cid:63) } = arg min (cid:36) ˜ p (cid:16) ˆ p ( f ) − − ˜ p ) H − f ) (cid:17) + (cid:36) ˜ µ (cid:0) ˆ p ( f ) ˆ µ ( f ) − ˜ p ˜ µ (cid:1) + (cid:36) ˜ σ (cid:16) ˆ p ( f ) (cid:16) ˆ σ f ) + (cid:0) − ˆ p ( f ) (cid:1) ˆ µ f ) (cid:17) − ˜ p (cid:0) ˜ σ + (1 − ˜ p ) ˜ µ (cid:1) H ( f ) (cid:17) (26)48iquidity Stress Testing in Asset Managementwhere ˆ p ( f ) , ˆ µ ( f ) and ˆ σ ( f ) are the empirical estimates of the parameters p , µ and σ , and H ( f ) is the Herfindahl index associated with the fund. In practice, the liability structure changesevery day, meaning that the Herfindahl index is time-varying. Therefore, we can use theaverage of Herfindahl indices. The weights (cid:36) ˜ p , (cid:36) ˜ µ and (cid:36) ˜ σ indicate the relative importanceof each moment condition. If we consider several funds, the previous criterion becomes: { ˜ p (cid:63) , ˜ µ (cid:63) , ˜ σ (cid:63) } = arg min (cid:36) ˜ p (cid:88) f (cid:36) ( f ) (cid:16) ˆ p ( f ) − − ˜ p ) H − f ) (cid:17) + (cid:36) ˜ µ (cid:88) f (cid:36) ( f ) (cid:0) ˆ p ( f ) ˆ µ ( f ) − ˜ p ˜ µ (cid:1) + (cid:36) ˜ σ (cid:88) f (cid:36) ( f ) (cid:16) ˆ p ( f ) (cid:16) ˆ σ f ) + (cid:0) − ˆ p ( f ) (cid:1) ˆ µ f ) (cid:17) − ˜ p (cid:0) ˜ σ + (1 − ˜ p ) ˜ µ (cid:1) H ( f ) (cid:17) (27)where (cid:36) ( f ) is the relative weight of the fund f .In practice, the estimates ˜ p (cid:63) , ˜ µ (cid:63) and ˜ σ (cid:63) are very sensitive to the Herfindahl index be-cause of the first and third moment conditions. To illustrate this point, we consider theinstitutional category and we assume that there is only one fund. On page 36, we foundthat ˆ p ( f ) = 8 . µ ( f ) = 3 .
23% and ˆ σ ( f ) = 10 . H ( f ) = 5, we obtain ˜ p (cid:63) = 1 . µ (cid:63) = 15 .
61% and ˜ σ (cid:63) = 53 . H ( f ) = 20, we obtain ˜ p (cid:63) = 0 . µ (cid:63) = 62 .
04% and˜ σ (cid:63) = 212 . p ( f ) = 45 . µ ( f ) = 0 . σ ( f ) = 2 . H ( f ) = 1 000, we obtain ˜ p (cid:63) = 0 . µ (cid:63) = 247% and ˜ σ (cid:63) = 2 489%. If H ( f ) = 10 000, we obtain ˜ p (cid:63) = 0 . µ (cid:63) = 2 472% and ˜ σ (cid:63) = 24 891%. These results are notrealistic since ˜ µ (cid:63) > σ (cid:63) > Using mandates and dedicated funds
Collective investment and mutual funds arepooled investment vehicles, meaning that they are held by several investors. We now consideranother type of funds with a single unitholder. They correspond to mandates and funds thatare dedicated to a unique investor. In this case, the Herfindahl index is equal to one, andthe solution of Problem (26) corresponds to the parameter set of the zero-inflated model: (cid:8) ˜ p (cid:63) = ˆ p ( f ) , ˜ µ (cid:63) = ˆ µ ( f ) , ˜ σ (cid:63) = ˆ σ ( f ) (cid:9) In our database, we can separate the observations between collective and mutual fundson one side and mandates and dedicated funds on the other side. In Tables 23, 24 and 25,we have estimated the parameters ˜ p , ˜ µ and ˜ σ by only considering mandates and dedicatedfunds. These results highly differ than those obtained for collective and mutual funds (Tables14, 15 and 16 on page 36). First, we can calibrate a smaller number of cells. Indeed, werecall that the estimates are not calculated if the number of observations is less than 200.Second, the magnitude of the estimates is very different. If we consider all fund and investorcategories, we obtain ˜ p = 3 . µ = 2 .
13% and ˜ σ = 10 . p = 31 . µ = 0 .
72% and σ = 4 . p (cid:28) p and ˜ σ (cid:29) ˜ σ because of the following reasons: • the redemption probability is larger in a collective fund than in a dedicated fundbecause they are several investors; • the volatility of the redemption severity is smaller in a collective fund than in a ded-icated fund because the behavior of the different investors is averaged, implying thatthe dispersion of redemption is reduced.49iquidity Stress Testing in Asset ManagementTable 23: Estimated value of ˜ p in %(1) (2) (3) (4) (5) (6) (7) (8)Central bank 0 .
13 0 .
21 0 .
73 2 .
99 0 . .
49 1 .
14 0 .
13 0 .
57 0 . .
16 1 .
40 1 .
60 3 .
06 0 .
41 0 .
47 1 . .
47 1 .
35 0 .
41 2 .
13 1 .
65 0 .
40 0 .
00 1 . .
09 2 .
12 1 .
52 0 .
59 0 .
13 1 . .
23 0 .
44 0 .
35 0 .
16 0 .
03 0 . .
71 8 .
07 3 .
46 25 .
40 11 .
68 7 .
17 14 . .
95 2 .
63 1 .
73 5 .
82 2 .
92 0 .
68 7 .
46 3 . µ in %(1) (2) (3) (4) (5) (6) (7) (8)Central bankCorporateCorporate pension fund 4 .
39 2 .
94 4 . .
88 4 .
05 3 .
29 4 . .
46 4 . .
77 1 .
52 0 .
44 0 . .
89 2 .
47 1 .
48 2 .
64 3 .
91 2 . σ in %(1) (2) (3) (4) (5) (6) (7) (8)Central bankCorporateCorporate pension fund 15 .
37 10 .
65 14 . .
42 13 .
88 12 .
30 14 . .
01 14 . .
28 5 .
84 3 .
29 4 . .
61 10 .
35 8 .
29 10 .
39 15 .
06 10 . (1) = balanced, (2) = bond, (3) = enhanced treasury, (4) = equity, (5) = money market, (6) = other, (7)= structured, (8) = total Curiously, we do not observe that ˜ µ ≈ µ . One explanation may be that investors in mandatesare not the same as investors in collective funds. Indeed, we may consider that they aremore sophisticated and bigger when they are able to put in place a mandate or a dedicatedfund. For instance, they can be more active.The results obtained with data from mandates and dedicated funds are more realisticthan those obtained with data from collective and mutual funds. The drawback is that theyare based on a smaller number of observations and there are many cells where we don’thave enough observations for computing the estimates. Therefore, these estimates must becompleted by expert judgements. 50iquidity Stress Testing in Asset Management Computing the stress scenarios
Once we have estimated the parameters ˜ p , ˜ µ and ˜ σ ,we can compute the stress scenarios using the Monte Carlo method. Nevertheless, we facean issue here, because the stress scenario is not unique to an investor category. Indeed, itdepends on the liability structure of the fund. While the individual-based model is morerealistic and relevant than the zero-inflated model, then it appears to be limited from apractical point of view. Nevertheless, it is useful to understand the importance of theliability structure on the redemption rate. We now consider an extension of the previous model by introducing correlations betweenthe investors. We obtain the same expression of the redemption rate: R = n (cid:88) i =1 ω i · E i · R (cid:63)i However, the random variables ( E , . . . , E n , R (cid:63) , . . . , R (cid:63)n ) are not necessarily independent. Wediscuss three different correlation patterns:1. We can assume that E i and E j are correlated. This is the simplest and most un-derstandable case. Indeed, we generally observe long periods with low redemptionfrequencies followed by short periods with high redemption frequencies, in particularwhen there is a crisis or a panic.2. We can assume that the redemption severities R (cid:63)i and R (cid:63)j are correlated. For example,it would mean that high (resp. low) redemptions are observed at the same time. Nev-ertheless, this severity correlation is different from the previous frequency correlation.Indeed, the severities are independent from the number of redemptions, implying thatthe severity correlation only concerns the unitholders that have already decided toredeem.3. We can assume that E i and R (cid:63)i are correlated. We notice that we can write theredemption rate for a given category as follows: R = n (cid:88) i =1 ω i · R i where R i = E i · R (cid:63)i is the individual redemption rate for the i th investor. The breakdownbetween the binary variable E i and the continuous variable R (cid:63)i helps us to modelthe “ clumping-at-zero ” pattern. But there is no reason that the value taken by theredemption severity R (cid:63)i depends whether E i takes the value 0 or 1, because R (cid:63)i isobserved only if E i = 1.Finally, only the first two correlation patterns are relevant from a financial point of view,because the third correlation model has no statistical meaning. Nevertheless, it is obviousthat the first correlation model is more appropriate because the second correlation modelconfuses low-severity and high-severity regimes. During a liquidity crisis, both the redemp-tion frequency and the redemption severity increase (Coval and Stafford, 2007; Duarte andEisenbach, 2013; Kacperczyk and Schnabl, 2013; Roncalli and Weisang, 2015a; Schmidt etal. , 2016). The first effect may be obtained by stressing the parameter ˜ p or by consider-ing a high-frequency regime deduced from the first correlation model, but the second effect51iquidity Stress Testing in Asset Managementcan only be obtained by stressing the parameter ˜ µ and cannot be explained by the secondcorrelation model. Therefore, we only consider the first correlation pattern by modelingthe random vector ( E , . . . , E n ) with a copula decomposition. We continue to assume that E i ∼ B (˜ p ) are identically distributed, but the dependence function of ( E , . . . , E n ) is givenby the copula function C ( u , . . . , u n ). The individual-based model is then a special case ofthis copula-based model when the copula function is the product copula C ⊥ .In what follows, we consider the Clayton copula : C ( θ c ) ( u , . . . , u n ) = (cid:16) u − θ c + · · · + u − θ c n − n + 1 (cid:17) − /θ c or the Normal copula: C ( θ c ) ( u , . . . , u n ) = Φ (cid:0) Φ − ( u ) + · · · + Φ − ( u n ) ; C n ( θ c ) (cid:1) The Clayton parameter satisfies θ c ≥ θ c lies in the range[ − , C (0) = C ⊥ ≺ C ( θ c ) ≺ C + = C ( ∞ ) meaning that the product copula is reached when θ c = 0 and the upper Fr´echet boundcorresponds to the limiting case θ c → ∞ . For the Normal copula, we restrict our analysisto θ c ∈ [0 ,
1] because there is no sense to obtain negative correlations. Therefore, we have: C (0) = C ⊥ ≺ C ( θ c ) ≺ C + = C (1) The Normal parameter θ c is easy to interpret because it corresponds to the Pearsonlinear correlation between two Gaussian random variables. The interpretation of the Claytoncopula θ c is more tricky. Nevertheless, we can compute the associated Kendall’s tau andSpearman’s rho rank correlations . Their expressions are given in Table 26. Therefore, wecan deduce the Pearson correlation ρ .Table 26: Relationship between the copula parameter θ c , the Kendall’s tau τ , the Spearman’srho (cid:37) and the Pearson correlation ρτ (cid:37) ρ Clayton θ c θ c + 2 sin (cid:18) πθ c θ c + 4 (cid:19) sin (cid:18) πθ c θ c + 4 (cid:19) Normal 2 π arcsin ( θ c ) 6 π arcsin (cid:18) θ c (cid:19) θ c The previous formulas can be used to map the copula parameter space into the Kendall,Spearman or Pearson correlation space. Some numeric values are given in Table 27. Forexample, the Clayton copula θ c = 2 corresponds to a Kendall’s tau of 50%, a Spearman’s rhoof 69% and a Pearson correlation of 70 . We use the notations of Roncalli (2020, Chapter 11). For the Clayton copula, we calculate an approximation of the Spearman’s rho: (cid:37) ≈ π arcsin (cid:18)
12 sin (cid:18) πθ c θ c + 4 (cid:19)(cid:19) ≈ sin (cid:18) πθ c θ c + 4 (cid:19) θ c τ (cid:37) ρ θ c τ (cid:37) ρ .
00 0 .
00% 0 .
00% 0 .
00% 0 .
00 0 .
00% 0 .
00% 0 . .
00 33 .
33% 48 .
26% 50 .
00% 0 .
20 12 .
82% 19 .
13% 20 . .
00 50 .
00% 69 .
02% 70 .
71% 0 .
50 33 .
33% 48 .
26% 50 . .
00 71 .
43% 89 .
25% 90 .
10% 0 .
75 53 .
99% 73 .
41% 75 . .
00 83 .
33% 96 .
26% 96 .
59% 0 .
90 71 .
29% 89 .
15% 90 . .
00 96 .
15% 99 .
80% 99 .
82% 0 .
99 90 .
99% 98 .
90% 99 . Remark 10
We denote the copula-based model by CM ( n, ω, ˜ p, ˜ µ, ˜ σ, ρ ) (or CM ( n, ˜ p, ˜ µ, ˜ σ, ρ ) when the vector of weights are equally-weighted). We have the following equivalence: IM ( n, ω, ˜ p, ˜ µ, ˜ σ ) = CM ( n, ω, ˜ p, ˜ µ, ˜ σ, In Appendix A.7.1 on page 87, we show that:Pr { R = 0 } = C ( θ c ) (1 − ˜ p, . . . , − ˜ p )Since C ⊥ ≺ C ( θ c ) ≺ C + , we obtain the following bounds :(1 − ˜ p ) n ≤ Pr { R = 0 } ≤ − ˜ p We notice that the probability to observe zero redemptions converges to zero only when thenumber n of unitholders tends to ∞ and the copula is the product copula. By assuming thatthe redemption frequency ˜ p is equal to 10%, we obtain the results given in Figure 45 on page102 and we verify the previous statistical property. In Figure 25, we show the relationshipbetween the Pearson correlation ρ and the probability Pr { R = 0 } . As expected, it is anincreasing function. We notice that the introduction of the correlation is very importantto understand the empirical results we have calculated in Table 14 on page 36 and someunrealistic values we have obtained in Table 18 on page 40. For instance, the fact thatPr { R = 0 } is equal to 54 .
39% for the retail category can only be explained by a significantfrequency correlation since the number n of unitholders is high for this category.By construction, the frequency correlation modifies the probability distribution of theredemption frequency F , which is defined as the proportion of unitholders that redeem: F = 1 n n (cid:88) i =1 {E i = 1 } F is a random variable, whose range is between 0 and 1. F depends on the frequencyparameter ˜ p , the number n of unitholders and the copula function C ( θ c ) (or the Pearson Because we have: C ⊥ (1 − ˜ p, . . . , − ˜ p ) = n (cid:89) i =1 (1 − ˜ p ) = (1 − ˜ p ) n and: C + (1 − ˜ p, . . . , − ˜ p ) = min (1 − ˜ p, . . . , − ˜ p ) = 1 − ˜ p { R = 0 } in % with respect to the fre-quency correlation ρ (˜ p = 10%) Figure 26: Redemption frequencies in % with respect to the frequency correlation ρ (˜ p =20% , n = 1 000) 54iquidity Stress Testing in Asset Managementcorrelation ρ ). When C ( θ c ) is the product copula C ⊥ , the redemption events are independentand we obtain: F ∼ B ( n, ˜ p ) n because the sum of independent Bernoulli random variables is a binomial random variable.Therefore, we obtain the following approximation when n is sufficiently large: B ( n, ˜ p ) n ≈ N ( n ˜ p, n ˜ p (1 − ˜ p )) n = N (cid:18) ˜ p, ˜ p (1 − ˜ p ) n (cid:19) When the copula C ( θ c ) corresponds to the upper Fr´echet bound C + , the redemption fre-quency follows the Bernoulli distribution and does not depend on the number of unitholders: F ∼ B (˜ p )We have represented these two extreme cases in Figure 26 when ˜ p = 20% and n = 1 000.We have also reported the probability distribution of F when the Pearson correlation of thecopula function is equal to 25% and 50%. We notice that the skewness risk increases withthe frequency correlation. Therefore, the parameter ρ will have a high impact on the stresstesting results. In particular, when the frequency correlation is high, the risk is to observea large proportion of redemptions even if the number of unitholders is large. In this case,the diversification effect across unitholders is limited. An illustration is provided in Figure46 on page 102 that shows the probability to observe 100% of redemptions when n is setto 20. The mean effect
In Appendix A.7.3 on page 88, we show that the frequency correlationhas no impact on the average redemption rate since we obtain the same expression aspreviously: E [ R ] = ˜ p ˜ µ Therefore, the redemption frequency changes the shape of the probability distribution of R ,but not its mean. The volatility effect
The volatility of the redemption rate is equal to: σ ( R ) = (cid:16) ˜ p ˜ σ + (cid:16) ˜ p − ˘C ( θ c ) (˜ p, ˜ p ) (cid:17) ˜ µ (cid:17) H ( ω ) + (cid:16) ˘C ( θ c ) (˜ p, ˜ p ) − ˜ p (cid:17) ˜ µ where ˘C ( θ c ) is the survival copula associated to C ( θ c ) . Since we have C (cid:62) ≺ C ( θ c ) ≺ C + , weobtain the following inequalities:˜ p (cid:0) ˜ σ + (1 − ˜ p ) ˜ µ (cid:1) H ( ω ) ≤ σ ( R ) ≤ ˜ p ˜ σ H ( ω ) + ˜ p (1 − ˜ p ) ˜ µ If we consider the equally-weighted case and assume that n tends to infinity, we obtain:0 ≤ σ ( R ) = (cid:16) ˘C ( θ c ) (˜ p, ˜ p ) − ˜ p (cid:17) ˜ µ ≤ ˜ p (1 − ˜ p ) ˜ µ This implies that the volatility risk is not equal to zero for an infinitely fine-grained liabilitystructure if the frequency correlation is different from zero. It corresponds to the statistic Pr { F = 1 } . R in % with respect to the number n of unithold-ers (˜ p = 10% , ˜ µ = 50% , ˜ σ = 30%) Figure 28: Volatility of the redemption rate R in % with respect to the frequency correlation(˜ p = 10% , ˜ µ = 50% , ˜ σ = 10%) ρ . Thesefigures confirm that the volatility risk is minimum when the frequency correlation is equalto zero. The consequence is that the frequency correlation is a key parameter when build-ing stress testing scenarios. This is perfectly normal since ρ can been seen as a parameterthat controls spillover effects and the magnitude of redemption contagion. All these resultscorroborate the previous intuition that the individual-based model without redemption cor-relation may be not appropriate for building a robust stress testing program. The shape effect
The impact of the frequency correlation on the skewness and the volatil-ity can then change dramatically the shape of the probability distribution of the redemptionrate. In Figure 17 on page 39, we have already studied the histogram of the redemptionrate in the case ˜ p = 50%, ˜ µ = 50% and ˜ σ = 10%. Let us reproduce the same exercise byassuming that the frequency correlation is equal to 50%. The results are given in Figure29. The shape of the probability distributions is completely different except in the case of asingle unitholder . To better illustrate the impact of the frequency correlation, we reportin Figure 30 the histogram of the redemption rate by fixing n = 10. In the case of a per-fect correlation of 100% and an equally-weighted liability structure, we obtain two differentcases:1. there is zero redemption with a probability 1 − ˜ p ;2. there are n redemptions with a probability ˜ p , and the redemption severity R (cid:63) is theaverage of the individual redemption severities: R (cid:63) = 1 n n (cid:88) i =1 R (cid:63)i It follows that the probability distribution of the redemption rate is equal to: F ( x ) = { x ≥ } · (1 − ˜ p ) + { x > } · ˜ p · ¯G ( x )We retrieve the zero-inflated model ZI (cid:0) ˜ p, ˜ µ, n − / ˜ σ (cid:1) or the individual-based model witha single unitholder IM (cid:0) , ˜ p, ˜ µ, n − / ˜ σ (cid:1) . The only difference is the severity distribution ¯G , whose variance is divided by a factor n . Spillover and contagion risks come then fromthe herd behavior of unitholders. Instead of having n different investors, we have a uniqueinvestor in the fund, because the decision to redeem by one investor induces the decision toredeem by all the other remaining investors. In order to illustrate that re-demption frequencies are correlated, we build the time series of the frequency rate F t for agiven category : F t = n (cid:88) i =1 ω i,t · {E i,t = 1 } = n (cid:88) i =1 ω i,t E i,t Other illustrations are provided in Appendix C on page 103. Figures 47, 48 and 49 correspond to thecases ρ = 25%, ρ = 75% and ρ = 90%. We can use an equally-weighted scheme ω i,t = 1 /n . n of unitholders(˜ p = 50% , ˜ µ = 50% , ˜ σ = 10% , ρ = 50%)Figure 30: Histogram of the redemption rate in % with respect to the frequency correlation(˜ p = 50% , ˜ µ = 50% , ˜ σ = 10% , n = 10) 58iquidity Stress Testing in Asset Managementwhere E i,t is the redemption indicator for the investor i at time t . Using the sample( F , . . . , F T ), we compute the empirical mean F and the standard deviation ˆ σ ( F ). Then,the copula parameter θ c can be calibrated by solving the following nonlinear equation : C ( θ c ) (cid:0) F , F (cid:1) = ˆ σ ( F ) − F (cid:0) H ( ω ) − F (cid:1) − H ( ω )The copula parameter θ c can be transformed into the Kendall, Spearman or Pearson corre-lation using the standard formulas given in Table 26 on page 52. For instance, if C ( θ c ) isthe Clayton copula, the Pearson correlation is equal to: ρ = sin (cid:18) πθ c θ c + 4 (cid:19) An example is provided in Table 28 when the fund liability structure is equally-weightedand has 20 unitholders. For instance, if the empirical mean F and the standard deviationˆ σ ( F ) are equal to 25% and 20%, the calibrated Pearson correlation is equal to 44 . H ( ω ) = 1 / σ ( F ) F .
0% 20 .
0% 25 .
0% 30 .
0% 40 . .
0% 39 .
1% 5 .
1% 1 . .
0% 93 .
9% 58 .
7% 44 .
5% 34 .
7% 23 . .
0% 100 .
0% 91 .
5% 82 .
3% 72 .
8% 57 . .
0% 100 .
0% 98 .
7% 95 .
6% 87 . Remark 11
At first sight, calibrating the frequency correlation seems to be an easy task.However, it is very sensitive to the different parameters F , ˆ σ ( F ) and H ( ω ) . Moreover, itdepends on the copula specification. For instance, we obtain the results given in Table 49 onpage 104 when the dependence function is the Normal copula. We observe that the Pearsoncorrelations calibrated with the Clayton copula are different from those calibrated with theNormal copula. Remark 12
Another way to illustrate the frequency correlation is to split a given investorcategory into two subsamples S and S and calculate the time series of the redemptionfrequency for the two subsamples S k ( k = 1 , ): F k,t = 1 (cid:80) i ∈S k ω i,t (cid:88) i ∈S k ω i,t E i,t Then, we can calculate the Pearson correlation ρ ( F , F ) and calibrate the associated copulaparameter θ c using Equation (58) on page 91. Correlation risk between investor categories
The correlation risk is present withina given investor category, but it may also concern two different investor categories. Inorder to distinguish them, we use the classical statistical jargon of inter-class and intra-class correlations. In Table 29, we report the intra-class Spearman correlation for four See Equation (56) on page 90. The correlations of retail/insurance and institutional/insurance for balanced funds and the correlationsof retail/third-party distributor and retail/insurance for money market funds are not significant at theconfidence level of 95%. .
0% 52 .
9% 52 .
1% 3 . .
4% 23 .
2% 22 . − . .
0% 18 .
8% 31 . − . .
5% 48 .
0% 54 .
1% 24 . .
1% 21 .
5% 22 .
8% 39 . .
5% 16 .
2% 16 .
4% 29 . .
6% 30 .
1% 33 .
2% 12 . et al. , 1992; Wermers, 1999; Sias, 2004; Wylie, 2005; Coval and Stafford, 2007;Shleifer and Vishny, 2011; Cai et al. , 2019).Figure 31: Dependogram of redemption frequencies between retail investors and third-partydistributors Remark 13
Another way to illustrate the intra-class correlation is to report the dependo-gram (or empirical copula) of redemption frequencies. An example is provided in Figure 31for retail investors and third-party distributors. We observe that these dependogram doesnot correspond to the product copula . Examples of dependogram with the Normal copula and different correlations are provided in Figure 50on page 105.
The parameters of the copula-based model is made up by the parameters of the individual-based model (˜ p , ˜ µ and ˜ σ ) and the copula parameter θ c (or the associated frequency correla-tion). Once these parameters are estimated for a given investor/fund category, we transformthe ˜ µ − ˜ σ parameterization into the a − b parameterization of the beta distribution andcompute the risk measures M , Q ( α ), C ( α ) and S ( T ) by using the following Monte Carloalgorithm:1. we set k ←− ( u , . . . , u n ) ∼ C ( θ c ) ;3. we compute the redemption events ( E , . . . , E n ) such that: E i = { u i ≥ − ˜ p }
4. we simulate the redemption severities ( R (cid:63) , . . . , R (cid:63)n ) from the beta distribution B ( a, b );5. we compute the redemption rate for the k th simulation iteration: R ( k ) = n (cid:88) i =1 ω i E i R (cid:63)i
6. if k is equal to n S , we return the simulated sample (cid:0) R (1) , . . . , R ( n S ) (cid:1) , otherwise we set k ←− k + 1 and go back to step 2.Figure 32 shows the relationship between the correlation frequency and C (99%) for dif-ferent parameter sets when the liability structure has 20 equally-weighted unitholders. Theimpact of the correlation risk is not negligible in some cases. This is particularly true whenthe frequency correlation is close to 100%, but its impact is also significant when the fre-quency correlation is larger than 20%. On average, we observe that the risk measure C (99%)increases by 15%, 20% and 35% when the frequency correlation is respectively equal to 20%,30% and 50% compared the independent case. Remark 14
The algorithm to simulate the copula-based model CM ( n, ω, ˜ p, ˜ µ, ˜ σ, ρ ) can beused to simulate the individual-based model IM ( n, ω, ˜ p, ˜ µ, ˜ σ ) by setting C ( θ c ) = C ⊥ . This isequivalent to replace step 2 and simulate n independent uniform random numbers ( u , . . . , u n ) . In the case of daily redemptions, the correlation risk only concerns the cross-correlationbetween investors for a given market day. When we consider fire sales or liquidity crisis,the one-day study period is not adapted and must be extended to a weekly or monthlybasis. In this case, we may face time aggregation risk, meaning that redemption flows forthe subsequent market days may depend on the current redemption flows. Clayton and Normal copulas are easy to simulate using the method of transformation (Roncalli, 2020,page 803). Generally, the generation of beta random numbers is present in mathematical programming languages(Matlab, Python). Otherwise, we can use the method of rejection sampling (Roncalli, 2020, pages 886-887). It corresponds to the Pearson correlation of the Clayton copula. C (99%) with respect to the frequency correlation We recall that the total net assets at time t + 1 can be decomposed as follows:TNA ( t + 1) = (1 + R ( t + 1)) · TNA ( t ) + F + ( t + 1) − F − ( t + 1)By assuming that F + ( t + 1) = 0, we obtain:TNA ( t + 1) ≈ (1 + R ( t + 1) − R ( t + 1)) · TNA ( t )This formula is valid on a daily basis. If we consider a period of n h market days (e.g. aweekly period), we have:TNA ( t + n h ) ≈ TNA ( t ) n h (cid:89) h =1 (1 + R ( t + h ) − R ( t + h ))Therefore, it is not obvious to decompose the difference TNA ( t + n h ) − TNA ( t ) into a“ performance ” effect and a “ redemption ” effect since the two effects are related. Indeed, themathematical definition of the n h -day redemption rate is: R ( t ; t + n h ) = (cid:80) n h h =1 F − ( t + h )TNA ( t )whereas the fund return over the period [ t, t + n h ] is given by the compound formula: R ( t ; t + h ) = n h (cid:89) h =1 (1 + R ( t + h )) − et al. , 1991), we cannot separate the two effects:TNA ( t + n h ) (cid:54) = (1 + R ( t ; t + n h ) − R ( t ; t + n h )) · TNA ( t )62iquidity Stress Testing in Asset Management In the case where the performance effect is negligible — R ( t + h ) (cid:28) R ( t + h ), we have: R ( t, t + n h ) ≈ − n h (cid:89) h =1 (1 − R ( t + h )) (28)We can then calculate the probability distribution of R ( t, t + n h ) by the Monte Carlomethod. A first solution is to consider that the redemption rates are time-independent.A second solution is to consider that redemption rates are auto-correlated: R ( t ) = ρ time R ( t −
1) + ε ( t ) (29)where ρ time is the autocorrelation parameter and ε ( t ) is a random variable such that R ( t ) ∈ [0 , ε ( t ). However, this approachcan be approximated by considering a time-series copula representation:( R ( t + 1) , . . . , R ( t + n n )) ∼ C (cid:16) ˜F ( x ) , . . . , ˜F ( x ) ; Σ time ( n h ) (cid:17) (30)where ˜F is the probability distribution of R ( t ) defined by the individual-based (or copula-based) model, C is the Normal copula, whose parameters are given by the Toeplitz cor-relation matrix Σ time ( n h ) such that Σ time ( n h ) i,j = ρ | i − j | time . To calculate the probabilitydistribution of R ( t, t + n h ), we first simulate the individual-based (or copula-based) modelin order to estimate the probability distribution ˜F ( x ) of daily redemptions. Then, we gen-erate the sample of the time-series ( R ( t + 1) , . . . , R ( t + n n )) by using the method of theempirical quantile function (Roncalli, 2020, pages 806-809). Finally, we calculate the re-demption rate R ( t, t + n h ) using Equation (28). An example is provided in Figure 33 whenthe correlation between investors is equal to zero . We have also measured the impact ofthe autocorrelation value ρ time on the value-at-risk and the conditional value-at-risk. Re-sults are given in Tables 30 and 31 for six different individual-based models IM ( n, ˜ p, ˜ µ, ˜ σ ).When the value of the risk measure is small, we notice that the impact of ρ time is high.For instance, when n = 500, ˜ p = 1%, ˜ µ = 25% and ˜ σ = 10%, the value-at-risk Q (99%) isequal to 1 .
9% in the independent case. This figure increases respectively by +9% and +19%when ρ time is equal to 25% and 50%. We also notice that the impact on the conditionalvalue-at-risk is close to that on the value-at-risk. Remark 15
The compound approach defined by Equation (28) certainly overestimates stressscenarios. Indeed, we implicitly assume that the redemptions rates R ( t + h ) are identicallydistributed, meaning that there is no time effect on the individual redemption behaviour.However, we can think that an investor that redeems at time t + 1 will not redeem at time t + 2 and t + 3 . In practice, we observe that redemptions of a given investor are mutuallyexclusive during a short period of time. This property is not verified by Equation (28). Attime t + h , we notice IS ( t + h ) the set of investors that have redeemed some units before For instance, in the case of a weekly period, the Toeplitz correlation matrix is equal to:Σ time (5) = ρ time ρ ρ ρ ρ time ρ time ρ ρ ρ ρ time ρ time ρ ρ ρ ρ time ρ time ρ ρ ρ ρ time The same example with a correlation of 50% between investors is given in Figure 52 on page 106. ρ time (˜ p = 50% , ˜ µ = 50% , ˜ σ = 10% , ρ = 0% , n = 10)Table 30: Impact of the autocorrelation ρ time on the value-at-risk Q (99%) n ˜ p ˜ µ ˜ σ ρ time
0% 25% 50% 75% 100%10 000 0 .
1% 25% 10% 0 .
2% +6% +14% +24% +36%500 1 .
0% 25% 10% 1 .
9% +9% +19% +33% +50%50 2 .
0% 50% 10% 10 .
5% +12% +29% +49% +79%100 5 .
0% 50% 30% 18 .
2% +8% +18% +29% +45%10 20 .
0% 50% 30% 65 .
8% +6% +13% +21% +28%10 50 .
0% 50% 30% 90 .
1% +2% +4% +6% +8%Table 31: Impact of the autocorrelation ρ time on the conditional value-at-risk C (99%) n ˜ p ˜ µ ˜ σ ρ time
0% 25% 50% 75% 100%10 000 0 .
1% 25% 10% 0 .
2% +6% +16% +27% +41%500 1 .
0% 25% 10% 2 .
0% +9% +21% +37% +56%50 2 .
0% 50% 10% 11 .
4% +13% +32% +54% +84%100 5 .
0% 50% 30% 19 .
2% +9% +20% +32% +50%10 20 .
0% 50% 30% 68 .
8% +6% +13% +21% +28%10 50 .
0% 50% 30% 91 .
3% +2% +4% +6% +7%64iquidity Stress Testing in Asset Management t + h . We have IS ( t + 1) = { , . . . , n } . The mutually exclusive property implies that : i ∈ IS ( t + h ) ⇒ E i ( t + 1) = . . . = E i ( t + n h ) = 0 It follows that: R ( t + h ) = (cid:88) i/ ∈IS ( t + h ) ω i ( t + h ) · E i ( t + h ) · R (cid:63)i ( t + h ) and: ω i ( t + h + 1) = ω i ( t + h ) (cid:80) i/ ∈IS ( t + h ) ω i ( t + h ) Because ω i ( t + h − (cid:54) = ω i ( t + h ) and IS ( t + h − (cid:54) = IS ( t + h ) , it is obvious that R ( t + h − (cid:54) = R ( t + h ) . Therefore, the redemption decisions taken in the recent past(e.g. two or three days ago) have an impact on the future redemptions for the next days.This is a limit of the compound approach. The solution would be to develop a comprehensiveindividual-based model, whose random variables are replaced by stochastic processes. Never-theless, the complexity of such model is not worth it with respect to the large uncertainty ofstress testing exercises. Herding risk is related to momentum trading. According to Grinblatt et al. (1995), herdingbehavior corresponds to the situation where investors buy and sell the same securities atthe same time. Herding risk happens during good and bad times, and is highly documentedin economic research (Wermers, 1999; O’Neal, 2004; Ivkovi´c and Weisbenner, 2009; Ferreira et al. , 2012; Lou, 2012; Cashman et al. , 2014; Chen and Qin, 2017; Goldstein et al. , 2017;Choi et al. , 2019; D¨otz and Weth, 2019). However, we generally notice that sell herding mayhave more impact on asset prices than buy herding. Therefore, the sell-herding behaviorrisk may be associated to a price destabilizing or spillover effect. In the case of redemptionrisk, the spillover mechanism corresponds to two related effects: • A first spillover effect is that the unconditional probability of redemption is not equalto the conditional probability of the redemption given the returns of the fund duringthe recent past period:Pr { R ( t + h ) ≤ x } (cid:54) = Pr { R ( t + h ) ≤ x | ( R ( t + 1) , . . . , R ( t + h − } • A second spillover effect is that the unconditional probability of return is not equal tothe conditional probability of the return given the redemptions of the fund during therecent past period:Pr { R ( t + h ) ≤ x } (cid:54) = Pr { R ( t + h ) ≤ x | ( R ( t + 1) , . . . , R ( t + h − } This implies that the transmission of a negative shock on the redemption rate R ( t + 1)may also impact the redemption rates { R ( t + 2) , R ( t + 3) , . . . } because of the feedback For instance, if the investor has done a redemption at time t + 1, the probability that he will perform anew redemption at time t + 2 is very small, meaning that: E i ( t + 1) = 1 ⇒ E i ( t + 2) = . . . = E i ( t + n h ) = 0 R ( t + 1) R ( t + 1) R ( t + 2) R ( t + 2)loop on the fund performance. An illustration is provided in Figure 34. A large negativeredemption R ( t + 1) may induce a negative abnormal performance R ( t + 1), and this neg-ative performance may encourage the remaining investors of the fund to redeem, becausenegative returns accelerate redemption flows. This type of behavior is generally observed inthe case of fire sales and less liquid markets.As explained in the introduction, an integrated model that combines liability risk andasset risk is too ambitious and too complex. Moreover, this means modeling the policyreaction function of other investors and asset managers. Nevertheless, if we want to takeinto account sell herding, spillover or fire sales, we must build an econometric model. Forexample, the simplest way is to consider the linear dynamic model: (cid:26) R ( t ) = φ R ( t ) + u ( t ) R ( t + 1) = R + φ R ( t ) + u ( t + 1)We obtain an AR(1) process: R ( t ) = R + φ R ( t −
1) + u ( t )where φ = φ φ and u ( t ) = u ( t ) + φ u ( t −
1) is a white noise process. It follows that: E [ R ( t + h )] = 11 − φ φ R Therefore, spillover scenarios can be estimated by applying a scaling factor to the initialshock . In order to illustrate the time dependency between redemptions, we build the time seriesof the redemption rate R ( j,k ) ( t ), the redemption frequency F ( j,k ) ( t ) and the redemptionseverities R (cid:63) ( j,k ) ( t ) for each classification matrix cell ( j, k ), which is defined by a fund cate-gory FC ( j ) and an investor category IC ( k ) . For that, we calculate R ( f,k ) ( t ) the redemptionrate of the fund f for the investor category IC ( k ) at time t . Then, we estimate the dailyredemption rate R ( j,k ) ( t ) as the average of the redemption rates of all funds that belong tothe fund category FC ( j ) : R ( j,k ) ( t ) = 1 (cid:12)(cid:12) S ( j,k ) ( t ) (cid:12)(cid:12) (cid:88) f ∈S ( j,k ) ( t ) R ( f,k ) ( t ) (31)where S ( j,k ) ( t ) = (cid:8) f : f ∈ FC ( j ) , TNA ( f,k ) ( t ) > (cid:9) . We also estimate the daily redemp-tion frequency as follows: F ( j,k ) ( t ) = 1 (cid:12)(cid:12) S ( j,k ) ( t ) (cid:12)(cid:12) (cid:88) f ∈S ( j,k ) ( t ) (cid:8) R ( f,k ) ( t ) > (cid:9) (32) The previous analysis can be extended to more sophisticated process, e.g. VAR ( p ) processes. We only consider funds which have unitholders that belong to the investor category IC ( k ) . This isequivalent to impose that the assets under management held by the investor category IC ( k ) are strictlypositive: TNA ( f,k ) > R (cid:63) ( j,k ) ( t ) = 1 (cid:12)(cid:12)(cid:12) S (cid:63) ( j,k ) ( t ) (cid:12)(cid:12)(cid:12) (cid:88) f ∈S (cid:63) ( j,k ) ( t ) R ( f,k ) ( t ) (33)where S (cid:63) ( j,k ) ( t ) = (cid:8) f : f ∈ FC ( j ) , TNA ( f,k ) > , R ( f,k ) ( t ) > (cid:9) .Table 32: Autocorrelation of the redemption rate in %Balanced Bond Equity Money marketInstitutional 25 . ∗∗ − . − . . ∗∗ Insurance − . . . . ∗∗ Retail 1 . − . . . ∗∗ Third-party distributor 2 . . . . ∗∗ The computation of R ( j,k ) ( t ), F ( j,k ) ( t ) and R (cid:63) ( j,k ) ( t ) does make sense only if there isenough observations (cid:12)(cid:12) S ( j,k ) ( t ) (cid:12)(cid:12) and (cid:12)(cid:12)(cid:12) S (cid:63) ( j,k ) ( t ) (cid:12)(cid:12)(cid:12) at time t . This is why we focus on the mostrepresentative investor categories (retail, third-party distributor, institutional and insur-ance) and fund categories (balanced, bond, equity and money market). In Table 32, wereport the maximum between the autocorrelation ρ ( R ( t ) , R ( t − ρ ( R ( t ) , R ( t − ∗∗ the matrix cells where the p -value of the autocorrelation is lower than 5%. Except for moneymarket funds and the institutional/balanced matrix cell, redemptions are not significantlyautocorrelated. If we consider redemption frequencies and severities, we observe more au-tocorrelation (see Tables 50 and 51 on page 106). However, for bond and equity funds, theresults show that the autocorrelation is significant and high for the redemption frequency,but low for the redemption severity. The last section of this article is dedicated to the factors that may explain a redemptionstress. First, we investigate whether it is due to a redemption frequency shock or a redemp-tion severity shock. Second, we study how market risk may explain extreme redemptionrates, and we focus on three factors: stock returns, bond returns and volatility levels.
We may wonder whether the time variation of redemption rates is explained by the timevariation of redemption frequencies or redemption severities. Using the time series built inSection 4.3.4 on page 66, we consider three linear regression models: R ( t ) = β + β F ( t ) + u ( t ) R ( t ) = β + β R (cid:63) ( t ) + u ( t ) R ( t ) = β + β F ( t ) + β R (cid:63) ( t ) + u ( t )In the first model, we explain the redemption rate using the redemption frequency. In thesecond model, the explanatory variable is the redemption severity. Finally, the third modelcombines the two previous models. For each classification matrix cell ( j, k ), we have reportedthe centered coefficient of determination R c in Tables 33, 34 and 35.67iquidity Stress Testing in Asset ManagementTable 33: Coefficient of determination R c in % — R ( t ) = β + β F ( t ) + u ( t )Balanced Bond Equity Money marketInstitutional 2 . . . . . . . . . . . . . . . . R c in % — R ( t ) = β + β R (cid:63) ( t ) + u ( t )Balanced Bond Equity Money marketInstitutional 87 . . . . . . . . . . . . . . . . R c in % — R ( t ) = β + β F ( t ) + β R (cid:63) ( t ) + u ( t )Balanced Bond Equity Money marketInstitutional 88 . . . . . . . . . . . . . . . . R c is greater than 50% onlyfor the institutional/equity category. R c takes a value around 35% for the retail/balanced,retail/bond and institutional/bond categories, otherwise it is less than 20%. Results forthe second linear regression are better. This indicates that the redemption severity is abetter explanatory variable than the redemption frequency. The only exception is the in-stitutional/equity category. The combination of the two variables allows us to improve theexplanatory power of the model, but we also notice that the redemption severity is theprimary factor. The matrix cell with the highest R c is retail/equity, whereas the matrixcell with the lowest R c is institutional/equity. The scatter plot between R ( t ), F ( t ) and R (cid:63) ( t ) for these two extreme cases are reported in Figures 35 and 36. For the retail/equitycategory, we verify that the redemption severity explains the redemption rate. For the in-stitutional/equity category, the redemption severity is not able to explain the high values ofthe redemption rate.The previous results are very interesting since the redemption severity is the primaryfactor for explaining the redemption shocks. Therefore, a high variation of the redemptionrate is generally due to an increase of the redemption severity. Nevertheless, there aresome exceptions where stress scenarios are also explained by an increase in the redemptionfrequency. Remark 16
We have used the coefficient R c to show the power explanation of the twovariables F ( t ) and R (cid:63) ( t ) without considering the effect of the constant. For some matrixcells, we notice that the constant may be important (see Tables 52, 53 and 54 on page 106). R ( t ), F ( t ) and R (cid:63) ( t ) (retail/equity) Figure 36: Relationship between R ( t ), F ( t ) and R (cid:63) ( t ) (institutional/equity) Numerous academic research papers suggest that investor flows depend on past performance.According to Sirri and Tufano (1998) and Huang et al. (2007), there is an asymmetryconcerning the flow-performance relationship: equity mutual funds with good performancegain a lot of money inflows, while equity mutual funds with poor performance suffer smalleroutflows. However, this asymmetry concerns relative performance. Indeed, according toIvkovi´c and Weisbenner (2009), “ inflows are related only to relative performance ” while“ outflows are related only to absolute fund performance ”. Therefore, these authors suggestthat investors sell the asset class when this one has a bad performance. In the case ofcorporate bonds, Goldstein et al. (2017) find that relative performance also matters interms of explaining outflows. In order to better understand these results, we consider thefollowing analytical model : (cid:26) R f ( t ) = α f ( t ) + β f ( t ) R mkt ( t ) + ε ( t ) R f ( t ) = γ f + δ f α f ( t −
1) + ϕ f R f ( t −
1) + η ( t )where R f ( t ) is the return of the fund f , R mkt ( t ) is the return of the market risk factorand R f ( t ) is the redemption rate of the fund f . ε ( t ) and η ( t ) are two independent whitenoise processes. Using the first equation, we can estimate the relative performance of thefund, which is measured by its alpha component α f ( t ). The second equation states that theredemption rate R f ( t ) of the fund depends on the past relative performance α f ( t −
1) andthe past absolute performance R f ( t − H : δ f < H : ϕ f <
0. Accepting H implies that outflows depend on the relative performance,while accepting H implies that outflows depend on the absolute performance. In both cases,the value of the coefficient is negative, because we expect that a negative performance willincrease the redemption rate. The previous framework can be extended to take into accounta more sophisticated model for determining the relative performance α f ( t ) or to considerlagged variables (Bellando and Tran-Dieu, 2011; Ferreira et al. , 2012; Lou, 2012; Cashman et al. , 2014; Barber et al. , 2016; Fricke and Fricke, 2017). More generally, we have: R f ( t ) = γ f + p (cid:88) h =1 (cid:16) φ ( h ) f R f ( t − h ) + δ ( h ) f α f ( t − h ) + ϕ ( h ) f R f ( t − h ) (cid:17) + η ( t ) (34)Even if this type of flow-performance relationship is interesting to understand the investorbehavior, it is however not adapted in the case of a stress testing program for two reasons.The first reason is that Equation (34) is calibrated using low frequency data, e.g. quarterlyor monthly data. Therefore, the goal of Equation (34) is to describe long-term behaviorof investors, whereas stress testing of liabilities concerns short-term periods. The secondreason is the inadequacy of this approach with macro stress testing approaches developedby regulators and institutional bodies. If we consider stress testing programs developed in the banking sector (Roncalli, 2020, pages893-922), we distinguish historical, probabilistic and macroeconomic approaches. While thefirst two methods have been developed in the previous sections, we focus on the third method,which is the approach used by the regulators (Board of Governors of the Federal Reserve See Arora et al. (2019). For instance, we can use the three-factor Fama-French model or the four-factor Carhart model. • the performance of the bond market; • the performance of the stock market; • market volatility.Therefore, we assume that there is a linear relationship between the redemption rate andthese factors: R ( t ) = β + β F bond ( t ) + β F stock ( t ) + β F vol ( t ) + u ( t ) (35)where F bond ( t ) and F stock ( t ) are the h -day total returns of the FTSE World Broad Investment-Grade Bond index and the MSCI World index, F vol ( t ) is the difference of the VIX indexbetween t − h and t , and h is the time horizon.In Table 36, we report the coefficient of determination R c for the one-day time horizon.These figures are disappointing since the impact of the market risk factors are very low .For instance, the highest R -squared is reached for the third-party distributor/money marketcategory, but it is equal to 4 . R c (cid:28)
5% (see Tables 55 and 56 on page 107).Table 36: Coefficient of determination R c in % — Equation (35), one-day time horizonBalanced Bond Equity Money marketInstitutional 0 . . . . . . . . . . . . . . . . Remark 17
The previous results suggest that redemption rates do not depend on marketrisk factors on a short-term basis. However, fund managers generally have the feeling thatredemption rates increase when there is a stress on market returns. Nevertheless, we knowthat returns are more or less independent from one day to another. Therefore, we consideranother approach using market sentiment. For that, we compute the average redemption ratewhen the VIX index is above 30, and calculate its relative variation with respect to the entireperiod. Results are given in Table 37. We observe an impact in particular for bond/equityfunds and institutional/third-party distributor investors.
Table 37: Relative variation of the redemption rate R ( t ) when VIX ≥ .
3% +54 .
7% +74 .
3% +64 . − . − . − .
2% +75 . .
1% +21 .
5% +13 . − . .
6% +43 .
6% +49 .
5% +22 . Nevertheless, we verify that β is negative for equity funds, even though the relationship betweenredemption rate and stock returns is not convincing as shown in Figure 53 on page 108. Liquidity stress testing is a recent topic in asset management, which has given rise to numer-ous publications from regulators (AMF, 2017; BaFin, 2017; ESMA, 2019; FSB, 2017; IOSCO,2015, 2018), investment management associations (AFG, 2015; EFAMA, 2020) and affiliatedresearchers from central banks and international bodies (Arora et al. , 2019; Baranova et al. ,2017; Bouveret, 2017; Fricke and Fricke, 2017; Gourdel et al. , 2018). On the academic side,few studies specifically concern liquidity stress testing in asset management . Therefore,we observe a gap between general concepts and specific measurement models. As such, thepurpose of our study is to propose several analytical approaches in order to implement LSTpractical programs.Besides the historical approach that considers non-parametric risk measures, we havedeveloped a frequency-severity model that is useful when building parametric risk measuresof liquidity stress testing. This statistical approach can be seen as a reduced-form modelbased on three parameters: the redemption frequency, the expected redemption severityand the redemption uncertainty. Like the historical approach, the frequency-severity modelrequires some expert judgements to correct some data biases. Nevertheless, both historicaland analytical approaches are simple enough to verify properties of risk ordering coherencybetween fund and investor categories.We have also developed an individual-based behavioral model, which is an extension ofthe frequency-severity model. We have shown that redemption risk depends on the fundliability structure, and is related to the Herfindahl index of assets under management heldby unitholders. Even if this model is hard to implement because it requires knowing thecomprehensive liability structure, it allows us to justify liquidity stress testing based onthe largest fund holders. Moreover, this model shows the importance of cross-correlationbetween unitholders of a same investor category, but also of several investor categories.Nevertheless, the individual-based behavioral model is flexible enough that it can easilytake into account dependencies between investors by incorporating a copula model. Again,the issue with this extended individual-based behavioral model lies in the knowledge of theliability structure.The production of stress scenarios can be obtained by considering a risk measure appliedto the redemption rate. For the historical approach, we can use a value-at-risk or a condi-tional value-at-risk figure, which is estimated with non-parametric statistical methods. Forthe frequency-severity and individual-based behavioral models, the estimation of the VaRor CVaR is based on analytical formulas. Moreover, these models may produce parametricstress scenarios for a given return time. Another issue concerns the choice of data betweengross or net redemption rates for calibrating these stress scenarios. For some categories, netredemption rates may be used to proxy gross redemption rates, because they are very closein stress periods. However, we also demonstrate that it is better to use gross redemptionrates for some investor or fund categories (e.g. retail investors or money market funds).The design of macro stress testing programs is more complicated than expected. Sincethe flow-performance relationship is extensively documented by academic research, it isvalid at low frequencies, typically on a quarterly or annual basis. In this case, we mayobserve inflows towards the best fund managers. However, this relationship mainly concernsrelative performance, whereas macro stress testing programs deal with absolute performance.Indeed, relative performance is a key parameter when we want to analyze the idiosyncraticliability liquidity risk at the fund level. Nevertheless, the liquidity risk in asset management Because data on liabilities are not publicly available. However, we can cite Christoffersen and Xu(2017) and Darolles et al. (2018), who specifically study asset management flows with respect to the liabilitystructure of the investment fund. regulation of asset managers has beenlagging behind that of banks since the global financial crisis ”. The implementation of theliquidity coverage ratio (LCR) and the net stable funding ratio (NSFR), the use of liquidityand high-quality liquid assets (HQLA) buffers and the definition of regulatory monitoringtools date back to 2010 for the banking industry (BCBS, 2010, 2013). The regulatoryframework on liquidity stress testing proposed by ESMA (2019) is an important step for thedevelopment of liquidity measurement in the asset management industry. In this paper, wedevelop an analytical framework and give some answers. However, it is still early days andmuch remains to be done. The LCR became a minimum requirement for BCBS member countries in January 2015.
References
Arora , R.,
B´edard-Pag´e , G.,
Leblanc , G.O., and
Shotlander , R. (2019), Bond Fundsand Fixed-Income Market Liquidity: A Stress-Testing Approach,
Bank of Canada Tech-nical Report , 115.Association Fran¸caise de la Gestion Financi`ere (2015), Code of Conduct on the Use of StressTests,
AFG Report , April.Autorit´e des March´es Financiers (2017), The Use of Stress Tests as Part of Risk Management,
AMF Guide , February.
Baranova , Y.,
Coen , J.,
Noss , J., and
Lowe , P. (2017), Simulating Stress Across theFinancial System: The Resilience of Corporate Bond Markets and the Role of InvestmentFunds,
Financial Stability Paper , 42, Bank of England.
Barber , B.M.,
Huang , X., and
Odean , T. (2016), Which Factors Matter to Investors?Evidence from Mutual Fund Flows,
Review of Financial Studies , 29(10), pp. 2600-2642.Basel Committee on Banking Supervision (1996),
Amendment to the Capital Accord toIncorporate Market Risks , January 1996.Basel Committee on Banking Supervision (1999),
A New Capital Adequacy Framework , FirstConsultative Paper on Basel II, June 1999.Basel Committee on Banking Supervision (2001),
The New Basel Capital Accord , SecondConsultative Paper on Basel II, January 2001.Basel Committee on Banking Supervision (2010),
Basel III: A Global Regulatory Frameworkfor More Resilient Banks and Banking Systems , December 2010.Basel Committee on Banking Supervision (2013),
Basel III: The Liquidity Coverage Ratioand Liquidity Risk Monitoring Tools , January 2013.Basel Committee on Banking Supervision (2016),
Interest Rate Risk in the Banking Book ,April 2016.Basel Committee on Banking Supervision (2017),
Stress Testing Principles , ConsultativeDocument, December 2017.
Bellando , R., and
Tran-Dieu , L. (20111), Fund Flow/Performance Relationship — ACase Study of French Mutual Funds,
Revue Economique , 62(2), pp. 225-276.
Blanqu´e , P., and
Mortier , V. (2019a), How Investors Should Deal with the LiquidityDilemna,
Amundi CIO Insights , February.
Blanqu´e , P., and
Mortier , V. (2019b), Liquidity Dilemna Needs A Regulatory Response,
Amundi Investment Talks , October.Board of Governors of the Federal Reserve System (2017), Supervisory Scenarios for AnnualStress Tests Required under the Dodd-Frank Act Stress Testing Rules and the CapitalPlan Rule, . Bouveret , A. (2017), Liquidity Stress Tests for Investment Funds: A Practical Guide,
IMFWorking Paper , 17/226. 74iquidity Stress Testing in Asset Management
Brinson , G.P.,
Singer , B.D., and
Beebower , G.L. (1991), Determinants of PortfolioPerformance II: An Update,
Financial Analysts Journal , 47(3), pp. 40-48.
Brunnermeier , M.K., and
Pedersen , L.H. (2009), Market Liquidity and Funding Liquid-ity,
Review of Financial Studies , 22(6), pp. 2201-2238.Bundesanstalt f¨ur Finanzdienstleistungsaufsicht (2017), Liquidity Stress Testing by GermanAsset Management Companies, December 2017.
Cai , F.,
Han , S., Li , D., and Li , Y. (2019), Institutional Herding and its Price Impact:Evidence from the Corporate Bond Market, Journal of Financial Economics , 131(1), pp.139-167.
Cashman , G.D.,
Deli , D.,
Nardari , F., and
Villupuram , S.V. (2014), Investor Behaviorin the Mutual Fund Industry: Evidence from Gross Flows,
Journal of Economics andFinance , 38(4), pp. 541-567.
Chen , Y., and
Qin , N. (2017), The Behavior of Investor Flows in Corporate Bond MutualFunds,
Management Science , 63(5), pp. 1365-1381.
Choi , J.,
Hoseinzade , S.,
Shin , S.S., and
Tehranian , H. (2019), Corporate Bond MutualFunds and Asset Fire Sales,
SSRN , . Christoffersen , S.K., and Xu, H. (2015), Investor Attrition and Fund Flows in MutualFunds,
Journal of Financial and Quantitative Analysis , 52(3), pp. 867-893.
Coval , J.D., and
Stafford , E. (2007), Asset Fire Sales (and Purchases) in Equity Markets,
Journal of Financial Economics , 86(2), pp. 479-512.
Darolles , S.,
Le Fol , G., Lu , Y., and Sun , T. (2018), A Self-Exciting Model for MutualFund Flows: Investor Behaviour and Liability Risk,
Paris December 2018 Finance MeetingEUROFIDAI-AFFI . D¨otz , N., and
Weth , M.A. (2019), Redemptions and Asset Liquidations in CorporateBond Funds,
Discussion Paper , 11, Deutsche Bundesbank.
Duarte , F., and
Eisenbach , T.M. (2013), Fire-sale Spillovers and Systemic Risk (RevisedDecember 2019),
Staff Report , 645, Federal Reserve Bank of New York.European Banking Authority (2020a), 2020 EU-wide Stress Testing,
Methodological Note ,January.European Banking Authority (2020b), Market Risk Shocks,
Methodological Note , January.European Central Bank (2019), Sensitivity Analysis of Liquidity Risk – Stress Test 2019,
Final Report , September.European Fund and Asset Management Association (2020), Managing Fund Liquidity Riskin Europe — Recent Regulatory Enhancements & Proposals for Further Improvements,
AMIC/EFAMA Report , January.European Securities and Markets Authority (2019), Guidelines on Liquidity Stress Testingin UCITS and AIFs,
Final Report , September.
Ferreira , M.A.,
Keswani , A.,
Miguel , A.F., and
Ramos , S.B. (2012), The Flow-Performance Relationship Around The World,
Journal of Banking & Finance , 36(6), pp.1759-1780. 75iquidity Stress Testing in Asset ManagementFinancial Stability Board (2010),
Reducing the Moral Hazard Posed by Systemically Impor-tant Financial Institutions , Consultation Document, October 2010.Financial Stability Board (2015),
Assessment Methodologies for Identifying Non-bank Non-insurer Global Systemically Important Financial Institutions , Second Consultation Docu-ment, March 2015.Financial Stability Board (2017),
Policy Recommendations to Address Structural Vulnera-bilities from Asset Management Activities , January 2017.Financial Stability Board (2020),
Global Monitoring Report on Non-Bank Financial Inter-mediation 2019 , January 2020.
Fricke , C., and
Fricke , D. (2017), Vulnerable Asset Management? The Case of MutualFunds,
Discussion Paper , 32, Deutsche Bundesbank.
Goldstein , I.,
Jiang , H., and Ng , D.T. (2017), Investor Flows and Fragility in CorporateBond Funds, Journal of Financial Economics , 126(3), pp. 592-613.
Gourdel , R.,
Maqui , E., and
Sydow , M. (2018), Investments Funds Under Stress,
ECBWorking Paper , 2323.
Grillet-Aubert , L. (2018), Macro Stress Tests: What do they mean for the Markets andfor the Asset Management Industry?,
Risk and Trend Mapping , Autorit´e des March´esFinanciers, June.
Grinblatt , M.,
Titman , S., and
Wermers , R. (1995), Momentum Investment Strate-gies, Portfolio Performance, and Herding: A Study of Mutual Fund Behavior,
AmericanEconomic Review , 85(5), pp. 1088-1105.
Huang , J.,
Wei , K.D. and
Yan , H. (2007), Participation Costs and the Sensitivity of FundFlows to Past Performance,
Journal of Finance , 62(3), pp. 1273-1311.International Organization of Securities Commissions (2015), Liquidity Management Toolsin Collective Investment Schemes,
Final Report , 28, December.International Organization of Securities Commissions (2018), Recommendations for Liquid-ity Risk Management for Collective Investment Schemes,
Final Report , 01, February.
Ivkovi´c , Z., and
Weisbenner , S. (2009), Individual Investor Mutual Fund Flows,
Journalof Financial Economics , 92(2), pp. 223-237.
Kacperczyk , M., and
Schnabl , P. (2013), How Safe are Money Market Funds?,
QuarterlyJournal of Economics , 128(3), pp. 1073-1122.
Lakonishok , J.,
Shleifer , A., and
Vishny , R.W. (1992) The Impact of InstitutionalTrading on Stock Prices,
Journal of Financial Economics , 32(1), pp. 23-43.
Lou , D. (2012), A Flow-based Explanation for Return Predictability,
Review of FinancialStudies , 25(12), pp. 3457-3489.
Min , Y., and
Agresti , A. (2002), Modeling Nonnegative Data with Clumping at Zero: ASurvey,
Journal of the Iranian Statistical Society , 1(1-2), pp. 7-33
Nelsen , R.B. (2006),
An Introduction to Copulas , Second edition, Springer.76iquidity Stress Testing in Asset Management
O’Neal , E.S. (2004), Purchase and Redemption Patterns of US Equity Mutual Funds,
Financial Management , 33(1), pp. 63-90.
Ong , L.L. (2014),
A Guide to IMF Stress Testing: Methods and Models , InternationalMonetary Fund.
Ospina , R., and
Ferrari , S.L. (2010), Inflated Beta Distributions,
Statistical papers , 51(1),pp. 111-126.
Persaud , A.D. (2003),
Liquidity Black Holes: Understanding, Quantifying and ManagingFinancial Liquidity Risk , Risk Books.
Roncalli , T. (2020),
Handbook of Financial Risk Management , Chapman & Hall/CRCFinancial Mathematics Series.
Roncalli , T., and
Weisang , G. (2015a), Asset Management and Systemic Risk,
SSRN , . Roncalli , T., and
Weisang , G. (2015b),
Response to FSB-IOSCO Second ConsultativeDocument, Assessment Methodologies for Identifying Non-Bank Non-Insurer Global Sys-temically Important Financial Institutions , May 28, . Schmidt , L.,
Timmermann , A., and
Wermers , R. (2016), Runs on Money Market MutualFunds,
American Economic Review , 106(9), pp. 2625-2657.Securities and Exchange Commission (2015), Open-End Fund Liquidity Risk ManagementPrograms; Swing Pricing; Re-Opening of Comment Period for Investment Company Re-porting Modernization Release,
Proposed Rule , 33-9922, September.Securities and Exchange Commission (2016), Investment Company Liquidity Risk Manage-ment Programs,
Final Rule , 33-10233, October.Securities and Exchange Commission (2018a), Investment Company Liquidity Disclosure,
Proposed Rule , IC-33046, March.Securities and Exchange Commission (2018b), Investment Company Liquidity Disclosure,
Final Rule , IC-33142, June.
Shleifer , A., and
Vishny , R. (2011), Fire Sales in Finance and Macroeconomics,
Journalof Economic Perspectives , 25(1), pp. 29-48.
Sias , R.W. (2004), Institutional Herding,
Review of Financial Studies , 17(1), pp. 165-206.
Sirri , E.R., and
Tufano , P. (1998), Costly Search and Mutual Fund Flows,
Journal ofFinance , 53(5), pp. 1589-1622.
Thompson , J. (2019), H2O, Woodford and GAM Crises Highlight Liquidity Risk,
FinancialTimes , 29 June 2019.
Wermers , R. (1999), Mutual Fund Herding and the Impact on Stock Prices,
Journal ofFinance , 54(2), pp. 581-622.
Wylie , S. (2005), Fund Manager Herding: A Test of the Accuracy of Empirical Resultsusing UK Data,
Journal of Business , 78(1), pp. 381-403.77iquidity Stress Testing in Asset Management
AppendixA Mathematical results
A.1 Granularity and the X -statistic We consider n funds whose redemption rate is equal to p . The assets under managementof each fund are set to $1. The maximum redemption rate of n funds is equal to themathematical expectation of n Bernoulli random variables: p (max) = E [max ( B ( p ) , . . . , B n ( p ))]= 1 − (1 − p ) n whereas the redemption rate of the sum of n funds is equal to the expected frequency of aBinomial random variable: p (sum) = E [ B ( n, p )] n = p In Table 38, we report the value taken by the ratio p (max) /p (sum). For example, this ratiois equal to 3 .
71 if p = 5% and n = 4. To understand this ratio, we can consider a large fundwhose redemption probability is p . This fund is split into n funds of the same size. The ratioindicates the multiplication factor to obtain the maximum of the redemption rates amongthe n funds. Table 38: Value of the ratio p (max) /p (sum) n Probability p .
00 1 .
00 1 .
00 1 .
00 1 .
00 1 .
00 1 .
002 2 .
00 2 .
00 1 .
99 1 .
95 1 .
90 1 .
80 1 .
503 3 .
00 3 .
00 2 .
97 2 .
85 2 .
71 2 .
44 1 .
754 4 .
00 3 .
99 3 .
94 3 .
71 3 .
44 2 .
95 1 .
885 5 .
00 4 .
99 4 .
90 4 .
52 4 .
10 3 .
36 1 . .
00 9 .
96 9 .
56 8 .
03 6 .
51 4 .
46 2 . .
88 48 .
79 39 .
50 18 .
46 9 .
95 5 .
00 2 . .
51 95 .
21 63 .
40 19 .
88 10 .
00 5 .
00 2 . A.2 Statistical moments of zero-inflated probability distribution
A.2.1 General formulas
A zero-inflated random variable Z can be written as the product of a Bernoulli randomvariables X ∼ B ( p ) and a positive random variable Y : Z = XY Let µ (cid:48) m ( Z ) for the m -th moment of Z . Using the previous relationship, we deduce that: µ (cid:48) m ( Z ) = E [ Z m ]= E [ X m Y m ]= E [ X m ] E [ Y m ]= pµ (cid:48) m ( Y ) (36)78iquidity Stress Testing in Asset Managementbecause X and Y are independent by definition, and X m = X , implying that X m followsa Bernoulli distribution B ( p ). From Equation (36), we can compute the m -th centeredmoment µ m ( Z ). For that, we recall that: µ = µ (cid:48) µ = µ (cid:48) − µ µ = µ (cid:48) − µ (cid:48) µ + 2 µ µ = µ (cid:48) − µ (cid:48) µ + 6 µ (cid:48) µ − µ We deduce the expression of the second moment: µ (cid:48) = µ + µ For the third moment, we have: µ (cid:48) = µ + 3 µ (cid:48) µ − µ = µ + 3 (cid:0) µ + µ (cid:1) µ − µ = µ + 3 µ µ + µ = γ µ / + 3 µ µ + µ where γ is the skewness coefficient. For the fourth moment, it follows that: µ (cid:48) = µ + 4 µ (cid:48) µ − µ (cid:48) µ + 3 µ = µ + 4 (cid:16) γ µ / + 3 µ µ + µ (cid:17) µ − (cid:0) µ + µ (cid:1) µ + 3 µ = µ + 4 γ µ / µ + 12 µ µ + 4 µ − µ µ − µ + 3 µ = µ + 4 γ µ / µ + 6 µ µ + µ = ( γ + 3) µ + 4 γ µ / µ + 6 µ µ + µ where γ is the excess kurtosis coefficient. We can then compute the moments of Z . For themean, we have: µ ( Z ) = µ (cid:48) ( Z )= pµ ( Y ) (37)We deduce that the variance of Z is equal to: µ ( Z ) = µ (cid:48) ( Z ) − µ ( Z )= pµ (cid:48) ( Y ) − p µ ( Y )= pµ ( Y ) + p (1 − p ) µ ( Y ) (38)For the third moment, we have: µ ( Z ) = µ (cid:48) ( Z ) − µ (cid:48) ( Z ) µ ( Z ) + 2 µ ( Z )= pµ (cid:48) ( Y ) − p µ (cid:48) ( Y ) µ ( Y ) + 2 p µ ( Y )= p (cid:16) γ ( Y ) µ / ( Y ) + 3 µ ( Y ) µ ( Y ) + µ ( Y ) (cid:17) − p (cid:0) µ ( Y ) + µ ( Y ) (cid:1) µ ( Y ) + 2 p µ ( Y )= pγ ( Y ) µ / ( Y ) + 3 p (1 − p ) µ ( Y ) µ ( Y ) + p (1 − p ) (1 − p ) µ ( Y )79iquidity Stress Testing in Asset ManagementIt follows that the skewness coefficient is equal to: γ ( Z ) = µ ( Z ) µ / ( Z )= ϑ ( Z )( pµ ( Y ) + p (1 − p ) µ ( Y )) / (39)where: ϑ ( Z ) = pγ ( Y ) µ / ( Y ) + 3 p (1 − p ) µ ( Y ) µ ( Y ) + p (1 − p ) (1 − p ) µ ( Y )For the fourth moment, we have: µ ( Z ) = µ (cid:48) ( Z ) − µ (cid:48) ( Z ) µ ( Z ) + 6 µ (cid:48) ( Z ) µ ( Z ) − µ ( Z )= pµ (cid:48) ( Y ) − p µ (cid:48) ( Y ) µ ( Y ) + 6 p µ (cid:48) ( Y ) µ ( Y ) − p µ ( Y )= p ( γ ( Y ) + 3) µ ( Y ) + 4 pγ ( Y ) µ / ( Y ) µ ( Y ) + 6 pµ ( Y ) µ ( Y ) + pµ ( Y ) − p γ ( Y ) µ / ( Y ) µ ( Y ) − p µ ( Y ) µ ( Y ) − p µ ( Y ) +6 p µ ( Y ) µ ( Y ) + 6 p µ ( Y ) − p µ ( Y )= p ( γ ( Y ) + 3) µ ( Y ) + 4 p (1 − p ) γ ( Y ) µ / ( Y ) µ ( Y ) +6 p (1 − p ) µ ( Y ) µ ( Y ) + p (1 − p ) (cid:0) − p + 3 p (cid:1) µ ( Y ) (40)We deduce that the excess kurtosis coefficient is equal to: γ ( Z ) = µ ( Z ) µ ( Z ) − ϑ ( Z )( pµ ( Y ) + p (1 − p ) µ ( Y )) (41)where: ϑ ( Z ) = p ( γ ( Y ) + 3) µ ( Y ) + 4 p (1 − p ) γ ( Y ) µ / ( Y ) µ ( Y ) +6 p (1 − p ) µ ( Y ) µ ( Y ) + p (1 − p ) (cid:0) − p + 3 p (cid:1) µ ( Y ) − p µ ( Y ) − p (1 − p ) µ ( Y ) µ ( Y ) − p (1 − p ) µ ( Y )= ( pγ ( Y ) + 3 p (1 − p )) µ ( Y ) + 4 p (1 − p ) γ ( Y ) µ / ( Y ) µ ( Y ) +6 p (1 − p ) (1 − p ) µ ( Y ) µ ( Y ) + p (1 − p ) (cid:0) − p + 6 p (cid:1) µ ( Y )We can deduce the following properties:1. The skewness of Z is equal to zero if and only if:(a) the skewness of Y is equal to zero and the frequency probability p is equal to one;(b) the frequency probability p is equal to zero, meaning that Z is always equal tozero.2. The excess kurtosis of Z is equal to zero if and only if:(a) the kurtosis of Y is equal to 3 and the frequency probability p is equal to one;80iquidity Stress Testing in Asset Management(b) the frequency probability p is equal to zero, meaning that Z is always equal tozero.In other cases, the skewness and excess kurtosis coefficients of Z are different from zero evenif the random variable Y is not skewed and has not fat tails. Remark 18
The previous results seem to be contradictory with the properties given in Equa-tion (17) on page 26. In fact, the limit case p → + is not equal to p = 0 , because there isa singularity at the point p = 0 . A.2.2 Application to the beta distribution
We assume that Y ∼ B ( a, b ). Since we have: µ ( Y ) = aa + b we deduce that: µ ( Z ) = p aa + b For the second moment, we have: µ ( Y ) = ab ( a + b ) ( a + b + 1)and: µ ( Z ) = p ab ( a + b ) ( a + b + 1) + p (1 − p ) (cid:18) aa + b (cid:19) = p ab + (1 − p ) a ( a + b + 1)( a + b ) ( a + b + 1)This formula has been already found by Ospina and Ferrari (2010). The skewness and excesskurtosis coefficients of the beta distribution are equal to: γ ( Y ) = 2 ( b − a ) √ a + b + 1( a + b + 2) √ ab and: γ ( Y ) = 6 ( a − b ) ( a + b + 1) ab ( a + b + 2) ( a + b + 3) − a + b + 3)We plug these different expressions into the general formulas to obtain γ ( Z ) and γ ( Z ). A.3 Maximum likelihood of the zero-inflated model
We consider a sample { x , . . . , x n } of n observations, and we assume that X follows azero-inflated model, whose frequency and probability distributions are p and G ( x ; θ ). Thelog-likelihood of the i th observation is equal to: (cid:96) i ( p, θ ) = ln Pr { X = x i } = ln f ( x i )= { x i = 0 } · ln (1 − p ) + { x i > } · ln ( pg ( x i ; θ ))= { x i = 0 } · ln (1 − p ) + { x i > } · ln p + { x i > } · ln g ( x i ; θ ) The formulas are not reported here because they don’t have a lot of interest. (cid:96) ( p, θ ) = n (cid:88) i =1 (cid:96) i ( p, θ )= n ln (1 − p ) + ( n − n ) ln p + (cid:88) x i > ln g ( x i )where n is the number of observations x i that are equal to zero. The maximum likelihoodestimator (cid:16) ˆ p, ˆ θ (cid:17) is defined as follows: (cid:110) ˆ p, ˆ θ (cid:111) = arg max (cid:96) ( p, θ )and satisfies the first-order conditions: ∂ p (cid:96) (cid:16) ˆ p ; ˆ θ (cid:17) = 0 ∂ θ (cid:96) (cid:16) ˆ p ; ˆ θ (cid:17) = We deduce that: ∂ p (cid:96) (cid:16) ˆ p ; ˆ θ (cid:17) = 0 ⇔ − n − ˆ p + n − n ˆ p = 0 ⇔ ˆ p = n − n n (42)The concentrated log-likelihood function becomes: (cid:96) (ˆ p, θ ) = n ln n + ( n − n ) ln ( n − n ) − n ln n + (cid:88) x i > ln g ( x i )Therefore, the ML estimator ˆ θ corresponds to the ML estimator of θ when considering onlythe observations x i that are strictly positive:ˆ θ = arg max (cid:96) (ˆ p, θ )= arg max (cid:88) x i > ln g ( x i ) (43) Remark 19
In the case of the zero-inflated beta model, we have θ = ( a, b ) and: (cid:110) ˆ a, ˆ b (cid:111) = arg max (cid:88) x i > (cid:18) ( a −
1) ln x i + ( b −
1) ln (1 − x i ) − ln B ( a, b ) (cid:19) (44) A.4 Statistical properties of the individual-based model
We define the random variable ˜ Z as the sum of products of two random variables:˜ Z = n (cid:88) i =1 ω i ˜ X i ˜ Y i where ˜ X i ∼ B (˜ p ) and ˜ Y i are iid random variables. Moreover, we assume that ω i > (cid:80) ni =1 ω i = 1. 82iquidity Stress Testing in Asset Management A.4.1 Computation of Pr (cid:110) ˜ Z = 0 (cid:111) This case corresponds to the situation where no client redeems:Pr (cid:110) ˜ Z = 0 (cid:111) = Pr (cid:40) n (cid:88) i =1 ω i ˜ X i ˜ Y i = 0 (cid:41) = Pr (cid:110) ˜ X = 0 , . . . , ˜ X n = 0 (cid:111) = n (cid:89) i =1 Pr (cid:110) ˜ X i = 0 (cid:111) = (1 − ˜ p ) n (45) A.4.2 Statistical momentsFirst moment
For the mean, we have: E (cid:104) ˜ Z (cid:105) = E (cid:34) n (cid:88) i =1 ω i ˜ X i ˜ Y i (cid:35) = n (cid:88) i =1 ω i E (cid:104) ˜ X i (cid:105) E (cid:104) ˜ Y i (cid:105) We deduce that: µ (cid:16) ˜ Z (cid:17) = ˜ pµ (cid:16) ˜ Y (cid:17) (46) Second moment
Since we have E (cid:104) ˜ X i (cid:105) = ˜ p and E (cid:104) ˜ Y i (cid:105) = µ (cid:48) (cid:16) ˜ Y (cid:17) , it follows that: E (cid:104) ˜ Z (cid:105) = E (cid:32) n (cid:88) i =1 ω i ˜ X i ˜ Y i (cid:33) = E n (cid:88) i =1 ω i ˜ X i ˜ Y i + 2 (cid:88) j>i ω i ω j ˜ X i ˜ X j ˜ Y i ˜ Y j = ˜ pµ (cid:48) (cid:16) ˜ Y (cid:17) n (cid:88) i =1 ω i + 2˜ p µ (cid:16) ˜ Y (cid:17) (cid:88) j>i ω i ω j We notice that: 1 = n (cid:88) i =1 ω i = (cid:32) n (cid:88) i =1 ω i (cid:33) = n (cid:88) i =1 ω i + 2 (cid:88) j>i ω i ω j µ (cid:16) ˜ Z (cid:17) = E (cid:104) ˜ Z (cid:105) − E (cid:104) ˜ Z (cid:105) = ˜ pµ (cid:48) (cid:16) ˜ Y (cid:17) n (cid:88) i =1 ω i + 2˜ p µ (cid:16) ˜ Y (cid:17) (cid:88) j>i ω i ω j − ˜ p µ (cid:16) ˜ Y (cid:17) = ˜ pµ (cid:48) (cid:16) ˜ Y (cid:17) n (cid:88) i =1 ω i + 2˜ p µ (cid:16) ˜ Y (cid:17) (cid:88) j>i ω i ω j − ˜ p µ (cid:16) ˜ Y (cid:17) n (cid:88) i =1 ω i + 2 (cid:88) j>i ω i ω j Therefore, the variance of ˜ Z is equal to: µ (cid:16) ˜ Z (cid:17) = (cid:16) ˜ pµ (cid:48) (cid:16) ˜ Y (cid:17) − ˜ p µ (cid:16) ˜ Y (cid:17)(cid:17) n (cid:88) i =1 ω i = ˜ p (cid:16) µ (cid:16) ˜ Y (cid:17) + (1 − ˜ p ) µ (cid:16) ˜ Y (cid:17)(cid:17) n (cid:88) i =1 ω i (47) Remark 20
In the equally-weighted case, we obtain: µ (cid:16) ˜ Z (cid:17) = ˜ p (cid:16) µ (cid:16) ˜ Y (cid:17) + (1 − ˜ p ) µ (cid:16) ˜ Y (cid:17)(cid:17) n Application to the beta severity distribution
If we assume that ˜ Y i ∼ B (cid:16) ˜ a, ˜ b (cid:17) , wehave: µ (cid:16) ˜ Y (cid:17) = ˜ a ˜ a + ˜ b and: µ (cid:16) ˜ Y (cid:17) = ˜ a ˜ b (cid:16) ˜ a + ˜ b (cid:17) (cid:16) ˜ a + ˜ b + 1 (cid:17) We deduce that: µ (cid:16) ˜ Z (cid:17) = ˜ p ˜ a ˜ a + ˜ b and: µ (cid:16) ˜ Z (cid:17) = ˜ pn ˜ a ˜ b (cid:16) ˜ a + ˜ b (cid:17) (cid:16) ˜ a + ˜ b + 1 (cid:17) + (1 − ˜ p ) ˜ a (cid:16) ˜ a + ˜ b (cid:17) = ˜ p ˜ an ˜ b + (1 − ˜ p ) ˜ a (cid:16) ˜ a + ˜ b + 1 (cid:17)(cid:16) ˜ a + ˜ b (cid:17) (cid:16) ˜ a + ˜ b + 1 (cid:17) A.5 Moment matching between the zero-inflated model and theindividual-based model
In order to calibrate the probability p , we match the redemption probability Pr { R > } .Using the results in Appendix A.4.1 on page 83, we obtain: p = 1 − Pr { R = 0 } = 1 − (1 − ˜ p ) n For the first moment, we have: E [ R ] = pµ = ˜ p ˜ µ We deduce that: µ = ˜ p − (1 − ˜ p ) n ˜ µ For the second moment, we have: σ ( R ) = pσ + p (1 − p ) µ = ˜ p (cid:0) ˜ σ + (1 − ˜ p ) ˜ µ (cid:1) n (cid:88) i =1 ω i It follows that: σ = ˜ p (cid:0) ˜ σ + (1 − ˜ p ) ˜ µ (cid:1) (cid:80) ni =1 ω i − p (1 − p ) µ p = ˜ p (cid:0) ˜ σ + (1 − ˜ p ) ˜ µ (cid:1) (cid:80) ni =1 ω i − (1 − ˜ p ) n − (1 − ˜ p ) n ˜ p (1 − (1 − ˜ p ) n ) ˜ µ = (cid:18) ˜ p H ( ω )1 − (1 − ˜ p ) n (cid:19) ˜ σ + (cid:32) ˜ p ((1 − ˜ p ) − (1 − ˜ p ) n ) H ( ω ) − ˜ p (1 − ˜ p ) n (1 − H ( ω ))(1 − (1 − ˜ p ) n ) (cid:33) ˜ µ where H ( ω ) = (cid:80) ni =1 ω i is the Herfindahl index. Remark 21
If we consider the equally-weighted case ω i = n − , we have H ( ω ) = n − and: σ = 1 n (cid:18) ˜ p − (1 − ˜ p ) n (cid:19) ˜ σ + 1 n (cid:32) ˜ p ((1 − ˜ p ) − (1 − ˜ p ) n ) − ˜ p (1 − ˜ p ) n ( n − − (1 − ˜ p ) n ) (cid:33) ˜ µ When ˜ p (cid:54) = 0, the limit cases are: lim n →∞ p = 1and: lim n →∞ µ = ˜ p ˜ µ For the parameter σ , we obtain:lim n →∞ σ = ˜ p (cid:0) ˜ σ + (1 − ˜ p ) ˜ µ (cid:1) H ( ω )For an infinitely fine-grained liability structure, we have:lim n →∞ σ = 085iquidity Stress Testing in Asset Management A.6 Upper bound of the Herfindahl index under partial informa-tion
Let π k be a probability distribution, meaning that π k ≥ (cid:80) nk =1 π k = 1. The Herfindahlindex is equal to: H = n (cid:88) k =1 π k = n (cid:88) k =1 π k : n = n (cid:88) k =1 π n − k +1: n where: 0 ≤ min π k = π n ≤ π n ≤ · · · ≤ π k : n ≤ π k +1: n ≤ · · · ≤ π n : n = max π k We have: H = m (cid:88) k =1 π n − k +1: n + n (cid:88) k = m +1 π n − k +1: n where k = 1 : m denotes the largest contributions that are known, meaning that we don’tknow the values taken by { π n , . . . , π n − m : n } . Since we have π n − k : n ≤ π n − k +1: n , we deducethat: n (cid:88) k = m +1 π n − k +1: n ≤ (cid:18) − (cid:80) mk =1 π n − k +1: n π n − m +1: n (cid:19) π n − m +1: n = (cid:32) − m (cid:88) k =1 π n − k +1: n (cid:33) π n − m +1: n and : H ≤ H + m = m (cid:88) k =1 π n − k +1: n + (cid:32) − m (cid:88) k =1 π n − k +1: n (cid:33) π n − m +1: n (48)An example is given in Table 39. The Herfindahl index is equal to 17 . . m π m (in %) 30 .
00 20 .
00 15 .
00 10 .
00 9 .
00 7 .
00 5 .
00 4 . H + m (in %) 30 .
00 23 .
00 20 .
50 18 .
75 18 .
50 18 .
18 18 .
00 17 . A.7 Correlated redemptions with copula functions
We define the random variable ˜ Z as previously:˜ Z = n (cid:88) i =1 ω i ˜ X i ˜ Y i We verify that H + n = H . Y i are iid random variables. We assume that ˜ X i ∼ B (˜ p ) are identically distributed,but not independent. We note C ( u , . . . , u n ) the copula function of the random vector (cid:16) ˜ X , . . . , ˜ X n (cid:17) and B ( x ) the cumulative distribution function of the Bernoulli random vari-able B (˜ p ). This means that B (0) = 1 − ˜ p and B (1) = 1.In practice we use the Clayton copula: C ( θ c ) ( u , . . . , u n ) = (cid:16) u − θ c + · · · + u − θ c n − n + 1 (cid:17) − /θ c or the Normal copula : C ( θ c ) ( u , . . . , u n ) = Φ (cid:0) Φ − ( u ) + · · · + Φ − ( u n ) ; C n ( θ c ) (cid:1) The Clayton parameter satisfies θ c ≥ θ c lies in the range[ − , C ( θ c ) ( u , u ) = C ( θ c ) ( u , u , , . . . ,
1) = (cid:16) u − θ c + u − θ c − (cid:17) − /θ c and: C ( θ c ) ( u , u ) = C ( θ c ) ( u , u , , . . . ,
1) = Φ (cid:0) Φ − ( u ) + Φ − ( u ) ; C ( θ c ) (cid:1) A.7.1 Joint probability of two ˜ X i ’s We consider the bivariate case. The probability mass function is described by the followingcontingency table: ˜ X = 0 ˜ X = 1˜ X = 0 π , π , π = 1 − ˜ p ˜ X = 1 π , π , π = ˜ pπ = 1 − ˜ p π = ˜ p (cid:110) ˜ X ≤ u , ˜ X ≤ u (cid:111) = C ( θ c ) ( B ( u ) , B ( u )), we deduce that: C ( θ c ) ( B (0) , B (0)) = C ( θ c ) (1 − ˜ p, − ˜ p ) C ( θ c ) ( B (0) , B (1)) = C ( θ c ) (1 − ˜ p,
1) = 1 − ˜ p C ( θ c ) ( B (1) , B (0)) = C ( θ c ) (1 , − ˜ p ) = 1 − ˜ p C ( θ c ) ( B (1) , B (1)) = C ( θ c ) (1 ,
1) = 1and: ˜ X = 0 ˜ X = 1˜ X = 0 C ( θ c ) (1 − ˜ p, − ˜ p ) 1 − ˜ p − C ( θ c ) (1 − ˜ p, − ˜ p ) 1 − ˜ p ˜ X = 1 1 − ˜ p − C ( θ c ) (1 − ˜ p, − ˜ p ) C ( θ c ) (1 − ˜ p, − ˜ p ) + 2˜ p − p − ˜ p ˜ p C ( θ c ) = C ⊥ , ˜ X and ˜ X are independent, we retrieve the results obtainedfor the individual-based model: ˜ X = 0 ˜ X = 1˜ X = 0 (1 − ˜ p ) (1 − ˜ p ) ˜ p − ˜ p ˜ X = 1 (1 − ˜ p ) ˜ p ˜ p ˜ p − ˜ p ˜ p The Normal copula depends on the correlation matrix Σ. Here, we assume a uniform redemptioncorrelation, implying that Σ is the constant correlation matrix C n ( θ c ) where θ c is the pairwise correlation. C ⊥ ( u , u ) = u u . In the case where C ( θ c ) = C + , ˜ X and ˜ X are perfectlydependent and we obtain the following contingency table:˜ X = 0 ˜ X = 1˜ X = 0 1 − ˜ p − ˜ p ˜ X = 1 0 ˜ p ˜ p − ˜ p ˜ p C + ( u , u ) = min ( u , u ). The contingency tables (51) and (52) represent the twoextremes cases. Remark 22
If we use a radially symmetric copula (Nelsen, 2006) such that: C ( θ c ) ( u , u ) = u + u − C ( θ c ) (1 − u , − u ) the contingency table (50) becomes: ˜ X = 0 ˜ X = 1˜ X = 0 1 − p + C ( θ c ) (˜ p, ˜ p ) ˜ p − C ( θ c ) (˜ p, ˜ p ) 1 − ˜ p ˜ X = 1 p − C ( θ c ) (˜ p, ˜ p ) C ( θ c ) (˜ p, ˜ p ) ˜ p − ˜ p ˜ p In the general case, we obtain a similar contingency table by replacing the copula function C ( θ c ) ( u , u ) by its corresponding survival function ˘C ( θ c ) ( u , u ) because we have (Nelsen,2006): ˘C ( θ c ) ( u , u ) = u + u − C ( θ c ) (1 − u , − u ) A.7.2 Computation of Pr (cid:110) ˜ Z = 0 (cid:111) This case corresponds to the situation where no client redeems:Pr (cid:110) ˜ Z = 0 (cid:111) = Pr (cid:40) n (cid:88) i =1 ω i ˜ X i ˜ Y i = 0 (cid:41) = Pr (cid:110) ˜ X = 0 , . . . , ˜ X n = 0 (cid:111) = C ( θ c ) (1 − ˜ p, . . . , − ˜ p ) (53)In the case where C ( θ c ) = C ⊥ , we retrieve the result Pr (cid:110) ˜ Z = 0 (cid:111) = (1 − ˜ p ) n . In the casewhere C ( θ c ) = C + , we obtain Pr (cid:110) ˜ Z = 0 (cid:111) = 1 − ˜ p . A.7.3 Statistical momentsFirst moment
For the mean, we have: E (cid:104) ˜ Z (cid:105) = E (cid:34) n (cid:88) i =1 ω i ˜ X i ˜ Y i (cid:35) = n (cid:88) i =1 ω i E (cid:104) ˜ X i (cid:105) E (cid:104) ˜ Y i (cid:105) We deduce that: µ (cid:16) ˜ Z (cid:17) = ˜ pµ (cid:16) ˜ Y (cid:17) (54)88iquidity Stress Testing in Asset Management Second moment
Using the contingency table (50), we have: E (cid:104) ˜ X ˜ X (cid:105) = C ( θ c ) (1 − ˜ p, − ˜ p ) + 2˜ p − ˘C ( θ c ) (˜ p, ˜ p )It follows that: E (cid:104) ˜ Z (cid:105) = E (cid:32) n (cid:88) i =1 ω i ˜ X i ˜ Y i (cid:33) = E n (cid:88) i =1 ω i ˜ X i ˜ Y i + 2 (cid:88) j>i ω i ω j ˜ X i ˜ X j ˜ Y i ˜ Y j = ˜ pµ (cid:48) (cid:16) ˜ Y (cid:17) n (cid:88) i =1 ω i + 2 ˘C ( θ c ) (˜ p, ˜ p ) µ (cid:16) ˜ Y (cid:17) (cid:88) j>i ω i ω j and µ (cid:16) ˜ Z (cid:17) = E (cid:104) ˜ Z (cid:105) − E (cid:104) ˜ Z (cid:105) = ˜ pµ (cid:48) (cid:16) ˜ Y (cid:17) n (cid:88) i =1 ω i + 2 ˘C ( θ c ) (˜ p, ˜ p ) µ (cid:16) ˜ Y (cid:17) (cid:88) j>i ω i ω j − ˜ p µ (cid:16) ˜ Y (cid:17) = ˜ p (cid:16) µ (cid:16) ˜ Y (cid:17) + µ (cid:16) ˜ Y (cid:17)(cid:17) H ( ω ) + ˘C ( θ c ) (˜ p, ˜ p ) µ (cid:16) ˜ Y (cid:17) (1 − H ( ω )) − ˜ p µ (cid:16) ˜ Y (cid:17) = ˜ pµ (cid:16) ˜ Y (cid:17) H ( ω ) + (cid:16) ˜ p H ( ω ) + ˘C ( θ c ) (˜ p, ˜ p ) (1 − H ( ω )) − ˜ p (cid:17) µ (cid:16) ˜ Y (cid:17) = (cid:16) ˜ pµ (cid:16) ˜ Y (cid:17) + (cid:16) ˜ p − ˘C ( θ c ) (˜ p, ˜ p ) (cid:17) µ (cid:16) ˜ Y (cid:17)(cid:17) H ( ω ) + (cid:16) ˘C ( θ c ) (˜ p, ˜ p ) − ˜ p (cid:17) µ (cid:16) ˜ Y (cid:17) (55)In the case where C ( θ c ) = C ⊥ , we have ˘C ( θ c ) (˜ p, ˜ p ) = ˜ p . Therefore, we retrieve the resultfound in Equation (47) on page 84: µ (cid:16) ˜ Z (cid:17) = ˜ p (cid:16) µ (cid:16) ˜ Y (cid:17) + (1 − ˜ p ) µ (cid:16) ˜ Y (cid:17)(cid:17) H ( ω )In the case where C ( θ c ) = C + , we have ˘C ( θ c ) (˜ p, ˜ p ) = ˜ p and we obtain: µ (cid:16) ˜ Z (cid:17) = ˜ pµ (cid:16) ˜ Y (cid:17) H ( ω ) + ˜ p (1 − ˜ p ) µ (cid:16) ˜ Y (cid:17) A.8 Statistical moments of the redemption frequency
We recall that ˜ X i ∼ B (˜ p ), meaning that E (cid:104) ˜ X i (cid:105) = E (cid:104) ˜ X i (cid:105) = ˜ p . The weighted redemptionfrequency is defined as follows: F = n (cid:88) i =1 ω i ˜ X i We have: E [ F ] = E (cid:34) n (cid:88) i =1 ω i ˜ X i (cid:35) = ˜ p E (cid:2) F (cid:3) = E (cid:32) n (cid:88) i =1 ω i ˜ X i (cid:33) = E n (cid:88) i =1 ω i ˜ X i + 2 (cid:88) j>i ω i ω j ˜ X i ˜ X j = ˜ p H ( ω ) + ˘C ( θ c ) (˜ p, ˜ p ) (1 − H ( ω ))We deduce that: µ ( F ) = ˜ p H ( ω ) + ˘C ( θ c ) (˜ p, ˜ p ) (1 − H ( ω )) − ˜ p Remark 23
We notice that the expected value and the volatility of the redemption frequencyare related in the following way: µ ( F ) = E [ F ] ( H ( ω ) − E [ F ]) + ˘C ( E [ F ] , E [ F ]) (1 − H ( ω )) (56) A.9 Pearson correlation between two redemption frequencies
We consider two redemption frequencies F and F . The redemption frequency F k isassociated to the liability structure ( ω k, , . . . , ω k,n k ) and corresponds to an investor category,whose redemption probability is ˜ p k and frequency correlation is characterized by the copulafunction C ( θ k ) ( k = 1 , C ( θ ) . It follows that we have threecopula functions: • C ( θ ) is the copula function that defines the frequency correlation between the investorsof the first category; • C ( θ ) is the copula function that defines the frequency correlation between the investorsof the second category; • C ( θ ) is the copula function that defines the frequency correlation between the in-vestors of the first category and those of the second category.In the case where the two categories are the same, we have C ( θ ) = C ( θ ) = C ( θ ) = C ( θ c ) .To compute the covariance between F and F , we calculate the mathematical expecta-tion of the cross product: E [ F F ] = E (cid:32) n (cid:88) i =1 ω ,i ˜ X ,i (cid:33) n (cid:88) j =1 ω ,j ˜ X ,j = E n (cid:88) i =1 n (cid:88) j =1 ω ,i ω ,j ˜ X ,i ˜ X ,j = E (cid:104) ˜ X ,i ˜ X ,j (cid:105) n (cid:88) i =1 n (cid:88) j =1 ω ,i ω ,j = ˘C ( θ ) (˜ p , ˜ p )90iquidity Stress Testing in Asset Managementbecause (cid:80) n i =1 (cid:80) n j =1 ω ,i ω ,j = 1. We deduce the expression of the Pearson correlation: ρ ( F , F ) = ˘C ( θ ) (˜ p , ˜ p ) − ˜ p ˜ p (cid:112) µ ( F ) µ ( F ) (57)where: µ ( F k ) = ˜ p k ( H ( ω k ) − ˜ p k ) + ˘C ( θ k ) (˜ p k , ˜ p k ) (1 − H ( ω k )) k = 1 , Remark 24
The Pearson correlation ρ ( F , F ) is equal to zero if only if C ( θ k ) is theproduct copula C ⊥ . Remark 25
In the case where the two investor categories are the same and the liabilitystructures are equally-weighted, we have ˜ p = ˜ p = ˜ p and C ( θ ) = C ( θ ) = C ( θ ) = C ( θ c ) ,and we obtain: ρ ( F , F ) = ˘C ( θ c ) (˜ p, ˜ p ) − ˜ p (cid:112) µ ( F ) µ ( F ) (58) where: µ ( F k ) = ˘C ( θ c ) (˜ p, ˜ p ) − ˜ p + ˜ p − ˘C ( θ c ) (˜ p, ˜ p ) n k k = 1 , The limiting case n k → ∞ is equal to ρ ( F , F ) = 1 . This is normal since F and F converges to ˜ p when the liability structure is infinitely fine-grained. B Data We recall that C ( θ k ) is the Clayton or the Normal copula. In the general case, this property does nothold. i q u i d i t y S t r e ss T e s t i n g i n A ss e t M a n ag e m e n t Table 40: Breakdown of the liability dataset by investor and fund categoriesTotal number n Balanced Bond Enhanced Equity Money Other Structured Totalof observations Treasury MarketAuto-consumption 22 762 46 651 3 784 46 678 6 175 34 064 0 160 114Central bank 2 791 7 400 0 4 730 602 0 0 15 523Corporate 10 780 13 457 2 305 6 962 7 812 6 164 0 47 480Corporate pension fund 14 827 24 429 427 17 975 3 029 5 474 427 66 588Employee savings plan 9 894 4 240 1 349 19 145 3 232 0 5 279 43 139Institutional 50 813 95 013 3 961 76 057 9 542 31 973 241 267 600Insurance 10 577 45 494 3 303 23 145 12 633 6 528 0 101 680Other 27 938 29 817 5 816 4 898 9 347 18 717 0 96 533Retail 140 023 86 937 7 531 99 624 15 418 31 370 83 496 464 399Sovereign 7 291 12 788 854 14 183 3 471 5 308 0 43 895Third-party distributor 63 792 86 716 5 247 123 004 11 160 15 407 5 126 310 452Total 361 488 452 942 34 577 436 401 82 421 155 005 94 569 1 617 403Total number n Balanced Bond Enhanced Equity Money Other Structured Totalof redemptions Treasury MarketAuto-consumption 3 744 8 796 1 135 11 871 3 040 883 0 29 469Central bank 4 16 0 38 18 0 0 76Corporate 324 484 144 159 3 110 20 0 4 241Corporate pension fund 460 513 17 447 213 17 2 1 669Employee savings plan 264 120 40 519 74 0 145 1 162Institutional 1 973 3 098 74 3 422 2 754 229 0 11 550Insurance 568 1 562 114 1 596 2 409 61 0 6 310Other 1 145 926 219 805 2 009 278 0 5 382Retail 54 095 36 018 3 932 67 862 6 882 5 030 22 783 196 602Sovereign 494 118 9 381 521 2 0 1 525Third-party distributor 19 837 29 140 2 277 54 689 7 127 4 569 334 117 973Total 82 908 80 791 7 961 141 789 28 157 11 089 23 264 375 959
Source : Amundi Cube Database (2020) and authors’ calculation. i q u i d i t y S t r e ss T e s t i n g i n A ss e t M a n ag e m e n t Table 41: Breakdown of the liability dataset by investor and fund categories (without mandates and dedicated mutual funds)Total number n Balanced Bond Enhanced Equity Money Other Structured Totalof observations Treasury MarketAuto-consumption 16 147 43 189 3 783 43 737 6 008 13 793 0 126 657Central bank 1 281 580 0 476 0 0 0 2 337Corporate 1 862 6 542 2 305 5 468 7 812 4 235 0 28 224Corporate pension fund 2 344 8 650 427 9 031 2 670 1 277 0 24 399Employee savings plan 9 894 4 240 1 349 19 145 3 232 0 5 279 43 139Institutional 6 858 36 792 3 716 41 104 8 329 16 029 0 112 828Insurance 3 436 13 011 3 303 21 832 8 543 5 750 0 55 875Other 7 577 12 751 5 428 4 155 9 333 11 788 0 51 032Retail 115 394 77 879 6 692 95 393 14 798 27 834 83 118 421 108Sovereign 2 969 2 261 854 3 405 2 853 1 746 0 14 088Third-party distributor 55 696 75 591 4 929 114 171 10 732 13 483 5 126 279 728Total 223 458 281 486 32 786 357 917 74 310 95 935 93 523 1 159 415Total number n Balanced Bond Enhanced Equity Money Other Structured Totalof redemptions Treasury MarketAuto-consumption 3 492 8 385 1 135 11 137 3 040 881 0 28 070Central bank 2 2 0 7 0 0 0 11Corporate 280 405 144 157 3 110 9 0 4 105Corporate pension fund 190 292 17 304 202 0 0 1 005Employee savings plan 264 120 40 519 74 0 145 1 162Institutional 1 328 2 312 73 2 677 2 734 166 0 9 290Insurance 419 874 114 1 576 2 385 60 0 5 428Other 733 493 200 804 2 008 262 0 4 500Retail 51 454 35 079 3 932 67 250 6 770 4 875 22 707 192 067Sovereign 484 72 9 343 520 1 0 1 429Third-party distributor 18 808 28 242 2 266 52 445 7 077 4 431 334 113 603Total 77 454 76 276 7 930 137 219 27 920 10 685 23 186 360 670
Source : Amundi Cube Database (2020) and authors’ calculation. iquidity Stress Testing in Asset Management C Additional results
Figure 37: Third-party distributor
20 40 60 80 100-100-50050100 0 20 40 60 80 10002040608010020 40 60 80 100-100-50050100 0 20 40 60 80 100020406080100 n equivalent small funds Figure 39: Relationship between the confidence level α of F − ( α ) and the confidence level α G of G − ( α G ) S ( T ) in % ( p = 5%) Figure 41: Stress scenario S ( T ) in % ( p = 50%) a (method of moments)(1) (2) (3) (4) (5) (6) (7) (8)Auto-consumption 0 .
02 0 .
05 0 .
03 0 .
03 0 .
09 0 .
06 0 . .
00 0 .
04 0 .
21 0 . .
01 0 .
01 0 .
21 0 . .
14 0 .
04 0 . .
01 0 .
04 0 .
06 0 .
10 0 . .
01 0 .
02 0 .
02 0 .
12 0 . .
05 0 .
05 0 .
01 0 .
05 0 .
01 0 . .
01 0 .
01 0 .
01 0 .
01 0 .
05 0 .
01 0 .
00 0 . .
05 0 .
02 0 .
11 0 . .
01 0 .
03 0 .
02 0 .
02 0 .
07 0 .
01 0 .
02 0 . .
01 0 .
02 0 .
02 0 .
01 0 .
07 0 .
02 0 .
00 0 . (1) = balanced, (2) = bond, (3) = enhanced treasury, (4) = equity, (5) = money market, (6) = other, (7)= structured, (8) = total Table 43: Estimated value of b (method of moments)(1) (2) (3) (4) (5) (6) (7) (8)Auto-consumption 1 .
23 2 .
86 1 .
20 2 .
25 2 .
81 2 .
23 2 . .
78 1 .
62 5 .
34 3 . .
41 0 .
50 2 .
69 1 . .
89 1 .
84 1 . .
21 1 .
52 2 .
13 2 .
15 1 . .
78 0 .
91 1 .
07 3 .
58 1 . .
58 1 .
84 1 .
04 1 .
35 1 .
17 1 . .
29 3 .
56 2 .
98 4 .
11 2 .
39 3 .
07 1 .
11 3 . .
69 0 .
83 0 .
90 0 . .
83 4 .
21 1 .
44 5 .
22 5 .
14 1 .
48 1 .
43 3 . .
68 2 .
80 1 .
01 2 .
89 2 .
58 1 .
60 0 .
97 2 . (1) = balanced, (2) = bond, (3) = enhanced treasury, (4) = equity, (5) = money market, (6) = other, (7)= structured, (8) = total a (method of maximum likelihood)(1) (2) (3) (4) (5) (6) (7) (8)Auto-consumption 0 .
20 0 .
26 0 .
26 0 .
23 0 .
32 0 .
25 0 . .
23 0 .
19 0 .
39 0 . .
13 0 .
13 0 .
37 0 . .
03 0 .
52 0 . .
22 0 .
19 0 .
21 0 .
28 0 . .
14 0 .
15 0 .
17 0 .
28 0 . .
26 0 .
21 0 .
27 0 .
25 0 .
28 0 . .
31 0 .
30 0 .
26 0 .
33 0 .
27 0 .
27 0 .
36 0 . .
68 0 .
17 0 .
31 0 . .
40 0 .
28 0 .
24 0 .
30 0 .
34 0 .
27 0 .
26 0 . .
29 0 .
25 0 .
23 0 .
27 0 .
29 0 .
24 0 .
32 0 . (1) = balanced, (2) = bond, (3) = enhanced treasury, (4) = equity, (5) = money market, (6) = other, (7)= structured, (8) = total Table 45: Estimated value of b (method of maximum likelihood)(1) (2) (3) (4) (5) (6) (7) (8)Auto-consumption 6 .
53 11 .
36 17 .
89 15 .
80 7 .
50 8 .
40 10 . .
32 6 .
03 8 .
96 6 . .
66 3 .
12 4 .
14 2 . .
62 24 .
41 30 . .
24 4 .
94 5 .
65 4 .
64 5 . .
56 5 .
42 5 .
26 7 .
14 5 . .
00 7 .
99 31 .
90 4 .
82 43 .
34 6 . .
99 82 .
20 51 .
08 116 .
86 10 .
51 48 .
15 309 .
26 65 . .
65 4 .
84 2 .
06 2 . .
38 39 .
19 15 .
23 64 .
70 21 .
61 36 .
15 12 .
11 47 . .
28 26 .
79 15 .
16 44 .
67 7 .
80 20 .
02 206 .
39 26 . (1) = balanced, (2) = bond, (3) = enhanced treasury, (4) = equity, (5) = money market, (6) = other, (7)= structured, (8) = total µ in % (method of maximum likelihood)(1) (2) (3) (4) (5) (6) (7) (8)Auto-consumption 2 .
92 2 .
23 1 .
45 1 .
43 4 .
08 2 .
91 2 . .
87 3 .
10 4 .
12 4 . .
47 4 .
03 8 .
29 5 . .
36 2 .
07 1 . .
32 3 .
78 3 .
55 5 .
77 4 . .
78 2 .
62 3 .
08 3 .
80 3 . .
93 2 .
56 0 .
83 4 .
99 0 .
64 3 . .
54 0 .
36 0 .
51 0 .
28 2 .
47 0 .
56 0 .
12 0 . .
06 3 .
47 13 .
01 6 . .
36 0 .
72 1 .
55 0 .
46 1 .
55 0 .
75 2 .
13 0 . .
66 0 .
92 1 .
46 0 .
59 3 .
53 1 .
16 0 .
15 0 . (1) = balanced, (2) = bond, (3) = enhanced treasury, (4) = equity, (5) = money market, (6) = other, (7)= structured, (8) = total Table 47: Estimated value of σ in % (method of maximum likelihood)(1) (2) (3) (4) (5) (6) (7) (8)Auto-consumption 6 .
05 4 .
16 2 .
74 2 .
88 6 .
66 5 .
41 4 . .
77 6 .
45 6 .
18 7 . .
74 9 .
53 11 .
74 11 . .
32 2 .
80 2 . .
73 7 .
70 7 .
06 9 .
59 7 . .
80 6 .
23 6 .
82 6 .
58 7 . .
78 5 .
21 1 .
58 8 .
84 1 .
20 6 . .
96 0 .
65 0 .
99 0 .
48 4 .
52 1 .
06 0 .
19 0 . .
07 7 .
46 18 .
33 11 . .
56 1 .
33 3 .
04 0 .
83 2 .
58 1 .
41 3 .
95 1 . .
20 1 .
80 2 .
97 1 .
13 6 .
13 2 .
33 0 .
27 1 . (1) = balanced, (2) = bond, (3) = enhanced treasury, (4) = equity, (5) = money market, (6) = other, (7)= structured, (8) = total .
47 3 .
11 5 .
42 3 .
06 6 .
45 2 .
40 3 . .
33 1 .
25 2 .
29 1 . .
16 2 .
45 3 .
43 3 .
07 5 .
08 2 .
22 3 . .
02 1 .
92 1 .
03 2 .
53 4 .
09 0 .
00 2 . .
57 0 .
41 2 .
75 1 .
42 0 .
59 2 .
65 1 . .
42 2 .
58 7 .
07 2 .
45 6 .
89 1 .
87 3 . .
05 2 .
79 1 .
49 2 .
77 4 .
53 1 .
89 3 . .
15 1 .
90 4 .
79 3 .
22 5 .
70 1 .
01 3 . .
88 1 .
74 2 .
55 1 .
76 5 .
18 1 .
36 1 .
38 1 . .
10 0 .
45 1 .
66 3 .
19 10 .
07 2 .
39 4 . .
57 2 .
15 5 .
22 1 .
76 3 .
88 3 .
37 1 .
80 2 . .
96 2 .
29 4 .
45 2 .
18 5 .
48 2 .
05 1 .
51 2 . (1) = balanced, (2) = bond, (3) = enhanced treasury, (4) = equity, (5) = money market, (6) = other, (7)= structured, (8) = total Figure 42: Liability weights in the case of the geometric liability structure ω i ∝ q i ˜F ( x | ω ) and ˜F ( x | H ) ( q = 0 . H ( ω ) − = 18)) Figure 44: Comparison of ˜F ( x | ω ) and ˜F ( x | H ) ( q = 0 . H ( ω ) − = 3)) { R = 0 } in % with respect to the number n of unitholders (˜ p = 10%) Figure 46: Probability to observe 100% of redemptions Pr { F = 1 } in % ( n = 20) n of unitholders(˜ p = 50% , ˜ µ = 50% , ˜ σ = 10% , ρ = 25%)Figure 48: Histogram of the redemption rate in % with respect to the number n of unitholders(˜ p = 50% , ˜ µ = 50% , ˜ σ = 10% , ρ = 75%) 103iquidity Stress Testing in Asset ManagementFigure 49: Histogram of the redemption rate in % with respect to the number n of unitholders(˜ p = 50% , ˜ µ = 50% , ˜ σ = 10% , ρ = 90%)Table 49: Calibrated Pearson correlation (Normal copula, H ( ω ) = 1 / σ ( F ) F .
0% 20 .
0% 25 .
0% 30 .
0% 40 . . .
0% 39 .
88% 24 . .
0% 50 .
00% 42 .
83% 38 .
88% 35 .
70% 31 . .
0% 50 .
00% 49 .
20% 47 .