Credit Crunch: The Role of Household Lending Capacity in the Dutch Housing Boom and Bust 1995-2018
CCredit Crunch: The Role of Household LendingCapacity in the Dutch Housing Boom and Bust1995-2018
Menno Schellekens and Taha Yasseri , , , Oxford Internet Institute, University of Oxford, Oxford, UK School of Sociology, University College Dublin, Dublin, Ireland Geary Institute for Public Policy, University College Dublin, Dublin, Ireland Alan Turing Institute for Data Science and AI, London, UK
January 5, 2021
Abstract
What causes house prices to rise and fall? Economists identify household access to creditas a crucial factor. "Loan-to-Value" and "Debt-to-GDP" ratios are the standard measures forcredit access. However, these measures fail to explain the depth of the Dutch housing bust afterthe 2009 Financial Crisis. This work is the first to model household lending capacity based onthe formulas that Dutch banks use in the mortgage application process. We compare the abilityof regression models to forecast housing prices when different measures of credit access areutilised. We show that our measure of household lending capacity is a forward-looking, highlypredictive variable that outperforms ‘Loan-to-Value’ and debt ratios in forecasting the Dutchcrisis. Sharp declines in lending capacity foreshadow the market deceleration.
Keywords—
Housing Price, Loan-to-Value, Dutch Market, Lending Capacity, Loan-to-Income
The flow of credit from the financial sector to the housing market is critical for understanding house prices,because households usually finance properties with mortgage debt. The availability of credit determines how * Corresponding author: Taha Yasseri, D405 John Henry Newman Building, University College Dublin, StillorganRd, Belfield, Dublin 4, Ireland. Email: [email protected]. a r X i v : . [ s t a t . A P ] J a n uch households can borrow and thus how much they can bid on properties. As relaxed credit constraintsallow all market participants to borrow more, households often have to borrow more to stay competitive, andhouse prices rise quickly (Bernanke and Gertler, 1995; Kiyotaki and Moore, 1997). After the Great FinancialCrisis, scholars started emphasising the importance of credit conditions to the development of housing prices.Studies of housing markets in advanced economies consistently find that empirical models that includemeasures of credit conditions outperform models that do not. Competing measures of credit conditions haveemerged in the literature. Some authors employ the average ‘loan-to-value’ (LTV) ratio of first time buyers(Duca and Muellbauer, 2016). Others choose mortgage debt to GDP ratios (Oikarinen, 2009), survey datafrom senior bank employees (van der Veer and Hoebrechts, 2016) and indeces of the ‘ease’ of credit policy(Chauvin and Muellbauer, 2014).The Dutch housing market experienced a boom and bust in the period 1995-2018. The Dutch case is apuzzle because existing models do not explain the size of the boom and the depth of the bust. Figure 1 showsthat house prices rose rapidly in the 1990’s and early 2000’s and entered a period of sustained decline after2009. Despite falling interest rates, housing prices fell 16% from their peak value. The crisis came at greatcost to many Dutch families. In 2015, 28% of homes were deemed ‘under water’: the home was worth lessthan the outstanding mortgage debt (Centraal Plan Bureau, 2014).Based on a qualitative analysis of Dutch housing reforms, we develop a novel approach to forecastinghouse prices that utilises a new measure of credit access. We study the ‘Loan-to-Income’ formulas thatbanks use to calculate how much money households can borrow relative to their income. We model theseformulas and calculate lending capacity for the average Dutch household from 1995 to 2018 with threeparameters: average household income, mortgage interest rates and regulatory changes. We find householdlending capacity is a more accurately predictor of house prices in the Netherlands than LTV ratios and ‘debt-to-GDP’ ratios. In a test on out-of-sample data, a univariate OLS model with household lending capacityprovides the most robust forecasts.The next section provides an analysis of the formulas that govern access to credit in the Netherlandsand how we model these formulas. In the methods section, we describe our dataset and the specification ofstatistical models. Lastly, we report our findings and discuss the implications and limitations of our approach.
The notion that household lending capacity (
HLC ) - the amount of mortgage debt one can legally borrowto finance a home - influences house prices is not new. Economists generally believe that the ability toafford loans is one of the channels through which income and interest rates affect prices (ESB, 2017). We Such studies have been conducted for the United States (Duca and Muellbauer, 2016), Ireland (Lyons, 2018),Finland (Oikarinen, 2009), Norway (Anundsen and Jansen, 2013), France (Chauvin and Muellbauer, 2014) and Sweden(Turk, 2016). The LTV is the mortgage for a property divided by the price paid. A low LTV means that the lender left a largemargin of safety between the mortgage and the market value of the home. Thus, the value of the collateral would begreater than the mortgage even if the house decreases in value. A high LTV indicates that lenders are willing to toleratemore risk. ypothesise that the details of how HLC is calculated matter. The rules that govern household borrowingrelative to income are called ‘Loan to Income’ (LTI) formulas. This section is divided in two parts: anintroduction to LTI formulas in the Netherlands and an elaboration on how we model LTI formulas.
Year A v e r age S a l e P r i c e Year G D P i n m illi on s o f E u r o s Year I n t e r e s t R a t e s Figure 1: Statistics From The Central Bureau of Statistics (CBS) in the Netherlands 1995-2018
Before 2009, the Netherlands had no national laws or regulations that restricted lending capacity. However,the associations of banks and insurers - the dominant mortgage providers - authored the ‘Gedragscode Hy-pothecaire Financieringen’ (Behavioral Code Mortgage Finance, GHF) in 1999. GHF specified norms thatthe members agreed to adhere to for fair and responsible mortgage finance, including norms pertaining tothe height of mortgages. The association specified that households would not be allowed to spend more thana specific percentage of their available income on mortgage repayment, depending on their family type andincome bracket (NIBUD, 2018). For the average family, the maximum percentage of income designated forincome (“woonquote”) was a range of 21-40% depending on their economic circumstances. Banks wereallowed to deviate from these norms in exceptional cases, but the Dutch regulator found that the norms weremostly adhered to (De Nederlandsche Bank, 2009). The norms were tightened in response to the FinancialCrisis and translated into law in 2011 (Rabobank, 2014, 2015).The housing market faced a number of reforms in the period 2011-2018. One key reform directly im-pacted lending capacity. After the advent of the Financial Crisis of 2007 - in which mortgage debt played akey role in the American crisis and to a lesser extent the near collapse of Dutch banks - the Dutch governmentaimed to decrease the burden of mortgage debt on the economy (International Monetary Fund, 2019). TheDutch Central Bank identified a category of mortgage products - ‘krediethypotheken’ (interest-only mort-gages) - as a source of increasing mortgage debt. Interest-only mortgages grew their market share from10% to 50% in the period 1995-2008 (De Nederlandsche Bank, 2009). Customers with interest-only planscould choose to pay off their mortgage at their own pace or not at all, whereas traditional annuity and linearpayment plans require households to amortise their mortgage every month. The fact that borrowers onlyhad to budget interest payments meant that households could borrow more with an interest-only mortgagethan with a traditional annuity. As of 2011, the government required lenders to use the maximum annuity asthe maximum mortgage, even for an interest-only mortgage (Centraal Plan Bureau, 2018; Rabobank, 2014). s of 2013, the government no longer allowed households to deduct interest payments from taxes for newinterest-only mortgages (De Nederlandsche Bank, 2019b). This removed a crucial fiscal benefit from the cat-egory and made interest-only mortgages more expensive than annuity mortgages . The next section detailshow the transition between regulatory regimes is modelled. LTI formulas govern how much a household can borrow given their income and mortgage interest rates.Lenders use formulas to calculate the legal limit that households can borrow based on income, interest rates,risk of default, expenses, debts, housing costs and more. We boil the formulas down into a single equationfor annuities and interest-only mortgages. LC k = I h r + c where I h is annual income allocated to housing expenses, r is annual interest rates and c is a fixed annual costexpressed as a percentage of the value of the home. The effect of interest rate changes on lending capacityis neither linear nor independent of income. This is significant, because independence and linearity are keyassumptions in most regression models. This is also true of annuities. The maximum annuity mortgage isthe mortgage for which households can pay the yearly interest and mandatory amortisation. The simplifiedequation for the maximum annuity mortgage LC a reads: LC a = I h ∗ f ( r + c ) where I h is income allocated to housing expenses, r is interest rates and c is a fixed annual cost expressed asa percentage of the value of the home. Function f ( x ) is the standard annuity formula: f ( x ) = − ( + x ) ( − ) ) x Annuity mortgages are calculated differently than interest-only mortgages, so they produce differentlending capacities for the same income and interest rates. In summary, the LTI formulas are non-linearfunctions in which the effect of a change in income and interest rates is dependent on the value of the othervariable.
Given that the dependencies and non-linearities cannot be modelled accurately in linear regression models,we construct the measure ‘average household lending capacity’ (
HLC ). The following section has threepurposes. First, it lays out the variables and their summary statistics. Second, it details the construction ofthe key variable
HLC . Finally, it describes the models and model evaluation metrics. The Dutch government decided to gradually phase out the mortgage interest deduction altogether in 2014. .1 Data We primarily utilise publicly available income and price data from Dutch Central Bureau of Statistics (CBS)and the ‘De Nederlandsche Bank‘ (DNB), the Dutch central bank. Summary statistics can be found in Table1; details on sources and variable construction in the Appendix.
Table 1: Summary StatisticsStatistic N Mean St. Dev. Min Pctl(25) Pctl(75) MaxAverage House Price 92 201.180 50,896 89.792 179.627 237.662 267.464Household Income 91 68.733 13.467 44.916 59.815 77.209 98.783Interest Rates 92 5,382 1,338 2,840 4,728 6,325 8,400Avg. LTV 88 101,3 2,6 96 100,2 103,4 103,9Interest-Only Marketshare 92 0,217 0,181 0,000 0,000 0,410 0,463
Three model are tested: (1) a benchmark model with LTV values as a measure of credit access, (2) a univari-ate model with average household lending capacity (
HLC ) and (3) a benchmark model plus
HLC . Each ofthe models has house prices ( HP ) as the dependent variable. The β coefficients are optimised to minimisethe squared prediction error (see Section C.5.1). The benchmark model is specified as follows: HP t = β + β I t + β r t + β LTV t where I is quarterly income, r are annual interest rates and LTV is the average ’loan-to-value’ ratio of firsttime buyers in quarter t . HLC
Model
The univariate model only has
HLC as a lagged independent variable.
HLC is the amount banks allowed theaverage household to borrow at time t : HP t = β + β HLC t − i where HLC is constructed as a weighted average between the maximum interest-only mortgage (
HLC k ) andthe maximum annuity mortgage ( HLC a ): LC = mHLC k + ( − m ) HLC a HLC k = IW ( − x ) r + cHLC a = I ∗ W ∗ a (( − x ) r + c ) where m represents the market share of interest-only mortgages (see Section C.4), I is quarterly income , W is the share of income used for housing expenses (see Section C.2), x is the tax rate at which households candeduct interest (see Section C.1), r are interest rates and c is annual costs of maintaining the house expressedas a percentage of the purchase value (see Section C.3). The lagged, unfitted variable HLC is shown inFigure 2. The similarities in variance between
HLC and house prices is evident.
Year H L C / A v e r age H ou s e P r i c e Figure 2: The correlation between the average household lending capacity (black) and averagehouse prices (purple) is strong.
HLC
Model
The final model includes both benchmark variables and
HLC . As
HLC is calculated based on income andinterest rates, it naturally correlates highly with those variables. To reduce the cross-correlation, we convert
HLC by dividing
HLC by income I . This creates a new variable that expresses the ratio between lendingcapacity and income. This variable reflects the multiple of their disposable income that the average householdcan borrow. The benchmark model including HLC has the form: HP t = β + β I t + β r t + β LTV t + β HLC t − i I t − i I is multiplied by 4 to obtain the yearly income or divided by 3 to obtain monthly income. .3 Fitting Approaches The model specifications above have the form of an Ordinary Least Squares regression. OLS is the pocketknife of econometric modelling. It fits a minimum number of parameters and is highly interpretable butdoes not correct for auto correlation. The standard model in the literature that corrects for autocorrelationis the Error Correction Model, a more complex regression model that distinguishes between longterm andshort-term predictors (Anundsen and Jansen, 2013; Turk, 2016). ECM has the following generic form: ∆ y = β + β ∆ x , t + ... + β i ∆ x i , t + γ ( y t − − ( α x , t − + ... + α i x i , t − )) Predicting ∆ y instead of y indicates that we predict the change in y rather than its absolute value. ECM wasdeveloped for econometric time-series analysis of variables that have trends on both the short and long term.ECM combines three estimates: it simultaneously estimates y t − based on long-term predictors, ∆ y based onshort-term predictors and γ to weight the longterm and short-term estimates. The strength of this model isthat it anticipates the existence of long and short term trends and adjusts for auto correlation. However, thefact that the ECM model is much less sparse and includes a past value of the dependent variable in the fittingprocess makes it less persuasive. As it is the standard test in the literature, we fit all models as both OLS andECM. We measure the quality of fit of every model with the Root Mean Squared Error (RMSE) and the MeanAbsolute Error (MAE). We evaluate the tendency of the model to overfit spurious correlations in the databy performing two sets of fits. We also conduct an ‘out of sample‘ test, where the models are fitted only ondata up to the second quarter of 2008. We choose that cut-off because it is the quarter in which the Dutcheconomy went into recession. The purpose of the test is to ascertain whether models are able to anticipatethe coming crisis based on data from another part of the business cycle.
In this section, we report the goodness of fit of 12 variants of the three model specifications above. We fitevery model specification as an OLS and as an ECM on the all quarters and only quarters up to 2009. Tables2, 3 and 4 present the measured quality of fit of all variants. Figure 5 displays the fitted values of the OLSmodels. We focus on the OLS results because the ECM models failed to generalise across the board. Afigure with all the fitted values from the ECM models can be found in Appendix E.
Year ( P r ed i c t ed ) A v e r age H ou s e P r i c e Year ( P r ed i c t ed ) A v e r age H ou s e P r i c e HLC
Year ( P r ed i c t ed ) A v e r age H ou s e P r i c e Year ( P r ed i c t ed ) A v e r age H ou s e P r i c e Benchmark incl.
HLC
Year ( P r ed i c t ed ) A v e r age H ou s e P r i c e Year ( P r ed i c t ed ) A v e r age H ou s e P r i c e Figure 3: Model Performance (OLS). These six diagrams displays the fitted values of regressionmodels (black) contrasted with observed average house prices (purple). Fitted values on the leftwere generated by models fit on all the available data whereas fitted values on the right were gen-erated by models trained only on quarters up to mid-2008. While the all three model specificationsseem reasonably accurate when fit on all data, the ‘out-of-sample‘ test reveals that only the
HLC model forecasts accurately on unseen data. 8able 2: Benchmark ModelModel Type RMSE MAE
Model: OLS
Fit on Quarters up to 2018 10.150 5.109Fit on Quarters up to 2009 31.173 8.146
Model: ECM
Fit on Quarters up to 2018 8.057 5.152Fit on Quarters up to 2009 13.864 3.868Table 3: Household Lending Capacity (
HLC ) ModelModel Type RMSE MAE
Model: OLS
Fit on Quarters up to 2018 7.451 4.891Fit on Quarters up to 2009 7.495 4.854
Model: ECM
Fit on Quarters up to 2018 6.700 4.681Fit on Quarters up to 2009 13.161 7.745Table 4: Benchmark Plus
HLC
ModelModel Type RMSE MAE
Model: OLS
Fit on Quarters up to 2018 5.375 3.606Fit on Quarters up to 2009 21.210 5.744
Model: ECM
Fit on Quarters up to 2018 5.284 3.989Fit on Quarters up to 2009 8.664 2.9949 he most striking finding is that all but one models show poor performance in the out-of-sample test.Only the OLS fit of the
HLC model achieves comparable performance on the out-of-sample quarters andthe in-sample quarters. By contrast, the benchmark models completely miss the direction of market in theout-of-sample quarters. The increase in robustness does not come at the expense of accuracy. When fit onall quarters, the OLS
HLC model has a 27% lower MAE than the OLS benchmark model. The models thatinclude both benchmark variables and
HLC have the best fit when fit on all quarters, but suffer from the samelack of robustness as the benchmark model.
Economists agree that access to credit is a crucial determinant of house prices. This paper suggests a newapproach to model credit conditions by modelling the structure of Loan-to-Income norms in the Netherlands.By modelling the formulas that Dutch banks use to calculate how much they can lend to households, weconstruct a regression model that does not overfit in the ‘out-of-sample‘ test. We compare the accuracy ofthe model with existing measures of credit access and find it delivers more accurate results. Our analysisindicates that the cause of the Dutch crisis was a Dutch phenomenon: the rise of interest-only mortgagesand their fall at the hands of the Dutch government in 2011. As the rise and fall of these mortgage productsis correlated with rising and falling LTV rates, this effect is partly represented in the benchmark model, butmodelling it explicitly allows for a better fit. The findings suggest that the ‘double dip‘ crisis was partly dueto the fact that Dutch government constricted lending capacity in 2011 by regulating interest-only mortgagesmore strictly. Whereas all economic indicators pointed towards price recovery, prices fell steeply soon afterthis legislation came into effect.More significantly, the model specification based on LTI formulas appears to generalise more robustlywhen only trained on a fraction of the available data. This observation suggests that the in-sample accuracyof econometric house price models may be a poor measure of their predictive power. These models maybe fitting spurious correlations that do not generalise to other economic circumstances. The field faces achallenge: how do we develop models that account for complex interactions between economic variables(low bias) whilst safeguarding against overfitting (low variance)? The methodology of this work lays out apossible path forward. We identify relationships between variables in qualitative research and construct ameasure based on those relationships for quantitive analysis. Rather than fit each variable independently , wefit a model where a function based on the independent variables is fitted . In this work, the function calculateshousehold lending capacity based on the formulas that Dutch banks use in the mortgage application process.Others authors can represent any hypothesised interaction between variables as functions. Thus, one canmodel any complex relationship between variables, hence lowering bias, whilst fitting fewer parameters,hence lowering variance. This route might offer an alternative route to the current direction of the field: As a robustness test, mortgage debt-to-GDP ratios were also included in the benchmark. This version of thebenchmark scored more poorly than the benchmark with LTV values. In the general form y = β + β x + ... + β i x i . In the general form: y = β + β f ( x , ..., x i ) lugging an increasing number of variables into increasingly complex statistical models. Competing interests
The authors declare that they have no competing interests.
Funding
TY was partially supported by the Alan Turing Institute under the EPSRC grant no. EP/N510129/1. Thesponsor had no role in study design; in the collection, analysis and interpretation of data; in the writing ofthe report; and in the decision to submit the article for publication.
Authors’ contributions
MS analyzed the data. MS and TY designed the study and the analysis. All authors read and approved thefinal manuscript.
References
Anundsen AK, Jansen ES (2013). “Self-reinforcing effects between housing prices and credit.”
Journal ofHousing Economics , (3), 192–212.Bernanke BS, Gertler M (1995). “Inside the black box: the credit channel of monetary policy transmission.” Journal of Economic perspectives , (4), 27–48.Centraal Bureau voor de Statistiek (2019a). “Bestaande koopwoningen.” [Online; ac-cessed 29-July-2019], URL https://opendata.cbs.nl/statline/ .Centraal Bureau voor de Statistiek (2019b). “Inkomen Huishoudens.” [Online; accessed 29-July-2019],URL https://opendata.cbs.nl/statline/ .Centraal Plan Bureau (2014). “Staat van de Woningmarkt, Jaarrapportage 2014.” URL . A possible critique of this method would be that it invites researchers to formulate endlessly complex versionsof f until they find some function that fits the dependent variable. To counter this, researchers should be required tocarefully justify their hypothesised function based on qualitative research. In this work, we base the construction of HLC on the LTI formulas of Dutch banks. Both the theoretical justification of the construction of f and the empiricalvalidation must be persuasive. entraal Plan Bureau (2018). “Doorrekening varianten aanpassing aflossingseis.”Chauvin V, Muellbauer J (2014). “Consumption, household portfolios and the housing market: a flow offunds approach for France.” In “Lyon Meeting,” .De Nederlandsche Bank (2009). “Risico op de Nederlandse Hypotheekmarkt.” [Online; ac-cessed 29-July-2019], URL .De Nederlandsche Bank (2019a). “Deposito’s en leningen van MFI’s aan huishoudens, rentepercentages.”[Online; accessed 29-July-2019], URL https://tinyurl.com/y8xkgzm4 .De Nederlandsche Bank (2019b). “Effecten van een verdere verlaging van de LTV-limiet.” .Duca J, Muellbauer J (2016). “How Mortgage Finance Reform Could Affect Housing †.” (5), 620–624.ESB (2017). “Lenen om te wonen.” ESB Dossier , (4749).Groot S, Vogt B, Van Der Wiel K, Van Dijk M (2018). “Oververhitting op de Nederlandse huizenmarkt?”International Monetary Fund (2019). “Country Report: Kingdom of the Netherlands.” (19).Kiyotaki N, Moore J (1997). “Credit cycles.” Journal of political economy , (2), 211–248.Lyons RC (2018). “Credit conditions and the housing price ratio: Evidence from Ireland’s boom and bust.” Journal of Housing Economics , , 84–96. ISSN 10960791. doi:10.1016/j.jhe.2018.05.002.NIBUD (2018). “Financieringslastnormen 2018.” [Online; accessed 29-July-2019], URL .Oikarinen E (2009). “Household borrowing and metropolitan housing price dynamics - Empirical evidencefrom Helsinki.” Journal of Housing Economics , (2), 126–139. ISSN 10511377. doi:10.1016/j.jhe.2009.04.001. URL http://dx.doi.org/10.1016/j.jhe.2009.04.001 .Rabobank (2014). “De gevolgen van de terugkeer van de annuïteitenhypotheek.” [Online;accessed 29-July-2019], URL https://economie.rabobank.com/publicaties/2014/maart/de-gevolgen-van-de-terugkeer-van-de-annuiteitenhypotheek/ .Rabobank (2015). “Ene LTV is de andere niet.” [Online; accessed 29-July-2019], URL https://economie.rabobank.com/publicaties/2015/maart/de-ene-ltv-is-de-andere-niet/ .Turk R (2016). “Housing Price and Household Debt Interactions in Sweden.” IMF Working Papers , (276),1. ISSN 1018-5941. doi:10.5089/9781513586205.001.van der Veer, Hoebrechts (2016). Journal of Banking and Finance . doi:10.1007/s11146-015-9531-2. Summary of Dutch Housing Market History
This section summarises the Dutch story, highlights existing theories and identifies the gap in understanding.The Dutch have consistently seen house prices rise in the post-war period. There has a been a near-permanentshortage of residential properties, so supply rarely meets demand. In a country with one of the highestpopulation densities in the world, population growth and rising incomes, households expect that residentialproperty values will continue to rise. This makes property a good investment - especially in popular citiessuch as Amsterdam, Utrecht and The Hague.Figure 1 shows that house prices rose rapidly in the 1990’s and early 2000’s and entered a period ofsustained decline after 2009. Despite falling interest rates, housing prices fell 16% from their peak value.The crisis came at great cost to many Dutch families. In 2015, 28% of homes were deemed ‘under water’: thehome was worth less than the outstanding mortgage debt (Centraal Plan Bureau, 2014). Families could notmove to a new house because they could not pay off the debt on their old property. Divorcees, the recentlyunemployed and other citizens with an urgent reason to sell their house could not fetch a price higher orequal to their mortgage. As a result, these citizens were left with substantial and long-term debts without themeans to repay them. Often, the costs of new housing and debt payments were more than these householdscould afford.The Dutch government attempted to boost home prices by lowering taxes on real-estate transactionsand allowing parents to assist their children to buy property by allowing a tax-free gift for the purpose ofacquiring a property. Interest rates fell to record lows (De Nederlandsche Bank, 2019a). However, thehousing market continued to slump. After years of falling prices, the market rose in 2015. Prices shot backup beyond their previous peak within a few years. In the large cities, prices rose with 10% per year, leadingthe Dutch Central Bank to call markets “overheated” that had been ‘cold’ just a few years before. The DutchCentral Bank expressed concern about the unprecedented speed of price increases (Groot et al. , 2018).The question remains which factors drove this dramatic cyclical movement in the market. Figure 1 showsthat The Netherlands experienced a ‘double dip’ in prices: prices fell in 2008 and 2009, stabilized from 2009to 2011 and sank further in 2013. The second dip occurred despite the fact that Gross Domestic Product hadresumed to grow (see Figure 1). Research remains to be done why the Dutch market experienced a ‘doubledip’ that prolonged the crisis on the housing market.The International Monetary Fund, the Dutch National Bank and the Centraal Plan Bureau have performeda great number of analyses of the housing market. These institutions express concern that Dutch householdsare amongst the most indebted in the world (International Monetary Fund, 2019). The bulk of householddebt consists of mortgage debts. Dutch households borrow because of fiscal incentives to do so. Householdseither rent or own property. Buying property with a mortgage is attractive, because monthly payments arepartially a form of savings, whereas rent is not. The Netherlands makes it fiscally attractive for householdsto switch from renting to owning property. This has three causes. The first is that renters in the middleand upper class pay some of the highest rents in Europe . The second is that the government allows home- This is a reason that supply-side factors are less relevant in the Dutch market. The cause of high rents in the Netherlands is a topic of much debate and beyond the scope of this work. wners to deduct interest from their taxable income, effectively subsidising mortgage debt. The third cause isthat the Dutch government guarantees mortgage debt for homes valued up to e Data Sources
B.1 House Prices
We collect the average transaction price of residential properties in the Netherlands ( HP ) as reported on aquarterly basis by the Dutch statistics office (Centraal Bureau voor de Statistiek, 2019a). Values before 1999were only reported on a yearly basis. For these years, we imputed the missing quarters by extrapolating thetrend between years. While this approach may miss quarterly variation, it captures the overall trend in prices. B.2 Household Income
Household Income after Taxes ( I ) is average after-tax household income as reported by the Dutch statisticsoffice (Centraal Bureau voor de Statistiek, 2019b). The variable has strong yearly seasonality, becausebonuses are added to income in the last quarter. Therefore, we smooth the variable by taking the average ofthe last 4 quarters. As a result, changes in the variable show changes in year-over-year income. B.3 Interest Rates
Banks calculate lending capacity using the nominal interest rate for the mortgage product. Mortgage interestrates differ from the interest rates on the capital markets, because banks factor in their margin, estimated riskof default, risk of devaluation of the property and future developments of the interest rates. Every bank setsdifferent interest rates for their products. We obtained the average mortgage interest rate for new contracts ona quarterly basis from the Dutch Central Bank for the period 1992-2019 (De Nederlandsche Bank, 2019a).We smooth the variable by taking the average of the last 4 quarters. Changes in the variable show changesyear-over-year.
B.4 Average LTV Values
Average LTV ratios were collected from reports by the Dutch Central Bank and the Rabobank (De Ned-erlandsche Bank, 2019b; Rabobank, 2015). The data was reported on a yearly level, hence we miss somequarter-to-quarter variance. Where data was missing, we imputed the value from the year before.
B.5 Market Share of Interest Only Mortgages
Market Share of Interest-Only mortgages was collected from two sources: a report by the Dutch CentralBank up to 2008 (De Nederlandsche Bank, 2009) and a Rabobank report up to 2011 (Rabobank, 2014) afterwhich the category became irrelevant. Setting Constants in Model Construction
C.1 Accounting for Mortgage Interest Deduction ( t ) The Dutch government allows to deduct mortgage payments from income tax. Thus, the effective interestpayed is a fraction of the nominal interest rate. For example, a household might pay a top rate of 50% taxover their income. If they have e e e t is set at 0.4. C.2 Share of Income for Housing Expenses ( W ) The ‘woonquote’ governs the share of income the bank assumes that buyers can use to pay for the costsassociated with their home. The precise ‘woonquote’ used for the average Dutch household is not known,because the rules and regulations only defines a range of 21% to 40% (NIBUD, 2018). Banks can choosea value within that range based on a large number of variables not available to me, such as family size,projected (energy) costs of the home, projections of future income, etc. It is possible that the ‘woonquote’was not the same for interest-only mortgages and annuities, because these mortgages may have been chosenby differing groups whose characteristics put them elsewhere on the range. Without any hard data, we chooseto set the parameter in the middle of the range: 30%. If the coefficient is not set appropriately, this will becompensated for in the estimation of β . C.3 Setting Maintenance Cost ( c ) LTI limits specify that homeowners must be able to pay all their costs from the portion of their incomeallocated for housing. We set the ‘other costs‘ parameter at 2,5% of the value of the home. The NIBUD -the government agency that sets LTI rules - estimates ‘other costs’ to be a significant fraction of the initialcost of the home. This matches a back-of-the-envelope calculation: homeowners face a tax of 0,75% of theirhome value, municipal taxes and approximately 1% yearly costs for maintanance. .4 Constructing Marketshare ( m ) For the construction of
HLC , we require the market share of interest-only mortgage over time. That data wasnot available over the entire time period. However, we estimate the market share of interest-only mortgagesin new transactions based on the changes in the total stock of Dutch mortgages. The rest of this section laysout how this measure was constructed.We can approximate the share of households who got a new mortgage by dividing the number of markettransactions by the total number of households for every year. This assumes that no household movedtwice. Based on the overall marketshare data, we calculate the increase in market share for interest-onlymortgages. If we know that 5% of households moved in a given year and interest-only mortgages captured2,5% of the market, we can deduce that 50% of households who moved switched from an annuity-likemortgage to an interest-only mortgage. What about the remaining 50%? These must be non-switchers. Wecan assume that non-switcher households renewed the mortgage type they already possessed. Therefore, in acase where interest-only mortgages have a marketshare of 40%, we can deduce that the share of non-switcherhouseholds who renewed an interest-only mortgage was the proportion of non-switchers times the overallmarket share (0 . ∗ . = . C.5 Calculating Optimal Lag for
H LC
This section concerns itself with choosing an appropriate time-lag for
HLC in our regression analysis. Intime-series analysis, the causal relationship between variables might be difficult to observe because of atime-lag. In those cases, we can shift the independent variables backward in time. If we believe that lendingcapacity today will influence housing prices 2 years from now, we regress lending capacity today againsthousing prices in two years. Timelags - like most other causal relationships - can be chosen on purelytheoretical grounds or inferred from the data. we employ a common, hybrid approach. We first specify arange of possible lags based on theory. This assures that any of the lags would accord with the theoreticalconception of the behaviours of home-buyers. We then test empirically which lag best fits the data. Weregress the different lagged versions of
HLC against house prices and take the R to measure goodness of fit.We choose the lag with the highest R .We chose a range of 0-6 quarters to lag HLC by, because homeowners often arrange financing first andthen acquire a home. The process of acquiring a home often involves bidding, bargaining, a waiting periodfor the owners to find a new home and finally the formal transaction. That process can take many months.Thus, we choose the range of 0-6 quarters as possible lags between lending capacity and house prices. Weregressed lending capacity lagged by 0-6 quarters against house prices and observed the R . The R is highestwith a time-lag of 6 quarters. We also tested lags greater than 6 and found the R is lower with a lag of 7 and 8 quarters. .9400.9450.950 2 4 6 lag r Figure 4: R of HLC with different lags.
C.5.1 Choosing the number of coefficients to estimate
Originally, we intended to fit separate coefficients for
HLC a and HLC k . However, we found that - remarkably- the value of both coefficients was nearly identical (on three decimals) for the optimal fit. This suggests thatthe best fit parameter values of constants c , t and W that are identical for HLC a and HLC k . For the sake ofrepresenting our model as clearly as possible, we therefore chose to simplify the model and only formallyestimate one coefficient on HLC in the regression analysis. However, for other values of c , t and W or inother markets, it may be necessary to fit separate β ’s for different mortgage product types. Model Fits
Table 5: OLS Results
Dependent variable:HP (1) (2) (3) I ∗∗∗ ∗∗∗ (0.218) (0.073) LTV ∗∗∗ ∗∗∗ (519.257) (326.972) r − ∗∗∗ (2,167.802) HLC t − ∗∗∗ (0.058) HLCI t − ∗∗∗ (4,132.603)Constant − ∗∗∗ − ∗∗∗ − ∗∗∗ (53,894.820) (10,537.810) (30,881.110)Observations 82 81 78R ∗∗∗ (df = 3; 78) 1,590.978 ∗∗∗ (df = 1; 79) 1,180.502 ∗∗∗ (df = 3; 74) Note: ∗ p < ∗∗ p < ∗∗∗ p < Dependent variable: ∆ HP (1) (2) (3) ∆ I ∆ LTV − − ∆ r − ∆ HLCI t − ∗ (12,426.520) I t − − ∗∗∗ (0.219) (0.371) LTV t − r t − − ∗∗∗ − ∆ HLC t − ∗∗∗ (0.239) HLC t − ∗∗∗ (0.092) HLCI t − ∗∗∗ (7,095.525) γ − ∗ − ∗∗∗ − ∗∗∗ (0.053) (0.065) (0.080)Constant 468.108 − − ∗∗ (38,566.940) (3,583.361) (44,634.110)Observations 83 83 83R ∗∗∗ (df = 7; 75) 12.188 ∗∗∗ (df = 3; 79) 4.844 ∗∗∗ (df = 9; 73) Note: ∗ p < ∗∗ p < ∗∗∗ p < Year ( P r ed i c t ed ) A v e r age H ou s e P r i c e Year ( P r ed i c t ed ) A v e r age H ou s e P r i c e HLC
Year ( P r ed i c t ed ) A v e r age H ou s e P r i c e Year ( P r ed i c t ed ) A v e r age H ou s e P r i c e Benchmark incl.
HLC
Year ( P r ed i c t ed ) A v e r age H ou s e P r i c e Year ( P r ed i c t ed ) A v e r age H ou s ee
Year ( P r ed i c t ed ) A v e r age H ou s e P r i c e Year ( P r ed i c t ed ) A v e r age H ou s ee P r i c ee