Analysis of Advisor Portfolio using Multivariate Time Series and Cosine Similarity
AAdvisors intelligence throughsales analytics
A dissertation submitted in partial fulfillmentof the requirements for the degree of
Master of Technology in Computer Science & Engineering with specialization in Big Data byGayatri Pradhan
May 2017 a r X i v : . [ q -f i n . GN ] J u l eclaration I hereby declare that the dissertation
Advisor Intelligence through Purchase Pat-terns and Sales Analytics submitted by me to the School of Computing Science andEngineering, VIT University Chennai, 600 127 in partial fulfillment of the requirementsfor the award of
Master of Technology in Computer Science & Engineering withspecialization in Big Data is a bona-fide record of the work carried out by me underthe supervision of
Prof. SyedIbrahim S. P .I further declare that the work reported in this dissertation, has not been submittedand will not be submitted, either in part or in full, for the award of any other degree ordiploma of this institute or of any other institute or University.Sign:Name & Reg. No.:Date: chool of Computing Science & Engineering
Certificate
This is to certify that the dissertation entitled
Advisor Intelligence through Pur-chase Patterns and Sales Analytics submitted by
Gayatri Pradhan (Reg. No.15MCB1008) to VIT University Chennai, in partial fulfullment of the requirement forthe award of the degree of
Master of Technology in Computer Science & Engi-neering with specialization in Big Data is a bona-fide work carried out under mysupervision. The dissertation fulfills the requirements as per the regulations of this Uni-versity and in my opinion meets the necessary standards for submission. The contentsof this dissertation have not been submitted and will not be submitted either in part orin full, for the award of any other degree or diploma and the same is certified.
Supervisor Program Chair
Signature: .................... Signature: ....................Name: .................... Name: ....................Date: Date:
Examiner
Signature: ....................Name: ....................Date: (Seal of the School)ii bstract
In mutual fund, an individual or a firm that is in the business of giving advice aboutsecurities to clients is an investment advisor. Investment advisers are individuals or firmsthat receive compensation for giving advice on investing in stocks, bonds, mutual funds,or exchange traded funds. Investment advisors manage portfolios of securities. Advisorscan use new cognitive and analytics capabilities to better understand their clients andtheir needs and have a stronger ability to deepen relationships with a better portfolio. Inthis paper, we analyze data points for each advisor, and distinguish the best prospects,obtain insight into their experience and credentials, and learn about their portfolio, inother words to recognize the pattern of portfolio of the advisors. Such analysis helpsthe sales people to sell the fund company products to the suitable advisors based onthe nature of the product they want to sell. This is done by investigating what kind ofproducts advisors have been buying, and what kind of products they might be lookingfor. This helps to increase the sales of the products as sales people will be reaching theappropriate advisors. cknow ledgements
I wish to express my sincere thanks to Dr.G.Viswanathan, Chancellor, Mr. SankarViswanathan, Vice President, Ms. Kadhambari S. Viswanathan, Assistant Vice Presi-dent, Dr. Anand A. Samuel, Vice Chancellor and Dr. P. Gunasekaran, Pro-Vice Chan-cellor for providing me an excellent academic environment and facilities for pursuingM.Tech. program. I am grateful to Dr. Vaidehi Vijayakumar, Dean of School of Com-puting Science and Engineering, VIT University, Chennai and to Dr. V. Vijayakumar,Associate Dean. I wish to express my sincere gratitude to Dr.Bharadwaja Kumar G,Program chair of M.Tech CSE with Specialization in Big Data for providing me an op-portunity to do my project work. I would like to express my gratitude to my internalguide Prof. SyedIbrahim S. P and my external guide Mr.Inigo Fernando who inspiteof their busy schedule guided me in the correct path. I am thankful to Broadridge Fi-nancial Solutions, Hyderabad for giving me an opportunity to work on my project andhelped me gain knowledge. I thank my family and friends who motivated me during thecourse of the project work. v ontents
Declaration iCertificate iiAbstract ivAcknowledgements vList of Figures viiiList of Tables ix1 Introduction 12 Literature Survey 6 ontents vii3.4 Module 3: Predicting Advisors Performance . . . . . . . . . . . . . . . . . 173.4.1 Fundamental Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 173.4.2 Technical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 173.4.3 Time Series Forecasting . . . . . . . . . . . . . . . . . . . . . . . . 18. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 ist of Figures ist of Tables or/Dedicated to/To my. . . x hapter 1 Introduction
The financial services industry was as of late recognized as the business well on the wayto be disturbed and changed by millennial in the US. There are comparable signs atthe worldwide level. The adjustments in the keeping money and budgetary administra-tions industry in the coming years will be seismic. For instance, worldwide banks aresetting up development focuses and concentrated groups to concentrate on block-chain,proclaimed as a problematic constrain that offers various open doors, for example, re-designing existing keeping money foundation, accelerating settlements and streamliningstock trades.Financial institutions are in effect consistently tested by contracting incomes and needto enhance operational cost efficiencies. Rising fintech new businesses and officeholderinnovation monsters are conveying new plans of action bringing on disturbance andtesting customary managing an account plans of action. Controllers in all topographiesare requesting stricter consistence and more grounded money related train, from theFederal Reserve in the US to the European Banking Authority (EBA) in the EU, thePrudential Regulation Authority (PRA) and the Financial Conduct Authority (FCA)in the UK.Financial institutions have the advantages of vast client bases and access to rich value-based information. Making more current business models or frameworks that use theaccessible information permits money related foundations to adapt information to conveyprevalent client esteem.Winning in this dynamic market will be supported by how the money related establish-ments can get an incentive from information. The union of machine and human insightis disturbing customary basic leadership by outfitting associations with learning and1hapter 1.
Introduction • To a great degree substantial informational indexes to analyze to reveal patterns,trends and correlations • Real-time predictive and prescriptive analytics for driving profound significant bitsof knowledge • Hazard and consistence requesting convenient accessibility of dependable, quan-tifiable and secure information • Appropriation of machine learning and intellectual abilities • Democratization of information empowering more self-administration • Consumerization of BI through best-of-breed information revelation, investigationand visualization tools • Advanced stages controlled by 360 perspective of clients • Data and BI scene change and modernization to decrease cost and grasp new-agetechnologies • Explanatory ace information administration abilities • Fortifying information administration capacitiesAmong financial specialists owning common store offers, the greater part of it holdsthe reserve shares through a middle person, for example, an agent merchant, bank,subsidize market or stage, insurance agency, speculation consultant. Speculators pickdelegate which best suits their necessities.Financial specialists utilize go-betweens to acquire various advantages. Speculators reg-ularly indicate premiums to the go-between for proposals on the best way to contributetheir cash. With the assistance of the venture counsel, the speculator may choose tohapter 1.
Introduction
Figure 1.1:
Role of Advisors.
Big data analytics helps to learn the portfolios of the investment advisors and recognizethe behavioral pattern of investment advisor. Such analysis will help the sales person orthe fund companies to find relevant advisor which will be beneficial for the sales personto increase the sales of his products.Further this recognized pattern can be mapped into ten global broad Morningstar cate-gory groups (Equity, Allocation, Convertibles, Alternative, Commodities, Fixed Income,Money Market, Tax Preferred, Property, and Miscellaneous).The Morningstar Global Category assignments were introduced in 2010 to help investorssearch for similar investments entertained across the globe. There are different flavorsof funds and the investment advisors can be categorized into different categories. Finan-cial time series analysis is worried with hypothesis and routine of benefit valuation aftersome time. It is a profoundly exact teach, yet like other logical fields hypothesis framesthe establishment for making induction. There is, in any case, a key component thatrecognizes money related time arrangement investigation from other time arrangementexamination. Both money related hypothesis and its exact time arrangement contain acomponent of vulnerability. For instance, there are different meanings of benefit unpre-dictability, and for a stock return arrangement, the instability is not straightforwardlyperceptible. Statistical theory and methods play an important role in financial timeseries analysis. Our financial time series data is a multivariate time series data set. AMultivariate time series data analysis is used when one wants to model and explainhapter 1.
Introduction
Figure 1.2:
Work flow.
This similarity measure can vary from problem to problem. For example, as author[4] proposed a separation work in view of the accepted autonomous Gaussian modelsand utilized a hierarchical clustering strategy to gathering regularity groupings intoan attractive number of clusters. Here an independent Gaussian model is a distancefunction which is a similarity measure and model is build using hierarchical clustering.Similarly n number of analysis can be done on multivariate time series data. In thispaper few of the analysis has been done on multivariate time series data. The firstsection of the paper explains an approach towards clustering multivariate time seriesdata. This will cluster advisors with similar behavior into same clusters. The secondsection of the paper explains detecting leaders from correlated advisors among n numberhapter 1.
Introduction hapter 2
Literature Survey
In recent past, there is an increased interest in time series clustering research, particularlyfor finding useful similar trends in multivariate time series in various applied areas suchas environmental research, finance, crime, etc. . Clustering multivariate time series haspotential for analyzing large volume of finance data at different time points as investorsare interested in finding market trends of various funds such as Equity, Allocation,Convertibles, Alternative, Commodities, Fixed Income, Money Market, Tax Preferred,Property, and Miscellaneous so that it will help them to invest.
In this paper, a novel approach in light of element time wrapping and parametricMinkowski display has been proposed to discover comparable wrongdoing patterns amongdifferent wrongdoing groupings of various wrongdoing areas and consequently utilizethis data for future wrongdoing patterns forecast. Investigation on Indian wrongdoingrecords demonstrate that the proposed strategy by and large beats the current strategiesin grouping of such multivariate time arrangement information.
Singhalet. al. [3] has proposed a procedure in view of ascertaining the level of closenessbetween multivariate timeseries datasets utilizing two similitude elements. One closenesselement depends on vital segment investigation and the edges between the main segmentsubspaces while the other depends on the Mahalanobis remove between the datasets. The6hapter 2.
Literature Survey
Creator Pooya Sobhe Bidari [6] exhibited two stage practical grouping as another ap-proach in quality bunching for grouping time arrangement quality expression informa-tion. The proposed approach depends on finding utilitarian examples of time arrange-ment utilizing Fuzzy C-Means and K-implies calculations.Pearson association equivalence measure is used to isolate the expression cases of char-acteristics. In this approach, qualities are assembled by K-means and FCM methodsaccording to theirs time course of action expression, then cases of value direct are ex-pelled. By then, new components are described for the qualities and by figuring Pearsonassociation between’s new segment vectors, qualities with practically identical expressionlead are procured which can incite find interconnections between qualities.
For identifying environmental change in multivariate information Hardy Kremer [7] pro-poses novel bunching and grouping following procedures. In this novel grouping ap-proach, time arrangement is part into disjoint, break even with length interims and af-ter that thickness based subsequence bunching methodology is connected, and dynamictime twisting is utilized as a separation measure.
K. Kalpakis [37] concentrated the grouping of ARIMA time course of action, by usingthe Euclidean partition between the Linear Predictive Coding cepstra of two time-planas their uniqueness measure. The cepstral coefficients for an AR(p) time course of actionhapter 2.
Literature Survey
Yupei Lin [41] attempted to enhance the forecast precision with amending two inadequa-cies, sub interims neglecting to well speak to the information appropriation structuresand a solitary precursor figure the fluffy relationship in current fluffy time arrangementdisplay. To begin with, he distributed universe of talk in subintervals with the midpointsof two neighboring gatherings centers, and the subintervals are used to fuzzily the timecourse of action into cushy time plan. At that point, the fluffy time arrangement displaywith multi calculates high request fluffy relationship is developed to conjecture the sharetrading system. The outcomes demonstrated that the model enhanced the expectationexactness when contrasted and the benchmark show.
Xiong and Yeung [38] proposed a model-based method for gathering univariate ARIMAcourse of action. They expected that the time course of action are created by k un-mistakable ARMA models, with each model identifies with one gathering of interest.A desire expansion (EM) calculation was utilized to take in the blending coefficientsand additionally the parameters of the part models that boost the desire of the totalinformation log-probability. What’s more, the EM calculation was enhanced so that thequantity of groups could be resolved consequently.hapter 2.
Literature Survey This paper is essentially from the meaning of VAR model, as VAR model is one the mostimperative approaches to gauge advertise chance. In light of VAR model Xinjie Ma andYongsheng Yang [41] proposed a way to deal with anticipate the danger of China’ssecurities exchange by means of observational examination made on Shanghai Indexand picks five examples for portfolio hazard look into so it has incredible centralityto China’s stock exchange chance measure. The investigation of the venture chancesin the share trading system gives an incredible reference an incentive to the financialspecialists. hapter 3
Experimental Design & Setup
Figure 3.1:
Methodology of predicting Advisors behavior
Experimental Design & Setup A novel approach based on cosine similarity with hierarchical clustering model has beenproposed to find similar advisors.The Morningstar Category groupings were acquainted in 1996 with help speculatorsmake important correlations between mutual funds. Morningstar found that the specu-lation objective recorded in a reserve’s plan regularly did not satisfactorily clarify how thestore really contributed. For instance, many assets asserted to look for ”development,”however some of those were putting resources into set up blue-chip organizations whileothers were putting resources into little top organizations. The Morningstar Categorycharacterizations tackled this issue by breaking portfolios into associate gatherings inview of their holdings. The classifications help investors distinguish the top-performingfunds, evaluate potential risk, and construct very much well-diversified portfolios. Morn-ingstar routinely surveys the classification structure and the portfolios inside every clas-sification to guarantee that the framework addresses the issues of speculators. Morn-ingstar doles out classes to a wide range of portfolios, for example, common assets,variable annuities, and separate records. Portfolios are set in a given classification inview of their normal possessions insights in the course of recent years. Morningstar’sarticle group additionally audits and supports all class assignments. On the off chancethat the portfolio is new and has no history, Morningstar gauges where it will fall beforegiving it a more lasting classification task. Whenever vital, Morningstar may change aclass task in view of late changes to the portfolio.In clustering, Euclidean distance measure is the most commonly used for non-time seriesdata clustering. While working with financial multivariate time series data, it is notsuitable for multivariate time series clustering. Instead of Euclidean distance measurein stand-alone mode, cosine similarity provides better results. The problem of findingsimilar advisor can be solved in two steps:1. Hierarchical clustering.2. Computing cosine similarityWe now discuss each step in detail.
Clustering can be viewed as the most vital unsupervised learning issue; along these lines,as each other issue of this kind, it manages finding a structure in an accumulation ofhapter 3.
Experimental Design & Setup
Using cosine similarity, advisor with strong similarity can be evaluated. Later, this willhelp to detect the leader and the follower among the strongly similar advisors.Since similar behavior advisors are bucketed into one cluster, a sample portfolio fromeach cluster is considered to compute the cosine similarity among each advisors.Cosine similarity is a measure of comparability between two non-zero vectors of aninternal item space that measures the cosine of the edge between them. The cosineof 0 is 1, and it is under 1 for some other edge. It is in this manner a judgmentof introduction and not size: two vectors with a similar introduction have a cosinehapter 3.
Experimental Design & Setup cos ( t , t
2) = 1 . / ( | || | ) = 0 . Table 3.1:
Cosine Similarity table t1 t2 t3 t4t1 cos(t1,t1) cos(t1,t2) cos(t1,t3) cos(t1,t4) t2 cos(t2,t1) cos(t2,t2) cos(t2,t3) cos(t2,t4) t3 cos(t3,t1) cos(t3,t2) cos(t3,t3) cos(t3,t4) t4 cos(t4,t1) cos(t4,t2) cos(t4,t3) cos(t4,t4)This result will give group or pairs of advisors whose portfolio behaviors are highlycorrelated. This outcome will give gathering or matches of counsels whose portfolio practices areprofoundly associated.An augmentation to above got comes about prompted find pioneers among varioustime arrangement information. This issue of recognizing pioneer among guide withcomparable portfolio is comprehended by breaking down lead-slack relations among thehapter 3.
Experimental Design & Setup
The initial phase in identifying pioneer among set to time arrangement information isto process the slacked connection between’s each match of time arrangement.We propose to add up to the effects of various slacks and describe an amassed slackedassociation. The amassed slacked relationship figuring can be lit up by the running withcase. Fig. 2(a) exhibits two time game-plan X (top) and Y (base) with a length of 150that recommends time learn of 150 x-focus and the time approach variable at y-focus.The window length is set to be 120 and the window set apart by the specked rectangle.Fig. 2(b) shows the slacked relationship among’s X and Y at each slack l enrolled byEq. (1) over the two windows. The condition is according to in figure 3.2:
Figure 3.2:
The lagged correlation Equation
The most extreme slack m = 60, i.e., mod(l) 60. At the point when the processed slackl ¡ 0 (that is considered as Y is deferred by X with slack l), the positive relationshipexists for l [60, 39] (the shadowed range). At the point when l 0 (i.e., X is deferredby Y with slack l), beginning from l = 1, we can watch a solid increment in positiveconnection and it accomplishes a pinnacle estimation of 0.81 at l = 32. We have tototal all the watched relationship values over the whole slack traverse keeping in mindhapter 3.
Experimental Design & Setup
Figure 3.3:
Aggregated lagged correlation between two time series
We say that Si leads Sj if l is less than 0, and Si led by Sj otherwise if l is greater than 0.Such leadership (Si leads Sj or vice versa depending on l value) is also called the lead-lagrelation between Si and Sj.
Figure 3.4:
Two Time Series
In light of this figured lead-slack connection esteem we can build a chart. This willseparate pioneer among set of time arrangement. Subsequently next stride is to buildedge-weighted guided chart in view of slacked connections to dissect the lead-slack con-nection among the arrangement of time arrangement.
With a specific end goal to outline the administration connections among an arrangementof time arrangement, building an edge-weighted diagram will learn lead-slack relationshipamong the arrangement of time arrangement information.hapter 3.
Experimental Design & Setup Figure 3.5:
Lagged Correlation
Straightforward edge-weighted coordinated chart, G (V, E), where the hubs V = S1, S2. SN speaks to N time arrangement, and the coordinated edges E speaks to lead-slackrelations between combine of time arrangement. An edge (Si, Sj) shows that Si is drivenby Sj and its weight is set as Ei j(r). Since we are keen on pivotal lead-slack relations,a connection edge g is set. The edge will reinforce the development of a chart to suchan extent that exclusive those sets Si and Sj with Ei j(r) ¿ g have edges in G. This willhelp in extricating solid pioneer among the time arrangement information by setting thesuitable limit esteem.
Figure 3.6:
Comparison of Leadership Score on Different Graph Structures hapter 3.
Experimental Design & Setup In perspective of the structure of G and the PageRank estimations of time course ofaction, remove the pioneers by discarding unsuitable specialists. The primitive believedis to first sort the time plan by the sliding solicitation of their PageRank values andafter that to oust iteratively the time course of action that is driven either by officiallyfound pioneers or by the relative of previously found pioneers. This will realize numberof pioneers.
For many years theorists have attempted to make a fiscal benefit in monetary showcasesby foreseeing the future cost of wares, stocks, remote trade rates and all the more asof late prospects and alternatives. In the course of the most recent couple of decadesthese endeavors have expanded particularly, utilizing an assortment of procedures (Hsu,which can be comprehensively ordered into three classes: • fundamental analysis • technical analysis • traditional time series forecasting Fundamental analysis makes utilization of essential market data keeping in mind the endgoal to anticipate future developments of a benefit. On the off chance that a financialspecialist was taking a gander at a specific stock’s basic information they would considerdata, for example, income, benefit estimates, supply, request and working edges andso on. Theorists taking a gander at commodities should seriously mull over weatherpatterns, political angles, government enactment and etc. Viably fundamental analysisis worried with full scale monetary and political components that may influence thefuture cost of a money related resource. Basic investigation is not viewed as further inthis review.
Technical analysis is the investigation of verifiable costs and examples with the point ofanticipating future costs. Experts of specialized investigation in the past were alluded tohapter 3.
Experimental Design & Setup
Financial time series data are a sequence of prices of some financial assets over a spe-cific period of time. Fundamental analysis is the examination of the underlying forcesthat affect the well-being of the economy, industry sectors, and individual companies.For example, in our problem to forecast the future performance of the individual advi-sors. Here we are forecasting the performance of multiple leading advisors of the financialmarket. Hence, the data used is multivariate time series data. Since its is a multivariatetime series data this problem can be solved using VAR models (vector autoregressivemodels).VAR model is used for multivariate time series. The vector autoregression (VAR) modelis a standout amongst the best, adaptable, and simple to utilize models for the inves-tigation of multivariate time arrangement. It is a characteristic augmentation of theunivariate autoregressive model to dynamic multivariate time arrangement. The VARdisplay has turned out to be particularly valuable for portraying the dynamic conduct ofmonetary and money related time arrangement and for estimating. It frequently givesbetter conjectures than those from univariate time arrangement models and expoundshypothesis based synchronous conditions models. Conjectures from VAR models arevery adaptable on the grounds that they can be made restrictive on the potential futureways of indicated factors in the model.Before proceeding with VAR model, modeling a time series requires certain criteria tosatisfy and it includes stationary series, random walks , rho Coefficient, Dickey FullerTest of Stationarity. I took up this segment first because that until unless your timearrangement is stationary, you cannot build a time series model. In situations where thestationary measure are damaged, the principal imperative moves toward becoming tostationarize the time arrangement and afterward attempt stochastic models to forecastthis time arrangement. There are different methods for bringing this stationarity. Somehapter 3.
Experimental Design & Setup
Figure 3.7:
Constant mean across time
Figure 3.8:
Constant variance across time
In statistics and econometrics, an Augmented DickeyFuller test (ADF) tests the nullhypothesis of a unit root is available in a period arrangement test. The option hypoth-esis is distinctively relying upon which rendition of the test is utilized, yet is normallystationarity or pattern stationarity. It is an increased adaptation of the DickeyFullertest for a bigger and more entangled arrangement of time arrangement models. Theaugmented DickeyFuller (ADF) measurement, utilized as a part of the test, is a nega-tive number. The more negative it is, the more grounded the dismissal of the hypothesisthat there is a unit root at some level of certainty. Once the series is found stationary,the model can be build on the multivariate time series data.hapter 3.
Experimental Design & Setup Figure 3.9:
Constant auto-covariance across time
The structure of VAR is that every variable is a straight capacity of past slacks of itselfand past slacks of alternate factors.As an example suppose that we measure three different time series variables, denotedby x ( t, , x ( t, , andx ( t, Figure 3.10:
Vector autoregressive model of order 1
Each variable is a linear function of the lag 1 values for all variables in the set.In a VAR(2) model, the slack 2 values for all factors are added to the correct sides ofthe conditions, For the situation of three x-factors (or time series) there would be sixindicators on the correct side of every condition, three slack 1 terms and three slack 2terms.In general, for a VAR(p) model, the principal p slacks of every variable in the frameworkwould be utilized as regression indicators for every variable. The value of p can becalculated using VARselect method. This helps to select lag value between two timeseries and further proceed with the VAR model as discussed earlier.hapter 3.
Experimental Design & Setup
Figure 3.11:
VAR Model
The underlying stride in the investigation is an examination stasioneritas data. To lookat the stationary of the information that can be utilized as a unit root test. The unitroot test utilized depends on the Augmented Dickey-Fuller () test. Numerically, thetype of ADF is shown as follows:
Figure 3.12:
ADF test hapter 3.
Experimental Design & Setup H : ρ = 0 (a unit root exists).At the hugeness level of (1 - α )100%, H is rejected, if the measurements is not as muchas the basic incentive at the time of α , or p value and not as much as the criticalnessestimation of α . Implying that, the information is stationary.The following stride is the determination of the slack request. The points of this pro-gression is to get the ideal slack request of the model. Slack request choice uses theaccompanying information criteria(Figure 3.13). Figure 3.13:
Information criteria for lag order selection
Where, p is lag, k is the number of endogenous variable.Estimation of slack p picked asthe estimation of p ∗ which limits the data criteria in the watched interims of 1, ..., p max .Slack is ideal in view of the littlest estimation of AIC , SC and HQ .The successive altered likelihood ratio (LR) test is completed. Beginning with the great-est slack, trailed by trial of the hypothesis to check whether the coefficients on slack pare mutually zero utilizing the X measurements.After the estimation of the model in view of the ideal slack order, indicative checking forthe lingering was done. It means to whether there is a serial connection (autocorrelation)in slack h on residuals VAR model is alluded to the best model, if the model meets the VAR analysis methodsthat is portrayed previously. The accompanying stride is forecasting of the future periodsutilizing the best of
VAR model. By and large, Mean Square Error (
MSE ) esteem isutilized to decide the precision of estimating results. The type of
MSE is appeared as:MSE = 1 n n X i =1 ( Y t − y t ) hapter 3. Experimental Design & Setup n is measure of information. A decent model will deliver smallest of esteem whichis identified with the precision of a forecast. In a few cases, the qualities can be utilizedto decide the execution of a model. hapter 4 Experiments & Results
The Morningstar Category classifications breaks portfolios into peer groups based ontheir holdings.Advisors can be clustered based on the their holdings.Portfolios are placedin a category based on their average holdings statistics over the past three years. Hereadvisors with similar category are clustered into a cluster.Below is the sample cluster of advisors with similar portfolio.Advisor at the centerrepresents as the leading advisor among the remaining advisors in the cluster.
Figure 4.1:
Sample Cluster of Advisor with Similar portfolio
Experiments & Results Leading advisors from each cluster are extracted from above step.Below figures are forecast results of sample leading advisors.
Figure 4.2:
Prediction using VAR model hapter 4.
Experiments & Results Figure 4.3:
Prediction using VAR model hapter 4.
Experiments & Results Figure 4.4:
VAR Estimation Results
Figure 4.5:
Diagram of Fit and Residuals of Sample Advisor hapter 5
Conclusions
This paper basically talks about analysis of investment advisors intelligence. A novelproblem of discovering leaders from set of time series data based on lagged correlationhas been proposed. A time series is learned as leader time series. The movement of leadertime series triggers many other time series which are called as followers. Behavior ofleader time series helps in learning the behavior of the followers time series. Proceedingwith further analysis of investment advisors intelligence led to pattern recognition of theinvestment advisor behavior. An approach towards this problem showed great interestin using EAs for pattern recognition tasks, and also came with other possible use ofEAs combined with other approaches for the development of fully automated patternrecognition systems 28 ibliographyibliography