Potential of Psychological Information to inform Credit Scoring

Transcription

1 Potential of Psychological Information to inform Credit Scoring Alexandros Ladas Supervisors: Uwe Aickelin Jon Garibaldi Eamonn Ferguson 21/02/2013 IMA Seminars

2 The Merge of Three Fields Potential of Psychological Information to inform Credit Scoring Psychology Problem Economics Data Mining 21/02/2013 IMA Seminars

3 Credit Risk Assessment Credit Score measures creditworthiness The likelihood that person will repay his/her debts. So far: Demographics Financial/ Economic Data Credit Score 21/02/2013 IMA Seminars

4 Psychological Information Personality Profiles based on: Psychological Factors Consumer behaviour Time horizons Social comparisons Economic socialization Social support Money management Personality Traits Locus of control Extraversion Sensation Seeking Traditionally measured by questionnaires and surveys 21/02/2013 IMA Seminars

5 Problem Demographics Personality Profiles Financial/ Economic Data Credit Score How do we build the Personality Profiles? How do we use the Personality Profiles in order to inform Credit Scoring? 21/02/2013 IMA Seminars

6 Data Mining Personality Profiling Extract Psychological Information Measure Traits and Factors Credit Score Modelling Incorporate additional knowledge Modify existing models 21/02/2013 IMA Seminars

7 CCCS Data Consumer Credit Counselling Service approx 58 variables Information for people in debt [1]R. Disney, and J. Gathergood, Understanding consumer over-indebtedness using counselling sector data: Scoping Study, Report to the Department for Business, Innovation and Skills (BIS), University of Nottingham, /02/2013 IMA Seminars

8 CCCS attributes CCCS data Demographics Financial Expenditure Debt Details Information on assets, debts, income leisure, travel, food, clothing, self employed, motoring, other, housing, priority Information on the amount of debt in 9 debt types, their monthly contract payments e.t.c 21/02/2013 IMA Seminars

9 Personality Profiling on CCCS Unsupervised Approach Clustering Aim: Discover different personality profiles of debtors Demographics, expenditure 21/02/2013 IMA Seminars

10 Issues of CCCS dataset Diverse types of attributes Categorical vs numerical Outliers Asymmetry in data Heavy weighted attributes House value, debts, income, housing spending Light weighted attributes #debt items, #remaining mortgage terms, etc 21/02/2013 IMA Seminars

11 Clustering on CCCS First Approach CCCS data Demographics Financial Expenditure Debt Details Kmeans CLARA debts(+), travel(+) travel(+), income(-) assets(+), travel(+) Selfish income(+), other(+), priority(+), housing(+), assets(+) income(+), assets(+) income(+), assets(+), priority(+), other(+) motoring(+) Non Selfish 21/02/2013 IMA Seminars

12 Consensus Clustering On CCCS Consensus Clustering Optimal number of clusters using different clustering algorithms and validation indices Sample of CCCS data Demographics Financial Expenditure Debt Details age #Debt items [2] Daniele Soria, Novel Methods to Elucidate Core Classes in Multi-Dimensional Biomedical Data, PhD Thesis, School of Computer Science, University of Nottingham, /02/2013 IMA Seminars

13 Consensus Clustering on CCCS Results Present Oriented Few Mortgages Singles Cluster 1 housingclothingservicestravel+ (*) big % in unemployment, income- Cluster 2 normal Behaviour travel+ (*) motoring+ (*) Typical Households Married with Children, Employed Cluster 5 Normal Behaviour Average income Cluster 6 food+ housing+ motoring+ services+ debt repayments+ Average to high income high mortgage debt Rich households with self control and money management issues Married with children Employed Full Time High spending except travel high levels of debts expensive assets High income Debt repayments+ Cluster 3 Cluster 4 New debt: Store cards Well paid jobs 21/02/2013 IMA Seminars

14 Clustering Conclusions Rational Behaviour Debtors with high income, have bigger houses, spend more and demonstrate higher levels of debt Over expression of travel in low income groups Problems Outliers Singleton clusters from Kmeans and Hierarchical Clustering Heavy valued attributes define the clustering How to incorporate Demographics into clustering methods 21/02/2013 IMA Seminars

15 Need for transformations Tackle the issues of this dataset But more importantly, move towards Behavioural Data 21/02/2013 IMA Seminars

16 Transformation of CCCS Homogeinity analysis to transform categorical attributes to numerical Factor analysis on Financial attributes to explore relations between income-assetsdebt Analysis on detailed expenditure to explore spending behaviours Factor analysis on monthly inflow and outflow to uncover money management groups 21/02/2013 IMA Seminars

17 Homals Returns a map representation in p dimensions of the n objects and the m categories. Minimization of the function: Object scores nxp Indicator matrix nxkj Categories quantifications kjxp Missing Values nxn [3] De Leeuw, Jan, and Patrick Mair. "Homogeneity Analysis in R: The package homals." (2007). 21/02/2013 IMA Seminars

18 2D representation of Categorical Variables 21/02/2013 IMA Seminars

19 Factor Analysis Factors A set of latent variables that determine the values of the observed variables Purpose Data reduction Summarization Analyze relationships 21/02/2013 IMA Seminars

20 Factor Analysis on Financial attributes Factor1 Factor2 Factor3 udebt mortdebt hvalue finasset carvalue mortterm income Factor 2 just defines income and slightly the debt. Factor 3 does not add any new knowledge Certain attributes should not be transformed However: 1 Housing Factor 21/02/2013 IMA Seminars

21 Money Management Factor Analysis Factor1 spending - income catalogues debt - income collection agency - income credit cards - income ge capital debt - income other kinds of debt - income overdrafts - income personal loans - income store cards debt - income factor model Good for dimensionality reduction 21/02/2013 IMA Seminars

22 Expenditure analysis Hierarchical Clustering on the correlation matrix between spending items [4]Otto, Philipp E., et al. "From spending to understanding: Analyzing customers by their spending behavior." Journal of Retailing and Consumer Services 16.1 (2009): /02/2013 IMA Seminars

23 4 Spending Behavioural Clusters Consensus clustering to find the optimal Spending Behavioural Clusters Necessity Spending Household Spending Excessive Spending Leisure Spending 21/02/2013 IMA Seminars

24 Summary of Transformations CCCS data Demographics Financial Expenditure Debt Details + remained unprocessed attributes 2d Spatial Coords 1 Housing Factor 4 Spending Behaviour al Clusters 1 Money Management Factor 21/02/2013 IMA Seminars

25 Future Steps Clustering on the processed group to uncover behavioural groups Level of Debt prediction or classification Semi supervised approach Combination of the above 21/02/2013 IMA Seminars

26 References 1. [1]R. Disney, and J. Gathergood, Understanding consumer over-indebtedness using counselling sector data: Scoping Study, Report to the Department for Business, Innovation and Skills (BIS), University of Nottingham, Daniele Soria, Novel Methods to Elucidate Core Classes in Multi-Dimensional Biomedical Data, PhD Thesis, School of Computer Science, University of Nottingham, De Leeuw, Jan, and Patrick Mair. "Homogeneity Analysis in R: The package homals." (2007). 4. Otto, Philipp E., et al. "From spending to understanding: Analyzing customers by their spending behavior." Journal of Retailing and Consumer Services 16.1 (2009): /02/2013 IMA Seminars

27 Any Questions? 21/02/2013 IMA Seminars