Conference on Risk Management in the Personal Financial Services Sector
London, 22-23 January 2009
Consumer lending has been seen as the Cinderella area of research in banking and finance. The recent Subprime mortgage and Buy-to-Let mortgage crises, and the subsequent credit crunch have changed that. This two day conference addressed current risks and risk management strategies in the sector. Invited papers covered topics including credit risk and the macroeconomy, portfolio models of credit risk, modelling loss given default, credit trends and the economy, dynamic models of customer behaviour, evaluating scorecards and other topics.
Programme
Credit card transaction fraud detection [ PDF ]
Niall Adams, Imperial College London
Fraud is a serious problem, impacting credit card providers, customers and vendors. I describe the character of the fraud detection problem. We have deployed various statistical and machine learning tools, both supervised and unsupervised, on suitably processed data. I will present key results and show methods of combining these tools. Finally, it is noted that credit card transaction data constitutes a data stream -- an unending sequence of non-stationary data. Standard data analysis approaches do not take this into account explicitly . I demonstrate how to modify certain methods to make them temporally adaptive
Calculating LGD for credit cards [ PDF ]
Tony Bellotti and Jonathan Crook, Uni. Edinburgh
Loss Given Default (LGD) is an important measure of credit loss used by banks to help compute risk within their credit portfolios, expected loss on an individual loan, and ultimately capital requirements. The Basel II Capital Accord gives banks the opportunity to calculate their own estimates of LGD. We investigate LGD models incorporating macroeconomic conditions. Based on UK data for retail credit cards with accounts defaulting between 1999 and 2005 we build models of LGD and test them as both explanatory and predictive models. We find that models that incorporate macroeconomic conditions at time of default perform better than those that do not include macroeconomic information. Additionally we show how these models can be used for stress testing.
Short-range prediction of individual customer behaviour [ PDF ]
Adam Brentnall, Martin Crowder and David Hand, Imperial College London
Francis Galton, Charles Stein and Herbert Robbins made important contributions to the development of statistics, including respectively in regression, shrinkage estimation and empirical Bayes. In this talk we appeal to these ideas in a method for the short-range prediction of individual behaviour. The approach is applied to debt-collection data, where the effect of actions on different individuals is to be predicted. The method helps to order the predictions and it is computationally attractive.
Dynamic Behavioural Models for Consumer Credit
Jonathan Crook and Tony Bellotti, Uni. Edinburgh
Recent economic events have emphasised the important effects that the macroeconomy may have on the probabilities of individual cases defaulting on loans. In this presentation we are interested in including characteristics either of borrowers or of the macroeconomy that vary over time in models of the risk of default of individual borrowers. Most credit lenders have data which has a panel structure. In this presentation we specify a generic model of consumer credit risk which incorporates time varying variables and contrast standard logistic regression models with survival models. We show how application and behavioural scoring models are nested within the generic model. The time varying variables can be behavioural credit scoring variables or macroeconomic variables. We illustrate the results of including both types of variables in discrete time survival models. We show that both types of variables can contribute to the prediction of default. We show how the results can be used to create an expected loss distribution when interest rates vary.
Big or Balanced? An empirical study of the effects of sample size and balancing on model performance [ PDF ]
Steve Finlay and Sven Crone, Jaywing / Lancaster University
There have been many empirical studies comparing the performance of different regression and classification techniques when applied to credit scoring. However, many of these studies have been based on small, low dimensional, data sets which are not representative of the data sets used by mainstream lending institutions in modern consumer credit markets. A second issue is the balancing of data. Many studies have used unbalanced data sets or (approximately) balanced data created from under sampling the majority class (the goods). This paper presents an empirical study of the effect of sample size and balancing on model performance, using two large data sets supplied by industry sources. One data set is an application scoring data set, the other a behavioural scoring data set. Results are presented for four of the most popular classification approaches applied to credit scoring problems. Logistic Regression, Linear Discriminant Analysis, Neural Networks and CART. The results show that both sample size and balancing are important contributors in determining the relative performance of each technique.
Credit Risk and Correlation in Retail Portfolios [ PDF ]
Alan Forrest, HBoS
This talk aims to show how ideas of correlation can change the way we model Retail Credit Risk and how correlation helps us understand cyclic Credit Risk phenomena that otherwise require ad hoc solutions. This is illustrated by analysis of long-term default time-series at portfolio level, and at obligor level. Unlike Wholesale Credit Risk, where data is thin, Retail Credit Risk can take advantage of huge datasets to apply data-hungry General Linear Mixed Modelling techniques. Never-the-less, Retail Credit Risk has some way to catch up with Wholesale in its acceptance of correlation, and this talk aims to encourage understanding among Retail Credit Risk analysts.
Evaluating Scorecards [ PDF ]
David Hand, Imperial College London
Choosing between scorecards, estimating parameters of scorecards, and deciding if scorecards need to be replaced all hinge on accurate performance evaluation. Various criteria are in widespread use for this purpose. However, some of the most widely used criteria have unrecognised deficiencies which may render them unfit for purpose. Moreover aspects of applications beyond the mere distributions of scores in the different classes may need to be taken into account in order to make effective decisions.
Markov chains for the dynamic risk of behavioural scores [ PDF ]
Madhur Malik and Lyn Thomas, Uni. Southampton
It is well known in corporate credit risk literature that ratings distributions vary over time and across different obligor types. Motivated by the above, in this paper we shall apply the Markov chain approach to study the customers migration between various behavioural scores segments over time. We shal l show that transition probabilities vary over time and model these transitions as a second-order Markov chain where the transition probabilities are made functions of the state of the economy, months on books and the second order behavioura l score. These transition matrices are applied across all accounts in the portfolio to predict how the current book will evolv e in terms of the score distribution in the future . From this, one may derive a distribution of the default states for various horizons which can help lenders to take long term lending decisions by estimating the risk associated with the change in the credit quality of portfolio of loans over time. These models may account for the current economic downturn and can be used to stress test for regulatory purposes and to produce forecasts for expected loss, risk weighted assets and impairments at the account level under different economic scenarios.
Credit scoring is fun [ PDF ]
Graham Platts
Graham Platts is past CEO of Experian-Scorex, and will set the scene for the conference by recounting some real sto ries from years in the industry.
Marginal Chi2 analysis: beyond goodness of fit for logistic regression models [ PDF ]
Gerard Scallan, ScorePlus
Logistic Regression is the most widespread technique for constructing credit risk scorecards, both for Basel and for internal portfolio management purposes. It can be interpreted as solving a set of "actual = expected" equations. For each predictor variable in a model, the average value over the actual sample goods must be the same as that on the average over the total sample weighted by the model-estimated probabilities of good. This leads to a set of distance and certainty measures for the discrepancy between the observed and model-estimated pattern on any potential predictor. With these measures once can identify areas where the model does not adequately estimate risk. This paper describes how the process facilitates model development: it identifies candidate variables for model selection, helps construct categorical variables and evaluates potential interaction terms. Once a model has been estimated, the approach provides a rigourous framework for model validation and performance tracking at the level of the individual covariates. It identifies the most important variables causing model deterioration - whether these are terms used in the model or not. Finally, the approach can be used to make "in flight adjustments" to models after deployment, as changes in risk patterns emerge.
Low default PD estimation revisited [ PDF ]
Dirk Tasche, Lloyds TSB
Pluto & Tasche (2005) proposed a methodology for PD (probability of default) estimation in low default portfolios. This methodology can be described as confidence bound calculation or quantile matching. Reasonable values for the confidence level and the correlation parameters have to be chosen to make the estimation work appropriately. We compare the low default methodology to Bayesian and Maximum Likelihood approaches to the problem. Based on these observations, we provide recommendations for the choices of the correlation and confidence level parameters. This is illustrated by an application to empirical data.
Credit risk for portfolios of consumer loans [ PDF ]
Lyn Thomas, Uni. Southampton
Credit scoring models have proved very successful for assessing the risk of individual loans for more than fifty years. Yet the fact that the Basel New Accord had to use a corporate credit risk based model to assess the capital needed for portfolios of consumer loans and the failure of the credit rating agencies to correctly price Residential Mortgage Backed Securities have both highlighted the lack of satisfactory models to asses the credit risk of portfolios of consumer loans based on credit scores. This paper investigates how this can be done by using a survival analysis model of the credit risk of individual loans. The correlation between the loans is given by incorporating the macroeconomic variables, which affect all loans and the vintage of the loans into the model. The model is developed for a portfolio of credit cards and shows how important it is to include the vintage effect which allows for changes in the market conditions for loans and in lender policy since otherwise the impact may be incorrectly ascribed to economic conditions.
Credit scoring - a social/financial exclusion project
George Wilkinson, George Wilkinson Associates
My work in Consumer Credit has been dominated by the practical application of credit scoring methodology. This has been to support many different lenders and products - with a lender perspective. The underlying source data for scoring has been basic to start with. More recently models have been dominated by credit reference bureau information. At an industry workshop in late 2007 I expressed an opinion that this trend probably disadvantaged those with few or no accounts on the CRAs. Nigel Kershaw, Chairman of BigIssue and Sarah Forster (ex-World Bank) attended. BigIssue had created a social banking arm intent on social change and investment (BigIssue (Invest) having had considerable experience of housing and financial difficulties for those on lower incomes). Kershaw and Forster had just initiated a project to investigate the capture of payments data from those in social tenanted accommodation. Within five working days I was on the team and now act as the technical advisor to the project which is concerned with the impact such data can have and with operational implications. We are essentially looking at credit decisions from a consumer point of view and particularly looking at those that are declined. I have changed camps! About 30% of housing is rented. Council houses are reducing - Housing Association properties increasing but private landlords are growing even faster albeit from a smaller base. Two major lenders have confirmed that (almost regardless of income) a substantial difference in decline rates exist between owners and tenants. We have also confirmed - albeit on preliminary data - that those in tenanted accommodation tend to have fewer accounts. Social lender payments are not trivial - a typical tenant in the social rental sector can easily pay rent of close to £3,000 per annum and this can represent a significant part of disposable income. This presentation is about the journey over the past year or so, learning experiences involved and the work yet to come. The project will essentially come to a natural crossroads when we have completed the 'proof of concept' phase. We are seeking to construct a viable test behaviour scorecard for and on behalf of social landlords and to overlay it on traditional CRA data to see if it makes a difference to tenant application decisions. We and many others believe it will - as appears to be the case in the US and we will be announcing our full results in April. This has not and will not be a fast project because it has to be funded from various enterprises supporting social change. Ironically - because the end results might advantage lenders - it is often seen by potential funders as a commercial project. But lenders see it as a social and financial exclusion one. There are substantial government funds available for financial education, debt advice, basic bank and savings accounts and other financial needs, but we believe that lack of data upon which to base credit decisions is a fundamental issue that requires addressing to help with financial inclusion initiatives. Our supporters include: Price Waterhouse Coopers, Clifford Chance, Joseph Rowntree Foundation, Friends Provident Foundation, Experian and two significant UK banking groups. We are indebted to them. Implementation has yet to be f ully considered but will soon need to be addressed if our hunches are right.
More information available from the conference website.
