Macroeconomic Factors and Retail Credit Risk Risk Managers Association Event: Istanbul 29th May 2013 Eric McVittie Director of Economic Research Experian Decision Analytics ©2013 Experian Ltd. All rights reserved. No part of this copyrighted work may be reproduced, modified, or distributed in any form or manner without the prior written permission of Experian. Experian Public. Macroeconomic factors & retail credit risk Agenda Economic conditions are fundamentally important to retail credit risk. Increased economic uncertainty makes it increasingly vital that lenders understand their exposures: ● ● ● ● ● Loss Forecasting Provisioning Concentration Risk Risk Appetite Scenario Analysis & Stress Testing Acquisitions / Account Management / Collections This is very challenging for data, models and processes. Important Difficult Possible (with care) Effective approaches: ► ► ► ► Maximise information extracted from available data Take proper account of economic and non-economic influences on loan performance Are robust, transparent and flexible Become integrated into lenders’ normal business practice ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 2 Macroeconomic factors & retail credit risk It’s Important: Value added to lenders Combining information from conventional credit risk data and models with economic data and models improves loss forecasting and helps optimise capital planning and allocation. Economic models provide a robust basis for stress testing and for enhancing decision processes within a rapidly evolving economic environment. Risk Scores & Portfolio Insight Loss Forecasting Provisioning Concentration Risk Risk Appetite Scenario Analysis & Stress Testing Economic forecasting ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian Build economics into decisions Provisioning / Capital Requirement Risk Committee / Regulators Operations 3 Macroeconomic factors & retail credit risk It’s Important: Economic challenges persist ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 4 Macroeconomic factors & retail credit risk It’s Important: Economic challenges persist GDP growth forecasts for the Eurozone, Spring forecast for following year ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 5 Macroeconomic factors & retail credit risk It’s Important: Economic challenges persist GDP Growth Forecasts for the Turkey, Spring forecast for following year ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 6 Macroeconomic factors & retail credit risk It’s Difficult: Requirements & challenges • Separate economic & non-economic factors 1 Data on trends / • Identify correct economic drivers 2 patterns in loan • Build robust & stable models for performance forecasting and stress testing 3 4 • Establish forecast & stress scenario assumptions for economic factors 5 • Loss forecasts / stressed losses / adjusted scores / etc. Challenges Limited data Many potentially relevant non-economic & economic drivers Uncertainty over model & forecast / scenario assumptions ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 7 Macroeconomic factors & retail credit risk It’s Difficult: The Problem ‘True’ Pr(good): Pijt Fj yit 1, Sit , X ijt where Pijt = probability of good (1-PD) for individual i in population j at time t Yit-1 = observable credit history for i up to t-1 Sit = (generally unobservable) individual ‘situational’ factors influencing i at t Xijt = economic conditions influencing i in population j at t Note: In general S and X vary by borrower, may include lags, and F may include interaction terms ‘Standard’ Credit Scorecard: Pijt G j yit 1 ijt e.g. Logistic regression Pijt T Pijt , ln yit 1 0 T yit 1 1 e 1 Pijt e T yit 1 Standard credit risk models are backward looking, and may be subject to biases and drifts in calibration due to changes in missing ‘situational’ factors which influence loan performance. ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 8 Macroeconomic factors & retail credit risk It’s Difficult: Possible Solutions A range of approaches have been attempted to incorporate economic factors in retail credit risk models. Aggregated models using macroeconomic factors are simple, but aggregation may lose much valuable information. Disaggregated approaches maximize information content of available data, and allows flexibility in selecting appropriate economic metrics, controlling for changes to lender and borrower behaviours, and allowing for interaction between economic & non-economic determinants of default. ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 9 Macroeconomic factors & retail credit risk It’s Difficult: Selection of economic risk factors Rich supply of historical economic data from official national and international sources: ► Primary data source (for Europe: Eurostat, national statistical offices) and statistics from secondary agencies (e.g. OECD, IMF, Bloomberg, Dow Jones) Model selection criteria: Estimation diagnostics In-sample fit The choice of economic factors should consider: ► ► ► ► Do accurate historical data exist, and are there credible forecasts of the economic data available to generate forward-looking scores? Is there a robust theoretical and empirical justification for including these factors as drivers of retail credit risk Predictive performance in holdout and ‘out of time’ samples Simulation/forecast properties Economic factors should directly relate to the ability to pay. These will be factors related to income, employment / unemployment and cost of repayment (i.e. interest rates – even this one is perhaps controversial) Care is required in model specification to capture as broad as possible a range of economic risk factors consistent with the data and while avoiding overfitting ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 10 Macroeconomic factors & retail credit risk It’s Difficult: Selection of economic risk factors Search space is enormous : ► Many potentially-relevant variables ► Functional forms? ► Lag structures? Finding ‘correct’ models hampered by: ► ► Data availability – particularly historical time series Estimation/Identification issues – collinearity, endogeneity High risk of building models based on spurious correlations. Grid Search / Data Mining Statistical Data Reduction Methods Theory, priors & ‘expert’ judgement Macroeconomic factors proxy more direct influences on borrowers – proxy quality varies. Restrict search space to variables with ‘clear’ link to borrower finances: income / net worth / affordability Need to make efficient use of available data, avoiding excessive aggregation and exploiting variation across sub-populations. Great care is needed to establish robust / stable models that can generate reliable forecasts ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 11 Macroeconomic factors & retail credit risk Example: UK Credit Card Delinquency & Unemployment Normalized series - levels 3 Covariance Analysis Card Delinquency Rate 2 ILO Unemployment Rate Covariance Correlation t-Statistic Probability Claimant Rate Covariance Correlation t-Statistic Probability 1 0 -1 -0.375 -0.242 -2.298 0.024 -0.808 -0.400 -4.020 0.000 -2 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 Credit Card Delinquency Rate ILO Unemployment Rate Claimant Unemployment Rate ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 12 Macroeconomic factors & retail credit risk Example: UK Credit Card Delinquency & Unemployment Moving correlations between ILO rate & card delinquency rate Moving Correlation Coefficient 1 Full Sample Card Delinquency Rate 0.5 0 5 year 7 year 10 year -0.5 15 year -1 -1.5 1997Q1 2000Q1 2003Q1 2006Q1 ILO Unemployment Rate Covariance Correlation t-Statistic Probability Claimant Rate Covariance Correlation t-Statistic Probability -0.375 -0.242 -2.298 0.024 -0.808 -0.400 -4.020 0.000 2009Q1 End of rolling sample period ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 13 Macroeconomic factors & retail credit risk Example: UK Credit Card Delinquency & Unemployment 4 Normalized series – delinquency rate levels, quarterly changes in unemployment rates (smoothed) Covariance Analysis 3 Card Delinquency Rate 2 Quarterly Change: ILO Rate Covariance Correlation t-Statistic Probability Quarterlly Change: Claimant Rate Covariance Correlation t-Statistic Probability 1 0 -1 -2 0.136 0.755 10.545 0.000 0.145 0.676 8.417 0.000 -3 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 Credit Card Delinquency Rate Quarterly Change in ILO Rate (Smoothed) Quarterly Change in Claimant Rate (Smoothed) ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 14 Macroeconomic factors & retail credit risk Example: UK Credit Card Delinquency & Unemployment Moving correlations between quarterly changes in ILO rate & card delinquency rate Full Sample 1 Moving Correlation Coefficient 0.8 0.6 5 year 0.4 7 year 10 year 0.2 15 year Card Delinquency Rate Quarterly Change: ILO Rate Covariance Correlation t-Statistic Probability Quarterlly Change: Claimant Rate Covariance Correlation t-Statistic Probability 0.136 0.755 10.545 0.000 0.145 0.676 8.417 0.000 0 -0.2 1997Q1 2000Q1 2003Q1 2006Q1 2009Q1 End of rolling sample period ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 15 Macroeconomic factors & retail credit risk Example: UK Credit Card Delinquency & Unemployment Moving regression coefficients between changes in ILO rate & card delinquency rate 3 Moving Regression Coefficient 2.5 2 Full Sample 1.5 5 year 1 7 year 0.5 Coefficient Std. Error t-Statistic Prob. 2.737 0.359 7.621 0.000 10 year 0 -0.5 -1 1997Q1 2000Q1 2003Q1 2006Q1 2009Q1 End of rolling sample period ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 16 Macroeconomic factors & retail credit risk Example: UK Credit Card Delinquency & Unemployment Predictive accuracy for model estimated on alternative 10 year sub-samples Forecast Error for CCD Delinq Rate 1 0.5 0 -0.5 2009 2010 -1 2011 -1.5 -2 all 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Model estimated on 10 year sample ending: ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 17 Macroeconomic factors & retail credit risk It’s Difficult Some conclusions on aggregate analysis: ► ► Estimated relationships between credit outcomes and macroeconomic factors vary greatly depending on the estimation sample In practice, estimating models linking short time series of historical credit data to macroeconomic factors is highly likely to establish spurious / over-estimated relationships to particular economic variables. Similar results in other applications using macroeconomic data: ► Loss models fitted on aggregated or account-level lender data ► Argues against explanations based on model specification ► Problem is (partially) mitigated in models using disaggregated economic data – e.g. unemployment by local area or by age To identify robust economic models for forecasting and stress testing need to maximise information extracted from available data – avoiding unnecessary and wasteful aggregation of credit or economic data ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 18 Macroeconomic factors & retail credit risk It’s Possible: Disaggregated Models ‘True’ Pr(good): Pijt Fj yit 1, Sit , X ijt where Pijt = probability of good (1-PD) for individual i in population j at time t Yit-1 = observable credit history for i up to t-1 Sit = (generally unobservable) individual ‘situational’ factors influencing i at t Xijt = economic conditions influencing i in population j at t Note: In general S and X vary by borrower, may include lags, and F may include interaction terms ‘Augmented ’ Credit Scorecard: Pijt G j yit 1 H j ( X ijt ) ijt where Gj(yit-1) = standard credit score Hj(Xijt) = economic risk score (calibration adjustment) Use appropriate estimation and inference methods to fit and test models, allowing for (a) potential endogeneity between economic factors and credit score; and (b) error structure given use of aggregated time varying economic data. This approach provides maximum flexibility and transparency while efficiently utilizing all available data ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 19 Macroeconomic factors & retail credit risk It’s Possible: Maximizing use of sub-national data Most countries publish rich economic data in far more detail than high level macroeconomic series – e.g. for regional and local geographies. This disaggregated data is potentially more relevant to the conditions of individual borrowers/accounts than are macroeconomic aggregates. Careful segmentation of accounts can also allow models to capture the variation in sensitivity of different borrowers/accounts to economic conditions. ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian www.ipa.sanayi.gov.tr 20 Macroeconomic factors & retail credit risk Conclusions There is a need reliably to link credit outcomes to economic conditions for loss forecasting, stress testing (and many other operational & strategic uses) This raises big challenges for model specification and selection of economic variables – given limited information available for estimating models and the statistical characteristics of candidate economic variables. Current approaches tend to emphasize a small set of macroeconomic factors (or functions of those factors) that correlated with historical trends in credit outcomes – but some of those relationships appear to be unstable and/or to have broken down in more recent history. Similar to an early-stage epidemiological study? – we can find correlations but we need to validate their robustness and to understand the mechanisms involved before we hang too much on the results Is it time to rethink this problem - focusing on identifying & understanding proximate economic drivers of default/delinquency and firmly linking those to reliably measurable proxies? ► Dig deeper into the economic data. ► Exploit variation in disaggregated data for model estimation and loss forecasting. ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 21 ©2013 Experian Experian Information Ltd. All rights reserved. ©2012 Solutions, Inc. All rights reserved. ExperianPublic. Public Experian 22