Model Risk With. Default. Estimates of Probabilities of. Dirk Tasche Imperial College, London. June 2015

Transcription

1 Model Risk With Estimates of Probabilities of Default Dirk Tasche Imperial College, London June 2015

2 The views expressed in the following material are the author s and do not necessarily represent the views of the Global Association of Risk Professionals (GARP), its Membership or its Management. 2

3 Outline Two forecasting problems A taxonomy of dataset shift Estimation under prior probability shift assumption Estimation under invariant density ratio assumption An application to the mitigation of model risk for PD estimation Concluding remarks References Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 2 / 18

4 Two forecasting problems Single borrower s probability of default (PD) Moody s corporate issuer and default counts in Grade Caa-C B Ba Baa A Aa Aaa All Issuers Defaults January 1, 2009: What is a Baa-rated borrower s probability to default in 2009? Natural (?) estimate: %. 2 Source: Moody s (2015) Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 3 / 18

5 Rating profile known Two forecasting problems Moody s corporate issuer proportions and default rates in 2008 and issuer proportions in All numbers in %. Grade Issuers Default rate Issuers Default rate Caa-C ? B ? Ba ? Baa ? A ? Aa ? Aaa ? All ? How to take account of the additional data? 3 Source: Moody s (2015) Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 4 / 18

6 Two forecasting problems Some thoughts Compared to 2008, the rating profile in January 2009 has changed. Why should the grade-level or total default rates remain the same? Invariant grade-level default rates would imply almost invariant discriminatory power: Observed accuracy ratio in 2008: 63.4% Forecast accuracy ratio for 2009: 66.6% What other ways are there to reflect almost invariant discriminatory power? Assuming an invariant accuracy ratio is not sufficient for inferring PDs if only the rating profile is known. Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 5 / 18

7 Two forecasting problems Alternatives to invariant grade-level default rates Geometric interpretations of accuracy ratio (for continuous score): Based on area under Receiver Operating Characteristic (ROC) Based on area between Cumulative Accuracy Profile (CAP) and diagonal Derivative of CAP is (essentially) the PD curve of the scores. Derivative of ROC is (essentially) the density ratio of the scores. Is invariant density ratio a viable alternative? We also look at invariant rating profiles of defaulters and non-defaulters as an alternative assumption on almost invariant discriminatory power. Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 6 / 18

8 A taxonomy of dataset shift The Machine Learning perspective Classification on datasets with changed distributions is a problem well-known in Machine Learning. Moreno-Torres et al. (2012) proposed a taxonomy for dataset shifts. Setting: Each item in a dataset has a class y and a covariates vector x. p test (x, y) and p trai (x, y) are the joint distributions of (x, y) on the test and training sets respectively. ptrai (x, y) is known from observation but for p test (x, y) only the marginal distribution p test (x) is observable now. How to determine unconditional class probabilities p test (y = c) and conditional class probabilities p test (y = c x)? Definition: Dataset shift occurs if p test (x, y) p trai (x, y). On Slide 4, y = default status, x = rating grade. Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 7 / 18

9 A taxonomy of dataset shift The Moreno-Torres et al. taxonomy Four types of dataset shift p test (x, y) p trai (x, y): Covariate shift: p test (x) p trai (x), but p test (y x) = p trai (y x). Prior probability shift: p test (y) p trai (y), but p test (x y) = p trai (x y). Concept shift: p test (x) = p trai (x), but p test (y x) p trai (y x), p test (y) = p trai (y), but p test (x y) p trai (x y). or Other shifts. Assuming invariant grade-level default rates on Slide 4 is equivalent to an assumption of covariate shift. Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 8 / 18

10 Estimation under prior probability shift assumption Moody s corporate rating profiles 2008 Frequency Defaulters All Non defaulters Caa C B Ba Baa A Aa Aaa Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 9 / 18

11 Estimation under prior probability shift assumption The least-squares estimator Setting as on Slide 4: y = default status (classes D default and N non-default), x = rating grade p test (x) known, but p test (x) p trai (x) Want to determine ptest = p test (y = D) and p test (y = D x) Prior probability shift assumption: p test (x y = c) = p trai (x y = c) for c = D, N. Hence, for all x, the class probability p test should satisfy p test (x) = p test p trai (x y = D) + (1 p test ) p trai (x y = N). This is unlikely to be achievable. Therefore least squares approximation ( )( ) p test (x) p trai (x y=n) p trai (x y=d) p trai (x y=n) dx p test = ) 2dx. (1) ( p trai (x y=d) p trai (x y=n) Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 10 / 18

12 Estimation under prior probability shift assumption Fitted and observed Moody s corporate rating profiles Fitted profile 2009 Observed profile 2009 Frequency Caa C B Ba Baa A Aa Aaa Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 11 / 18

13 Estimation under invariant density ratio assumption The invariant density ratio estimator Setting as on Slide 4: y = default status (classes D default and N non-default), x = rating grade p test (x) known, but p test (x) p trai (x) Want to determine ptest = p test (y = D) and p test (y = D x) Invariant density ratio assumption: p test (x y = D) p test (x y = N) = p trai(x y = D) def = λ(x). p trai (x y = N) Then the class probability p test must satisfy λ(x) 1 0 = 1+(λ(x) 1) p test d p test (x). (2a) There is a unique solution to (2a) if and only if λ(x) d p test (x) > 1 and λ(x) 1 d p test (x) > 1. (2b) Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 12 / 18

14 Estimation under invariant density ratio assumption Properties If condition (2b) is not satisfied then the profiles p test (x) and p trai (x) are so different that any inheritance of discriminatory power seems questionable. Exact fit: With the invariant density ratio estimate p test come estimates of the conditional densities p test (x y = c), c = D, N, such that λ(x) = p test (x y=d) p test (x y=n), and p test (x) = p test p test (x y = D) + (1 p test ) p test (x y = N). Call p test = p trai (y = D x) d p test (x) the covariate shift estimator of p test. Then it holds that p test = (1 π) p trai (y = D) + π p test. (3) Where 0 π 1 and π is the closer to 1 the more discriminatory power the scores x have on the training set. Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 13 / 18

15 Estimation under invariant density ratio assumption Conditional profiles 2009: Invariant from 2008 vs. fitted Defaulters Frequency Frequency Observed 2008 Fitted 2009 Observed 2009 Caa C B Ba Baa A Aa Aaa Non defaulters Observed 2008 Fitted 2009 Observed 2009 Caa C B Ba Baa A Aa Aaa Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 14 / 18

16 An application to the mitigation of model risk for PD estimation Different forecast methods Three methods of forecasting portfolio-wide default rate p test : Covariate shift estimator p test (Slide 13) Prior probability shift estimator p test, (1) Invariant density ratio estimator p test, (2a) p test and p test give the same estimate under a prior probability shift. Estimates p test and p test of p test provide estimates of the conditional default rates p test (y = D x) by p test (y = D x) = p test λ(x) 1 + (λ(x) 1) p test. (3) suggests that max(p test, p test ) could be an upper bound for next year s portfolio-wide default rate. Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 15 / 18

17 An application to the mitigation of model risk for PD estimation Observed vs. forecast corporate default rates 4 Per cent Observed Covariate shift forecast Invariant Density Ratio forecast Year 4 Source of observed rates: Moody s (2015). Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 16 / 18

18 Concluding remarks Straightforward covariate shift (or invariant conditional default rate ) PD estimates sometimes may seriously underestimate future default rates. The invariant density ratio approach often provides very different estimates that may be used for model risk mitigation. The invariant density ratio approach can also be applied for the estimation of loss rates. Rating agency data like Moody s (2015) possibly are subjective, making results of the approach conservative. Further reading: Background and more details: Tasche (2013), Tasche (2014) More sophisticated approaches: Hofer and Krempl (2013), Hofer (2015) Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 17 / 18

19 References V. Hofer. Adapting a classification rule to local and global shift when only unlabelled data are available. European Journal of Operational Research, 243(1): , V. Hofer and G. Krempl. Drift mining in data: A framework for addressing drift in classification. Computational Statistics & Data Analysis, 57(1): , Moody s. Annual Default Study: Corporate Default and Recovery Rates, Special comment, Moody s Investors Service, March J.G. Moreno-Torres, T. Raeder, R. Alaiz-Rodriguez, N.V. Chawla, and F. Herrera. A unifying view on dataset shift in classification. Pattern Recognition, 45(1): , D. Tasche. The art of probability-of-default curve calibration. Journal of Credit Risk, 9(4):63 103, D. Tasche. Exact fit of simple finite mixture models. Journal of Risk and Financial Management, 7(4): , Dirk Tasche (Imperial College) Model Risk With Estimates of Probabilities of Default 18 / 18