This paper develops new, dynamic and conditional versions of Vasicek’s widely used single factor, default rate distribution. We employ our new class of distributions in modelling US bank loan losses. We analyze the implications for risk, capital, diversification and cyclical effects in loan portfolios and investigate how observed macroeconomic factors such as shocks to industrial production and unemployment affect the distribution of credit losses. A strength of our approach is the simplicity with which one may incorporate rich patterns of autocorrelation and dependence on observable factors into default rate distributions. JEL classification: G33; E44; G21 Keywords: Default; Bank loan; Loss distribution, Capital, Stress test
Research Paper
Dynamic Default Rates
Date: May 2008 Reference Number:8/4
DYNAMIC DEFAULT RATES
Robert LambWilliam Perraudin
LondonLondon
First version: August 2005
1 Introduction
(iii) assessments of structured products by ratings agencies, and (iv) the pricing of
hazard rate models includes Duffie, Saita, and Wang (2007), Duffie, Eckner, Horel,
Vasicek shows that, as the number of exposures in the portfolio increases to infinity,
the capital charge formulae or “capital curves” contained in the Basel II proposals.
Cram, and Lundstedt (2004) and Bartram, Brown, and Hund (2007). Such models
probabilities and ratings transitions are related to business cycles. Nickell, Per-
saran, Schuermann, and Weiner (2004) develop a multi-country macro model and
data for individual firms (rather than on aggregate default or transition rates) is chal-
failures. Lando and Skodeberg (2002) models corporate ratings transitions using a
the only source of correlation between firms is observable variables. They apply this
a hazard-based framework. Das, Duffie, Kapadia, and Saita (2007) investigate factor
olds. For example, Schumway (2001) uses a discrete time logit model to investigate
patterns in bank lending standards, tracing the evolution of standards to the incen-
loss series, examining correlation across different loan market sectors, for example.
2 Loan Loss Models
2.1 Loan Losses with Autocorrelation and Trends
loss rates, appropriately transformed. Vasicek’s model has been extended in other
when underlying risk factors are non-Gaussian.
t if a latent variable, Zi,t satisfies Zi,t<c for a constant c. Suppose that the Zi,t for
p
Zi,t=ρXt+1−ρϵi,t.
(1)
Assume that:
Xt=βXt−1+ληt
(2)
where ϵi,t and ηt are standard normal and independent for pairs of obligors i and j
formulated, the shocks are assumed to have unit variance. We construct our model
Xt=i=0∑∞(β)iληt−i
(3)
Variance(Xt)=i=0∑∞βiλ=1−βλ
(4)
if we choose λ=1−β, the unconditional variance of Xt is unity and
p p
Xt=βXt−1+1−βηt.
(5)
of McNeil and Wendin (2006) and McNeil and Wendin (2007). If Xt has unit vari-
Φ−1(q)=c.
(6)
Now, consider the probability of default for i in our model conditional on information
where ηt is the shock in equation (2) above. Note that the negative sign before the
Solving (21) for Xt−1, we obtain:
Xt−1=ρβ1[Φ−1(q)−1−ρθ~t−ρ1−βηt].
(22)
Lagging this equation and substituting in
p p
Xt−1=βXt−2+1−βηt−1
(23)
yields the following result:
Proposition 1 Suppose that an individual obligor who is not in default at t−1 defaults at t if a latent variable Zi,t<c. Assume that the Zi,t satisfy the Gaussian autoregressive factor structure given by equations (1) and (2). Let θt denote the loss rate, i.e., the fraction of obligors that default. As n→∞, the transformed loss rate, θ~t=Φ−1(θt) converges to the following Gaussian order-1 autoregressive process:
that an obligor not in default at t−1 defaults at t if Zi,t<c(t) for a function c(.).Define qt≡Φ(c(t)). Then, one may easily show that θ~t follows the process:
This process is autoregressive around a non-linear trend.
2.2 The General AR Case
factor, Xt. Suppose that:
Xt=αB(L)Xt−1+1−αηt
(27)
whereB(L)Xt−1≡i=1∑MξiXt−i.
(28)
Choose α,ξii=1,2,…,M so that the unconditional variance of Xt is unity.
Similar arguments to those used before imply:
θ~t=1−ρΦ−1(qt)−ραB(L)Xt−1−1−ρρ1−αηt.
(29)
ton (1994)), then:
Xt−1=B−1(L)(ρα1[Φ−1(qt)−1−ρθ~t−ρ1−αηt])
(30)
Lagging this equation by one period, substituting for Xt−1 and Xt−2 in:
Xt−1=αB(L)Xt−2+1−αηt−1,
(31)
and then rearranging, one obtains:
Proposition 2 Under the above assumptions, when Xt follows the general autoregressive process in equation (27), the transformed loss rate, θ~t≡Φ−1(θt) , converges to the process
Note, that the data is presented net of recoveries. We therefore scale each of the loss
delinquent loans in order to be able to demonstrate a better performance subsequently.
inspection of the series seems to confirm the presence of trends and cyclical behavior
with a standard deviation that depends in a simple way on the asset correlation.So regressing θ~t≡Φ−1(θt) on a constant, one may infer the unconditional default
It is easy to extend this to cases in which the transformed loss rate is autoregressive
All Real Estate Loss Rates Credit Card Consumer Loss Rates Other Consumer Loss Rates 1 3 0.8 0.7 0.8 2.5 0.6 0.6 2 0.5 0.4 0.4 1.5 Default rate (%) Default rate (%) Default rate (%) 0.3 0.2 1 0.2 0 0.5 0.1 1990 1995 2000 2005 1990 1995 2000 2005 1990 1995 2000 2005 Year Year Year
Lease Loss Rates C&I Loss Rates Agricultural Loans Loss Rates 1 1.4 3 1.2 2.5 0.8 1 2 0.6 0.8 1.5 0.6 0.4 Default rate (%) Default rate (%) 0.4 Default rate (%) 0.2 0.2 0.5 0 0 0 1990 1995 2000 2005 1990 1995 2000 2005 1990 1995 2000 2005 Year Year Year
Figure 1: Calibration Loan Loss Series
were calibrated using this data, one without autocorrelation, so a static model, and then another
the AR(1) case, one may infer the unconditional default probability q, the square of
3.2 Results
loan categories in that the mean loss is 1.63%. By contrast, Agricultural loans, (A),
Table 1: Parameter Estimations
Parameter estimates based on various aggregate quarterly US bank loan loss series. Estimates were
Unconditional default probability q (%)0.251.630.390.280.480.35
on default probability and default correlation, ρ, and cannot be summarized simply
other series. This then translates into a smaller correlation parameter for Credit
relation. Introducing autocorrelation substantially reduces the standard errors of the
ρ are broadly comparable to those from the model without autocorrelation. This does
4 Factors and Correlations
4.1 Specification Tests
From each transformed loss rate time series, one may use the equation
θ~t=1−ρΦ−1(q)−ρXt.
(33)
both the static and autocorrelation models for each loan category. In the autocorre-
ηt=1−βXt−βXt−1.
(34)
and the shocks should follow a white noise process. Inspection suggests that there
Watson statistic tests for autocorrelation in the residuals of a regression. Table 2
As additional specification tests, we perform normality tests on the innovations
All Real Estate Factor Shocks Credit Card Consumer Factor Shocks Other Consumer Factor Shocks 4 4 4
2 2 2
0 0 0 Factor Shock Value −2 Factor Shock Value −2 Factor Shock Value −2
−4 −4 −4 1990 1995 2000 2005 1990 1995 2000 2005 1990 1995 2000 2005 Year Year Year
Plots of the common factor autoregressive innovations extracted from the US banks aggregate quar-
cant degree from their values under normality. We generate confidence levels for
The Durbin-Watson and the Jarque-Bera tests suggests no rejections of the null
Figure 2
Autoregressive Factor Shocks Plots of the common factor autoregressive innovations extracted from the US banks aggregate quar- terly loan loss rates. The inferred shocks are from 1985 to fourth quarter 2007.
4.2 Factor Correlations
different sectors of the loan market will help a bank reduce its risk. In asking such
factor innovations from the dynamic model, in both cases based on quarterly data.
Table 2: Hypothesis Testing on the Autoregressive Shocks
the confidence level of 1.679 implies the null is not rejected at 95%. The Jarques-Bera test assesses
correlated with Other Consumer Loans. All Real Estate is fairly highly correlated
Loans. This reflects the fact that the shocks in the static model may be thought
components (see Table 4). To analyze how far the correlation matrix is from one with
factors are present in the factors driving the individual loan loss rates. More details
First the eigenvalues are arranged in ascending order. The smallest p eigenvalues
Table 3: Loss Rate Correlation Matrices
using quarterly US bank loss rates. In the static case the correlation between the common factors
ACorrelation Matrix with AutocorrelationRE0.371.00−0.33−0.18−0.230.110.410.160.550.281.000.40
dence levels. The tests suggest that in both the static and dynamic model, multiple
most important. The corporate loan categories, C&I and Leases mainly have positive
Table 4: Principal Component Analysis of Factor Correlations
both a model with and without autocorrelation in the factor. Each row is an unobserved factor that
Common Factor 6Eigenvalue and Factor Weights with AutocorrelationCommon Factor 12.600.410.290.61−0.09−0.76−0.020.330.30−0.140.35−0.260.28−1.15
5 Conditional Default Rate Distributions
5.1 Conditioning on Macroeconomic Variables
VaRs or Expected Shortfalls for the distribution of the future portfolio value. These
growth). Such analysis is often referred to as stress testing. Research on stress
Table 5: Hypothesis Testing of Eigenvalues
lation matrix. Appendix A describes the testing hypothesis. Values at each of the confidence limits
Confidence LimitModel with No Autocorrelation75%485%390%395%2
macroeconomic shocks and hence, in a sense, devise stress test methods for credit
Duffie, Saita, and Wang (2007) suppose that fluctuation in observable variables induce
obligor i is assumed to default at time t based on equation (1). Now, suppose that
dard deviation and a shock. This can then be rearranged, lagged and substituted
Proposition 3 When individual-borrower latent variables, Zit, and the unobserved common factor Xt, follow the processes given by equations (35)and (5), respectively, as the number of loans n→∞, the transformed loss rate, θ~t≡Φ−1(θt) , converges to the process
θ~t=βθ~t−1+1−ρ1−βΦ−1(q)
−1−ρρλa(Yt−βYt−1)−1−ρρ1−λ21−βηt.
(40)
Unemployment Rate
Percent (%)
1950 1960 1970 1980 1990 2000 Year
Industrial Production
Index based (2002 = 100) 20
1920 1930 1940 1950 1960 1970 1980 1990 2000 Year
Figure 3: Macro-Economic Stress Parameters
is the US Unemployment Rate over the period 1948 to 2007 and the lower panel is the US Industrial
5.2 Estimation
described: Credit Card losses and C&I losses. As before, the series were re-scaled by
stresses involving the US Employment Rate and US Industrial Production.2 The
changes in unemployment and growth in production so we expressed the two series as
same period as the loss rate data, 1985 to 2007, and both series have been transformed to standard
dividing by their sample standard deviations. Figure 4 shows the normalised series
ployed Maximum Likelihood. In a multivariate setting the weights on the observed
Figure 4
Normalised Stress Series Stress series used as observable conditioning variables in the model. The series are both over the same period as the loss rate data, 1985 to 2007, and both series have been transformed to standard normal. The top panel is the annual percentage change in US Employment Rate and the lower panel is the annual percentage change in US Industrial Production.
5.3 Stress Estimates
the annual percentage change in the US Unemployment Rate, (ii) the C&I loss rate
For the Credit Card losses, the factor correlation, ρ, is 0.60%. The unconditional
Table 6: Univariate Stress Parameterisation
US Unemployment Rate over the period 1985 to 2007. The values in parenthesis are the t-statistics
Factor correlation ρ(%)0.60
(3.83)
Unconditional default probability q (%)1.67
(16.15)
AR(1) parameter β(%)68.14
(8.32)
Stress coefficient λ(%)−37.56(−2.88)−26.22(−2.34)
coefficient is correct. The stress coefficient is negative. This means that as the Unem-
much lower at 0.38% while the factor reversion is higher at 89.58%. Again, the
dynamic model. The parameters estimates are all slightly smaller. The correlation
Table 7: Multivariate Stress Parameterisation
5.4 Stressed Default Rate Distributions
Unemployment Rate stress, Figure 5, one can see how the loss rates are driven by
Table 8 shows statistics for these distributions. Looking first at the four year hori-
The one-year horizon loss rates present a different picture. Note that in Figure 5,
employment Rate. Stress period found to be 1950 to 1954. The series has been transformed to
case is higher, at 1.81%, than the unconditional case, at 1.69%. But, due to the drop
Figure 5
Unemployment Recession Stress The four year period of maximum positive change in the annual percentage differences in US Un- employment Rate. Stress period found to be 1950 to 1954. The series has been transformed to standard normal using the same normalisation parameters as those in the estimation exercise.
6 Capital Implications
6.1 Implied Capital
Marginal Value at Risk (MVaRα) for a single loan may be calculated as the expected
MVaRα=LGD×Φ(1−ρΦ−1(q)−ρβXt−1−ρ1−βΦ−1(α)),
(41)
Credit Card Conditional Loss Rate Paths 4.5
3.5
2.5 Loss Rate (%) 2 1.5
2008 2008.5 2009 2009.5 2010 2010.5 2011 2011.5 2012 Years
Credit Card Unconditional Loss Rate Paths 3.5
2.5 Loss Rate (%) 1.5
Credit Card loss rates stressing on the annual percentage change in US Unemployment Rate between
where the factor shock, ηt is taken at its α-quantile.
The probability of default conditional on information at t−1 is given by equation
Proposition 4 Under the above assumptions, the Marginal Value at Risk for a single loan, denoted MVaRα, is:
¶
MVaRα=LGD×Φ(1−ρβ1−ρΦ−1(qt)−ρ1−βΦ−1(α)).
(42)
the expected loss:
Basel Capital Formula=LGD×Φ(1−ρΦ−1(q)+ρ1−βΦ−1(α∗))−LGD
(43)
default probability qt. When β=0 so the factor is no longer autocorrelated, the
Credit Card Conditional One Year Loss Distribution Credit Card Conditional Four Year Loss Distribution 0.12 0.12
0.1 0.1
0.08 0.08
0.06 0.06 Probability Probability 0.04 0.04
0.02 0.02
0 0 0 1 2 3 4 5 0 1 2 3 4 5 Loss Rate (%) Loss Rate (%)
Credit Card Unconditional One Year Loss Distribution Credit Card Unconditional Four Year Loss Distribution 0.12 0.12
0.1 0.1
One and four year conditional and unconditional simulated loss distributions of US Credit Card loss
Figure 6
Stress Monte Carlo Paths Example loss rate paths over a four year period for the conditional and unconditional cases of US Credit Card loss rates stressing on the annual percentage change in US Unemployment Rate between 1950-1954.
Figure 7
Stressed Loss Distributions One and four year conditional and unconditional simulated loss distributions of US Credit Card loss rates stressing on the annual percentage change in US Unemployment Rate between 1950-1954.
6.2 Regulatory Implications
banks if the data period is one year. We therefore convert the data from quarterly
assumes a point-in-time default probability as a parameter for their formulae. We
that the capital implied by the autocorrelation case is less than the static case for all
Table 8: Stressed Loss Distribution Statistics
stressed. In the conditional case the annual percentage differences in US Unemployment Rate over
Altman, E. I. (1968): “Financial Ratios, Discriminant Analysis, and the Prediction
Anderson, T. W. (2003): An Introduction to Multivariate Statistical Analysis.
atic Risk in the International Financial System,” Journal of Financial Economics,
Basel Committee on Banking Supervision (1999): “A New Capital Adequacy
Becker, B. (2007): “Geographical Segmentation of US Capital Markets,” Journal
Berkowitz, J. (1999): “A Coherent Framework for Stress-Testing,” Journal of Risk,
Carey, M. (1998): “Credit Risk in Private Debt Portfolios,” Journal of Finance,
Chava, S., and R. Jarrow (2004): “Bankruptcy Prediction with Industry Effects,”
Couderc, F., and O. Renault (2004): “Times-to-Default: Life Cycle, Global and
Das, S. R., D. Duffie, N. Kapadia, and L. Saita (2007): “Common Failings:
de Servigny, A., and O. Renault (2002): “Default Correlation: Empirical Evi-
Duffie, D., A. Eckner, G. Horel, and L. Saita (2006): “Frailty Correlated
Prediction with Stochastic Covariates,” Journal of Financial Economics, 83(3),
Gordy, M. (2003): “A Risk-Factor Model Foundation for Ratings-Based Bank Cap-
Gross, D. B., and N. S. Souleles (2002): “An Empirical Analysis of Personal
Hamilton, J. D. (1994): Times Series Analysis. Princeton University Press, Prince-
“Assessing the Probability of Bankruptcy,” Review of Accounting Studies, 9(1), 5–
Kealhofer, S. (2003): “Quantifying Credit Risk I: Default Prediction,” Financial
Koopman, S. J., and A. Lucas (2005): “Business and default cycles for credit
Koopman, S. J., A. Lucas, and P. Klaassen (2005): “Empirical credit cycles
Kupiec, P. (1998): “Stress Testing in a Value at Risk Framework,” Journal of
Drift with Continuous Observations,” Journal of Banking and Finance, 26(2-3),
the Cox Proportional Hazards Model to Bank Failure,” Journal of Banking and
in the Property-Liability Insurance Industry: A Comparison of Logit and Hazard
Longin, F. (2000): “From Value at Risk to Stress Testing: The Extreme Value
McNeil, A. J. M., and J. Wendin (2006): “Dependent Credit Migrations,” Work-
——- (2007): “Bayesian Inference for Generalized Linear Mixed Models of Port-
Nickell, P., W. R. Perraudin, and S. Varotto (2000): “The Stability of
Ohlson, J. A. (1980): “Financial Ratios and the Probabilistic Prediction of
Pesaran, M., T. Schuermann, B. Treutler, and S. Weiner (2006): “Macroeconomic Dynamics and Credit Risk: A Global Perspective,” Journal of Money, Credit and Banking, 38(5), 1211–1261.
Pesaran, M., T. Schuermann, and S. Weiner (2004): “Modeling Regional Interdependencies Using a Global Error-Correcting Macroeconomic Model,” Journal of Business and Economic Statistics, 22, 129–162.
Peura, S., and E. Jokivuolle (2004): “Simulation-based stress testing of banks
tegrated Risk Management with Skewed, Fat-Tailed Risks,” Journal of Financial
Ruckes, M. (2004): “Bank Competition and Credit Standards,” The Review of
Schonbucher, P. J. (2002): “Taken to the Limit: Simple and Not So Simple Loan
Schumway, T. (2001): “Forecasting Bankruptcy More Accurately: A Simple Hazard
Vasicek, O. (1991): “Limiting Loan Loss Probability Distribution,” Unpublished
Zhou, C. (2001): “The term structure of credit spreads with jump risk,” Journal of
Notes
∗The authors’ may be contacted at [email protected] or [email protected]. An earlier version of this paper was circulated under the title ‘Dynamic Loan Loss Distributions: Esti- mation and Implications’. 1An initial value of X[0] = 0 was assumed for each factor series. 2Das, Duffie, Kapadia, and Saita (2007) find some evidence that growth in US Industrial Pro- duction explains covariation in individual borrower default rates. 3We performed Dicky-Fuller tests on the stress series, not reported here, that confirmed the series are stationary.
We would like to use Google Analytics cookies to see how visitors use this website. They are set only if you accept. Cookie policy