Mindy Leow

The stability of survival model parameter estimates for predicting the probability of default: Empirical evidence over the credit crisis

Using a large portfolio of credit card loans observed between 2002 and 2011 provided by a major UK bank, we investigate the stability of the parameter estimates of discrete survival models, especially since the start of the credit crisis of 2008. Two survival models are developed for accounts that were accepted before and since the crisis. We find that the two sets of parameter estimates are statistically different from each other. By applying the estimated parameters onto a common test set, we also show that they give different predictions of probabilities of default. The changes in the predicted probability distributions are then investigated. We theorise them to be due to the quality of the cohort accepted under different economic conditions, or due to the drastically different economic conditions that was seen in the UK economy, or a combination of both. We test for each effect.

Intensity models and transition probabilities for credit card loan delinquencies

We estimate the probability of delinquency and default for a sample of credit card loans using intensity models, via semi-parametric multiplicative hazard models with time-varying covariates. It is the first time these models, previously applied for the estimation of rating transitions, are used on retail loans. Four states are defined in this non-homogenous Markov chain: up-to-date, one month in arrears, two months in arrears, and default; where transitions between states are affected by individual characteristics of the debtor at application and their repayment behaviour since. These intensity estimations allow for insights into the factors that affect movements towards (and recovery from) delinquency, and into default (or not). Results indicate that different types of debtors behave differently while in different states. The probabilities estimated for each type of transition are then used to make out-of-sample predictions over a specified period of time.

The Economy and Loss Given Default: Evidence from Two UK Retail Lending Data Sets

Loss given default (LGD) models predict losses as a proportion of the outstanding loan, in the event a debtor goes into default. The literature on corporate sector LGD models suggests LGD is correlated to the economy and so changes in the economy could translate into different predictions of losses. In this work, the role of macroeconomic variables in loan-level retail LGD models is examined by testing the inclusion of macroeconomic variables in two different retail LGD models: a two-stage model for a residential mortgage loans data set and an ordinary least squares model for an unsecured personal loans data set. To improve loan-level predictions of LGD, indicators relating to the macroeconomy are considered with mixed results: the selected macroeconomic variable seemed able to improve the predictive performance of mortgage loan LGD estimates, but not for personal loan LGD. For mortgage loan LGD, interest rate was most beneficial but only predicted better during downturn periods, underestimating LGD during non-downturn periods. For personal loan LGD, only net lending growth is statistically significant but including this variable did not bring any improvement to R2.

A new Mixture model for the estimation of credit card Exposure at Default

Using a large portfolio of historical observations on defaulted loans, we estimate Exposure at Default at the level of the obligor by estimating the outstanding balance of an account, not only at the time of default, but at any time over the entire loan period. We theorize that the outstanding balance on a credit card account at any time during the loan is a function of the spending by the borrower and is also subject to the credit limit imposed by the card issuer. The predicted value is modelled as a weighted average of the estimated balance and limit, with weights depending on how likely the borrower is to have a balance greater than the limit. The weights are estimated using a discrete-time repeated events survival model to predict the probability of an account having a balance greater than its limit. The expected balance and expected limit are estimated using two panel models with random effects. We are able to get predictions which, overall, are more accurate for outstanding balance, not only at the time of default, but at any time over the entire default loan period, than any other particular technique in the literature.