What is probability of default (PD)?

PD sits under most credit decisions. In underwriting it helps set the spread and the lending limit so that pricing reflects expected credit losses, across business loans, asset finance and consumer credit. In the accounts it drives provisions: expected loss equals PD times exposure at default times loss given default, so a 2% PD on a $100,000 exposure with a 45% LGD gives an expected loss of $900.

Banks using internal ratings-based approaches feed PD into regulatory capital and stress testing. Across a portfolio, the spread of PDs informs concentration limits, scenario analysis, counterparty limits and collateral requirements, and the models themselves are subject to validation and governance.

A point-in-time (PIT) PD is conditioned on today's borrower signals and economic conditions. It responds quickly, which makes it the right input for pricing, provisioning and forward-looking stress tests, but it is volatile and needs frequent recalibration and macro overlays.

A through-the-cycle (TTC) PD is smoothed over the credit cycle to represent an average default likelihood. It is stable and comparable, which suits long-term capital allocation, but it can understate near-term risk. Lenders convert between the two using a cycle factor or a macro model, and model governance has to record which basis is reported and how the conversion is done.

Statistical models dominate retail and SME lending. Logistic regression scores a borrower on inputs such as financial ratios, payment history, utilisation, industry and economic variables, then converts the score to a PD. Survival models look at when a default is likely rather than only whether one happens, allowing for loans repaid early or brought back up to date. Machine learning can rank borrowers well but often gets the probabilities themselves wrong, so the scores are recalibrated before they are used.

Structural models compare a listed company's share market value with its debts to judge how close it is to default. Market prices imply a PD as well, though they also reflect liquidity and what investors charge for taking on risk. Mappings from rating bands to PDs, drawn from historical default studies, round out the toolkit.

A PD model is tested on two things: discrimination, whether it ranks riskier borrowers higher, and calibration, whether the predicted probabilities match the defaults that actually happen. Backtesting compares average predicted PD with the observed default rate, overall and by score band, and stress tests re-run PDs under adverse scenarios.

Point-in-time PDs are typically recalibrated at least quarterly or when performance drifts; through-the-cycle PDs less often. The usual pitfalls are having too few defaults to learn from, building the model only on the borrowers who stayed, and letting it drift out of calibration as conditions change. APRA expects credible governance, independent validation and documented calibration for models used in capital, in line with the Basel framework.