Credit risk · Allowance for credit losses

Credit Loss Reserve

Expected credit loss for a $291M consumer installment loan book at September 30, 2026, under US CECL (lifetime losses on every loan) and IFRS 9 (12-month losses unless credit risk has risen sharply). Move the scenario weights to see how much reserve each view of the economy needs.

Source code on GitHub Loan Scorecard Studio Budget Variance Pack Treasury Cash Forecast Settlement & Liquidity Dashboard

Scenario weights

Weights are rescaled to sum to 100%. Stages are held at the committee-weighted assessment.

Unemployment rate scenarios, %

Allowance by grade

Stage 1: performing. Stage 2: significant increase in credit risk (30–89 days past due, or 12-month PD at least double its origination level and up 0.5 points or more). Stage 3: 90+ days past due.

Allowance roll-forward, Q3 2026, $M

Ending = beginning − charge-offs + recoveries + provision. The provision is split by driver: new loans; payoffs (reserve released on loans that paid off); replenishing what charge-offs used beyond the reserve already held; credit migration and seasoning; and the macro outlook (the Q3 book valued under June's versus September's scenarios).

Backtest: 12-month charge-off rate by grade

ActualTransition modelHistorical loss rate

Cumulative charge-off rate by vintage

Share of each year's originations charged off by months on book. Losses build fastest in the first 18 months. The 2025 vintage, with a riskier grade mix, runs above 2023–2024.

How it works

Probability of default (PD): four multinomial logistic models, one per delinquency state (current, 30, 60, 90 days), predict next month's state from grade, months on book, unemployment and recent delinquency. Chaining them month by month gives each loan's chance of charge-off in every future month, under any unemployment path.
Loss given default (LGD): 1 − present value of recoveries ÷ balance at charge-off, by grade. Only charge-offs with a full 12-month recovery window are used (LGD ).
Exposure at default (EAD): the scheduled balance three months before charge-off, discounted at each loan's own rate.
Scenarios: four 24-month unemployment paths, reverting to a 4.6% long-run rate over the following 12 months, weighted 15/45/30/10. The September paths are worse than June's, which drives the macro line in the roll-forward.

All data is synthetic: 70,000 loans from 2021–2026, simulated with known grade, seasoning, macro and re-default effects so the models can be checked against the truth. Built with AI-assisted coding (Claude); the model design, validation and review are my own.