ECL for consumer and microfinance portfolios with limited loss history

Consumer and microfinance books at finance companies are high-volume, small-ticket and often young: a new personal finance product, a BNPL book or a microfinance programme may have only one or two years of performance, none of it through a downturn. A PD model fitted on that history usually understates losses, because early vintages have not yet reached their peak default months and the portfolio has never been tested by a rise in unemployment, inflation or interest rates. The ECL figure therefore depends heavily on proxies, vintage curves, roll rates and overlays, and each of these can move the allowance, and so profit and covenant ratios with bank lenders, by a large amount.

How many months of performance does each product have, and have the oldest vintages reached the age at which cumulative defaults normally flatten? Where own history is thin, what proxy is used (credit bureau loss rates, a comparable product in another market, a bank's published data for a similar segment), and has the company documented why that proxy is comparable in customer profile, ticket size, tenor and collection practices? Are roll rates between delinquency buckets (current, 1 to 30, 31 to 60, 61 to 90, over 90 days) measured monthly and used to project losses on the current book, or is a single historic loss rate applied? How are forward-looking scenarios linked to the portfolio when there is no downturn in the data, for example through a link to unemployment, inflation or real wage growth borrowed from a proxy? Are management overlays quantified, approved, tracked over time and released only when the model captures the risk?

1. Segment the book by product, channel and risk drivers (for example salary-assigned versus non-salary-assigned personal finance, new-to-credit customers, microfinance group versus individual loans), because loss behaviour differs sharply between them.
2. Build vintage curves showing cumulative default or loss by months on book for each origination cohort, and identify whether young vintages are still on the steep part of the curve.
3. Measure monthly roll rates between delinquency buckets and use them, together with cure rates, to estimate the probability that today's balances reach default within 12 months and over their lifetime.
4. Where own data does not cover the lifetime or a full cycle, extend the curves with a documented proxy and adjust it for known differences.
5. Apply the 30-days-past-due and 90-days-past-due presumptions unless rebutted with evidence, and stage each account under the general approach (staging mechanics are covered in BANK-002).
6. Derive forward-looking adjustments and at least an upside, base and downside scenario; if no statistical link can be estimated from own data, use a link from the proxy or a reasoned sensitivity.
7. Add management overlays only for identified risks the model cannot capture (new product, data gaps, a sudden policy change), quantify each one, and set the trigger for its release.

Measure ECL as the probability-weighted, discounted shortfall across scenarios, using PD derived from vintage and roll-rate analysis, LGD from actual post-default recoveries (including salary deduction, guarantor and collection agency recoveries, net of collection costs) and EAD reflecting scheduled amortisation and prepayment. Where recovery data is immature, use observed recoveries to date plus a proxy for the remaining recovery period rather than assuming the long-run recovery rate. Overlays are part of the ECL measurement, not a separate reserve, and should be expressed in the same PD, LGD or scenario terms where possible so that they can be tested and unwound.

Disclose: (1) the inputs, assumptions and techniques used to estimate ECL for each consumer product, including the use of proxy data and why it is considered comparable; (2) how forward-looking information and scenario weights are incorporated; (3) the amount, rationale and movement of each material management overlay; (4) the loss allowance reconciliation by stage; (5) credit risk exposure by delinquency bucket or internal grade and by stage; (6) sensitivity of ECL to key assumptions, such as a downside scenario weighting or a change in PD.

• Applying an average historic loss rate to a young, fast-growing book, which dilutes losses because most balances are new
• treating the absence of a downturn in the data as evidence that macroeconomic adjustments are unnecessary
• using a proxy without adjusting for differences in salary assignment, tenor or collection strength
• recording overlays as round numbers with no link to an identified risk and no release trigger
• keeping overlays for years after the model has been recalibrated, so that the same risk is counted twice
• estimating LGD only from loans that have completed recovery, which excludes the worst cases still in collection
• ignoring microfinance features such as group guarantees and repeat-loan incentives that change default behaviour