IFRS 9 changed provisioning from an incurred-loss model to an expected-loss one: you provision for losses you expect, not just losses already realized. That forward-looking view is more accurate, but it’s also more demanding — it needs staging logic, macroeconomic inputs and a defensible, reproducible calculation.
The four stages
- Stage 1 — performing exposures. Provision for 12-month expected credit loss (ECL).
- Stage 2 — significant increase in credit risk since origination. Provision for lifetime ECL.
- Stage 3 — credit-impaired. Lifetime ECL, with interest on the net carrying amount.
- POCI — purchased or originated credit-impaired assets, tracked separately from day one.
The hard part is the Stage 1 → Stage 2 transition: defining “significant increase in credit risk” consistently across a portfolio.
ECL in three components
Expected credit loss is, at its core, PD × LGD × EAD — probability of default, loss given default, exposure at default — discounted appropriately and adjusted for forward-looking scenarios. Each component is a model, and each model needs governance.
Why on-premise matters here
ECL runs on your full loan book — some of the most sensitive data you hold. KeplerAI computes staging and provisioning inside your data plane, so exposures never leave. The data-science agent maps the calculation to your schema and the results land in monitoring, fully documented.
Auditable and reproducible
Auditors will ask you to reproduce a provisioning number months later. Every scenario in KeplerAI is versioned and traceable, so the same inputs give the same outputs — with the model logic documented for review.
IFRS 9 rewards institutions that treat provisioning as a governed modeling problem, not a spreadsheet. See the IFRS 9 solution.