AI & Data Analytics Application · Practice two

ECL
Model

An IFRS 9 expected credit loss model is usually one analyst’s spreadsheet: judgment calls made once, never written down, and rebuilt from scratch every reporting cycle. Ours is a staged, governed methodology instead, from the definition of default to disclosure-ready reporting.

Live model · Portfolio ECL run
Stage 1 to Stage 2 migrations142 facilities
Scenario-weighted ECL movement+2.3%
SICR triggered · 3 facilities reclassified

A spreadsheet is not a model.

Most IFRS 9 ECL models live in one analyst’s workbook: the definition of default, the observation window, the cure policy, decided once and never written down anywhere else. Data gaps get quietly treated as zero. A small portfolio gets forced into a calibration it cannot statistically support. Macroeconomic variables get chosen by testing combinations until one fits, with no check for how easy that is to do by chance.

We built the layer that turns that into a documented, staged methodology instead: every governance decision recorded with its rationale, every economic relationship tested before it is trusted, and every final figure traceable back to the assumption it came from.

How the platform works

Two stages, run as one pipeline. Model the risk builds the parameters; Calculate & Govern turns them into a defensible number.

Model the risk

Historical data in, governed risk parameters out
01

Establish the rules

Definition of default, observation window and cure policy are agreed and documented before a single number is calculated, not assumed along the way.

02

Calibrate parameters

Historical migration and default behaviour are converted into probability-of-default, loss-given-default and exposure parameters, by risk grade.

03

Link to the economy

Macroeconomic variables are tested for a genuine statistical relationship to default behaviour, not just the best-fitting combination, then weighted into best, base and worst-case scenarios.

Calculate & Govern

Governed parameters in, a disclosure-ready figure out
01

Stage & extend

Every exposure is classified by stage and its risk parameters extended across its full remaining life, not just the next twelve months.

02

Calculate the loss

Probability, exposure, severity and discounting combine into the expected credit loss figure, weighted across every economic scenario.

03

Reconcile & defend

Every judgment call is documented and every figure ties back to a control total, so the model is ready for an auditor before one ever asks.

Where the judgment lives today

The problem, the way it’s handled today, and what changes once the ECL Model is running.

The problem

  • Governance decisions (definition of default, observation window, cure treatment) made once and never written down anywhere else.
  • Data gaps quietly treated as zero rather than flagged as missing.
  • Small portfolios forced into a calibration they cannot statistically support.
  • Macroeconomic variables chosen by testing combinations until one fits, with no check for how easy that is to do by chance.
  • Judgment calls (correlation assumptions, loss severity sourcing, scenario weights) presented as measured facts, not disclosed assumptions.

Current state

  • ECL rebuilt from scratch, or lightly patched, in a spreadsheet every reporting cycle.
  • Consistency checks (monotonicity, scenario ordering, the sign of an economic relationship) left to manual review, if they happen at all.
  • Model logic lives in one analyst’s workbook, documented nowhere else.
  • Replacing a legacy model means starting over, with no explanation of what changed or why.
  • Audit and regulatory review treated as a defence exercise after the fact, not built in from the start.

What ALP delivers

  • A documented, staged methodology: every governance decision recorded with its rationale before calculation begins.
  • Statistical safeguards against spurious or overfit economic relationships, not just best-fit selection.
  • Automated consistency checks that catch what manual spreadsheet review misses.
  • Full traceability from the final expected credit loss figure back to its source assumptions.
  • A two-pass approach to replacing a legacy model: reproduce it faithfully first, then apply corrections one at a time, so every change is an explained bridge, not a black box.
Staged methodologyEvery governance decision documented upfront
Statistically validatedEconomic relationships tested, not just best-fit selected
Full traceabilityEvery figure traces back to its source assumptions
Work with us

Bring us the model you currently have to defend alone.

Speak to the team