Field card · Credit risk
alextouvras.com
Loan origination and IFRS 9

Credit risk is a lifetime, not a cutoff

Decide → watch → stage → hold

Originate, monitor, stage, and provision are layers. Models are lanes. A score at application is not the loss you hold — name SICR before anyone books the number. How the number is built lives on the Analytics card (grain, truth, consume).

ORIGINATEDecision — creditworthiness, policy, engine
MONITORBook — arrears, emerging risk, in-production scores
STAGEIFRS 9 — SICR, stage 1 / 2 / 3
PROVISIONECL — 12-month or lifetime, PD · LGD · EAD, overlay

Problem → use → example

If the real problem is… Use Example case
New credit without a named affordability test Origination standard B2B partner or consumer loan: creditworthiness, data used, and who can override — not a gut “looks fine”
Cut-off set once, never watched In-production scorecard Application PD: cut-off, through-the-door mix, and override rate on a named cadence — a launch deck is not monitoring
Each cut-off trial is a full book backtest Bayesian opt Spend the next portfolio run where last week’s surface still moves the decision. Don’t grid forty cut-offs because the engine can.
Policy lives in a person, not the engine Credit engine + owner Rule or score change in the engine: test, named owner, rollback — chat is not the policy
Book looks calm; early arrears are not Portfolio watch Segment PD vs observed default; vintage and product — steering is a change, not a slide
Everyone is still “performing” SICR / staging Significant increase since origination → lifetime ECL (stage 2); credit-impaired → stage 3. A missed payment is a trigger, not a debate
Allowance is a round number from last year ECL methodology PD × LGD × EAD, 12-month vs lifetime, forward-looking info — a spreadsheet total is not IFRS 9
Model is blind to the downturn you can see Judgement overlay Named overlay, size, and exit test; experienced credit judgement is a control, not a plug
PD/LGD/EAD don’t match the grain of the book Named risk parameters Auto-loan lifetime ECL: account grain, remaining term, prepayment — Analytics card owns the gold table; this card owns the stage
Override rate is “how we actually decide” Override governance Who can stretch policy, why, and how it is sampled — silent overrides are a second scorecard
Capital model and IFRS 9 model never speak Know the two jobs IRB EL ≠ IFRS 9 ECL. Reconcile the difference; don’t paste one into the other

Default order: creditworthiness + policy in the engine → watch the book and the score → stage on SICR → hold 12-month or lifetime ECL (overlay named). Analytics owns grain and gold; this card owns the decision and the stage.

Tool picker

EBA origination GLCreditworthiness, monitoring, model governance
IFRS 9Staging, SICR, 12-month vs lifetime ECL
EBA ECL GLSound ECL practices next to the Standard
BCBS ECL guidanceJudgement, data, and allowance quality
IRB differentiationPD that actually ranks risk — not a label
IFRS 9 supportImplementation notes, not a vendor deck
PD · LGD · EADName the three; freeze the cut you used

Decision vs number

Decision · originate Creditworthiness, policy, engine, override. A score without a cut-off owner is not a decision.
Number · hold Stage, then ECL. Analytics can build the table; this card still owns SICR and the overlay.
Memory check ORIGINATE = decision. MONITOR = book. STAGE = SICR. PROVISION = the loss you hold.

Ladder + gates

  1. Creditworthiness + policy in the engine
  2. Cut-off + override owner
  3. In-production monitoring
  4. SICR / stage 1–3
  5. PD · LGD · EAD at the right grain
  6. Overlay named — or no plug
Kill switch If you can’t name the stage, the parameters, and who owns the overlay, you don’t book the number.

Anti-patterns

Always on

SecurityWho can see PII and flip a policy rule
GovernanceModel, cut-off, and overlay owners
ObservabilityScore, arrears, and stage movement
EvalsBacktest and through-the-door mix
Human ApprovePolicy, SICR, overlay — not auto-booked