IFRS 9 ECL Engine

A practitioner-grade expected-credit-loss (ECL) engine for IFRS 9 impairment, built end to end across five phases: from a single Stage 1 loan through PD term structures, three-stage classification with a full SICR waterfall, macroeconomic scenario overlays, and finally a forward-looking point-in-time PD derived from a one-factor Vasicek/ASRF model. It is built around the IFRS 9 measurement identity — discounted PD × LGD × EAD — and shows how each regulatory concept (staging, SICR, forward-looking information, probability-weighted scenarios) maps onto code that a credit-risk or impairment team would recognise.

Every phase ships in parallel Python and Excel, with a Streamlit dashboard for interactive exploration and a methodology deck for the narrative. The PD anchors are real (S&P Global), the macro paths are real (EBA 2025 EU-wide stress test), and the regulatory citations are woven in at each site, not bolted on.

What it demonstrates

The rigorous core is replacing the ad-hoc "notch-shift" macro overlay with a single-factor Vasicek/ASRF point-in-time PD, then computing probability-weighted ECL across baseline/adverse/upside scenarios (IFRS 9 §B5.5.42). On the sample book this reproduces the characteristic IFRS 9 dynamics: Stage 3 carries the overwhelming majority of ECL (≈11% coverage versus well under 1% in Stage 1), and the forward-looking scenario weighting lifts ECL roughly 52% above the baseline-only number — exactly the cliff a single-scenario model understates.

Left: ECL by IFRS 9 stage, with coverage ratios. Right: portfolio ECL under each macro scenario and the probability-weighted aggregate (the +52% forward-looking overlay vs baseline). Computed by the deterministic engine — python make_figures.py.

Status

Phase	Scope	Module	Status
1	12-month ECL for a single Stage 1 exposure; S&P 1Y PD anchors	ecl_engine_phase1.py	Complete
2	PD term structures from cumulative defaults; bullet & annuity amortisation; lifetime ECL (Stage 2/3)	ecl_engine_phase2.py	Complete
3	Three-stage classification, SICR waterfall, 30/90-DPD backstops, portfolio aggregation	ecl_engine_phase3.py	Complete
4	Macro scenarios (baseline/upside/adverse), transition-based PD stress, probability-weighted ECL, sensitivity	ecl_engine_phase4.py	Complete
5	Forward-looking PIT PD via single-factor Vasicek/ASRF (the rigorous replacement for the Phase 4 notch-shift)	macro_pd.py	Complete
—	Two-tab interactive dashboard (single-loan + portfolio)	app.py	Complete
—	Parallel Excel implementation, one workbook per phase	IFRS9_ECL_Phase1–4.xlsx	Complete
—	Methodology deck	IFRS9_ECL_Methodology.pptx	Complete

The measurement chain

   rating + exposure + LGD/EAD                         (Phase 1)
            │
            ▼
   PD term structure  (marginal PDs from S&P cumulative)   (Phase 2)
            │            + amortisation → EAD profile
            ▼
   lifetime ECL = Σ_t  discount_t · marginal_PD_t · LGD · EAD_t
            │
            ▼
   staging:  Stage 1 (12m)  →  Stage 2 (lifetime, SICR)  →  Stage 3 (impaired)   (Phase 3)
            │   SICR = relative PD change ∪ absolute PD change ∪ qualitative ∪ 30-DPD
            ▼
   forward-looking overlay:  baseline / upside / adverse                          (Phase 4)
            │   PD stressed by rating migration (notch shift), LGD by downturn uplift
            ▼
   probability-weighted ECL  =  Σ_s  w_s · ECL_s     (IFRS 9 B5.5.42–44)
            │
            ▼
   PIT PD, derived not asserted:  PD(z) = Φ( Φ⁻¹(PD_ttc) − √(ρ/(1−ρ))·z )         (macro_pd)
                where z is the systematic macro factor implied by GDP/unemployment paths

The two-line story for the macro layer: Phase 4 stresses PD by asserting a rating migration of N notches per scenario — the standard, transparent teaching form. macro_pd.py then derives that same stress from first principles: it maps the scenario's GDP and unemployment deviations to a systematic factor z, runs z through the one-factor Vasicek model that underpins the Basel IRB framework, and reads off both the conditional PIT PD and the implied notch shift. The asserted shift becomes a special case of the derived one (and z = 0 returns the through-the-cycle PD exactly, so baseline leaves PD unchanged).

Architecture

ifrs9-ecl-engine/
├── ecl_engine_phase1.py        # 12m ECL, Stage 1; S&P Table 26 1Y PDs (global + Europe)
├── ecl_engine_phase2.py        # PD term structure (marginal from cumulative), amortisation, lifetime ECL
├── ecl_engine_phase3.py        # Stage 1/2/3 classification + SICR engine + DPD backstops + portfolio ECL
├── ecl_engine_phase4.py        # Macro scenarios, transition-based PD stress, downturn LGD, weighted ECL
├── macro_pd.py                 # Single-factor Vasicek/ASRF PIT-PD; macro→z bridge; implied notch shift
├── app.py                      # Streamlit dashboard (Tab 1: single-loan; Tab 2: portfolio)
├── IFRS9_ECL_Phase1.xlsx       # Parallel Excel — single-loan 12m ECL
├── IFRS9_ECL_Phase2.xlsx       # Parallel Excel — PD term structure + lifetime ECL
├── IFRS9_ECL_Phase3.xlsx       # Parallel Excel — staging + portfolio aggregation
├── IFRS9_ECL_Phase4.xlsx       # Parallel Excel — scenarios + probability-weighted ECL
├── IFRS9_ECL_Methodology.pptx  # Methodology deck
├── tests/                      # pytest suite
│   ├── conftest.py
│   ├── test_phase1_2.py        # 12m + lifetime ECL, term-structure construction
│   ├── test_phase3_4.py        # staging, SICR triggers, scenario weighting
│   └── test_macro_pd.py        # Vasicek monotonicity, z=0 ⇒ TTC, calibration
├── requirements.txt            # streamlit, matplotlib, numpy
├── requirements-dev.txt        # pytest
└── sources.md                  # Data provenance + methodology references

A deliberate engineering choice: the only runtime dependencies are numpy, matplotlib, and streamlit. The normal CDF/inverse-CDF used by the Vasicek model are implemented in closed form in macro_pd.py (Abramowitz–Stegun / Acklam), so there is no SciPy dependency — the engine is self-contained and the statistics are visible rather than hidden behind a library call.

The SICR waterfall (Phase 3)

A loan moves from Stage 1 (12-month ECL) to Stage 2 (lifetime ECL) on a significant increase in credit risk since origination. The engine tests four triggers in sequence and stages on the first that fires:

1. Quantitative — relative — the lifetime PD has risen by more than a configurable multiple of its origination value (IFRS 9 B5.5.17–18).
2. Quantitative — absolute — an absolute PD increase threshold, to catch high-quality names where a large relative jump is still a small absolute one.
3. Qualitative — watchlist / forbearance / covenant flags.
4. 30-DPD backstop — the rebuttable presumption of IFRS 9 §5.5.11.

Stage 3 (credit-impaired) is triggered by the 90-DPD default definition of CRR Art. 178 or an explicit default flag. Migration distance is measured in notches on the full S&P scale, so the staging logic is rating-aware rather than PD-only.

Quick start

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Each phase runs standalone and prints a worked example:
python ecl_engine_phase1.py     # 12-month ECL, single Stage 1 loan
python ecl_engine_phase2.py     # PD term structure + lifetime ECL
python ecl_engine_phase3.py     # staging + SICR + portfolio ECL
python ecl_engine_phase4.py     # scenarios + probability-weighted ECL
python macro_pd.py              # Vasicek PIT-PD + scenario calibration

# Interactive dashboard:
streamlit run app.py

Limitations

The modelling choices a validator would want flagged:

• PD calibration uses agency through-the-cycle averages. S&P 1-year PDs are mapped to a point-in-time view via the Vasicek model; the asset correlation ρ and the macro-to-systemic-factor mapping are calibrated to EBA scenario severity, not estimated from the bank's own default history.
• LGD is exogenous and largely static. Downturn LGD is applied only where set; collateral dynamics and cure rates are not modelled.
• Three discrete scenarios approximate the macro distribution. Fixed probabilities on baseline/adverse/upside stand in for the full distribution; tail states beyond "adverse" are not represented.
• SICR thresholds are policy choices. The notch-downgrade and 30/90-DPD backstops are judgemental — the standard deliberately leaves them to the institution.
• No behavioural modelling. Amortisation is bullet/annuity only; prepayment and revolving-exposure behaviour are out of scope.

Run the test suite with:

pip install -r requirements-dev.txt
pytest -q

The 41-test suite also runs automatically on every push via GitHub Actions (.github/workflows/tests.yml).

The Excel workbooks (IFRS9_ECL_Phase1–4.xlsx) mirror each Python phase cell-for-cell, so the same numbers can be traced in a spreadsheet — useful for walking a non-technical reviewer through the calculation.

Headline figures used (real anchors)

Input	Source	Example value
1Y PD, BBB / BB / B / CCC (global)	S&P 2024 Default Study, Table 26	0.13% / 0.44% / 2.69% / 26.12%
1Y PD, BBB / BB / B / CCC (Europe)	S&P 2024 Default Study, Table 25	0.05% / 0.36% / 1.75% / 26.26%
LGD (senior unsecured / subordinated)	CRR Art. 161 (F-IRB)	45% / 75%
Adverse macro path (EU, 2025–27)	EBA 2025 EU-wide stress test	GDP −6.3%, unemployment +6.1pp, equities −50%

Known simplifications

Deliberate scope choices, documented inline, not oversights:

• PDs are through-the-cycle anchors (S&P long-run averages). The point-in-time adjustment lives in the macro layer; Phases 1–3 use the TTC values directly, which is the standard academic/interview baseline.
• LGD is a regulatory anchor with a downturn uplift, not an empirically calibrated curve. sources.md lists the EBA IRB benchmarking and Banque de France recovery studies that would replace it in a production calibration.
• Scenario probability weights are expert-set (illustrative), with a sensitivity panel showing how the weighted ECL moves as the adverse weight is varied — rather than estimated from a transition model.
• The macro→z bridge in macro_pd.py uses linear sensitivities to GDP and unemployment deviations; the asset correlation ρ follows the Basel corporate formula. Both are visible parameters, not fitted.

Regulatory references

• IFRS 9 — Financial Instruments (IASB). Key paragraphs: §5.5.3–5.5.5 (staging & measurement), §5.5.9 / §5.5.11 (SICR & 30-DPD presumption), §5.5.17(c) (forward-looking information), §B5.5.17–18 (relative vs absolute PD change), §B5.5.42–44 (probability-weighted ECL), Appendix A (credit-impaired).
• CRR (Regulation (EU) No 575/2013) — Art. 161 (F-IRB LGD), Art. 178 (definition of default / 90-DPD backstop).
• EBA GL/2017/06 — Guidelines on ECL accounting practices for credit institutions (scenario design and weighting, §30–32).
• EBA 2025 EU-wide stress test — macro-financial scenario (Jan 2025) and methodological note (staging under stress, §73–75).
• ACPR — French supervisory expectations on IFRS 9 implementation.

See sources.md for the full provenance of every PD, LGD, and macro input, including document titles, dates, tables, and URLs.