Enhance Risk Assessment: Machine learning for LGD/EAD

Why this matters for IFRS 9 ECL

Under IFRS 9, the accuracy of ECL depends on robust estimates of Probability of Default (PD), LGD and EAD. LGD and EAD determine the scale and timing of expected losses and therefore directly affect provisions, capital planning and reported profitability. Machine learning for LGD/EAD offers potential improvements in predictive power, segmentation and scenario responsiveness, but it also introduces new model validation, transparency and governance requirements.

For credit risk and finance teams, failing to improve LGD/EAD estimates can lead to two common pain points: (1) volatile provisioning with swings to profit and loss that stress earnings and capital ratios, and (2) insufficient sensitivity testing and explanation at the Risk Committee and auditor level. Machine learning techniques can reduce error, but require disciplined Risk Model Governance and clear Risk Committee Reports to be accepted by auditors and regulators.

Core concepts: LGD and EAD with ML

Definitions and targets

LGD is the percentage loss the lender suffers if a borrower defaults, after recoveries and net of costs. EAD is the expected exposure at the time of default — for revolving products this includes projected drawdowns. In ML projects you must define precise target variables (e.g., realized discounted loss for LGD, cumulative drawdown for EAD) and align them with your ECL Methodology and accounting treatment.

Model types and techniques

Common machine learning models used for LGD/EAD include gradient boosting (XGBoost, LightGBM), random forests, penalised linear models and neural networks. For EAD and recovery timing, survival analysis and time-to-event models — including ML-enhanced survival forests — often work well. For contrast with classical methods, see the approaches in statistical ECL modeling.

Feature engineering and data sources

Feature engineering is critical: include loan characteristics (product type, seasoning, collateral), borrower metrics, macroeconomic indicators and behavioural signals (transactional patterns). When possible, enrich models with vendor or alternative data; read about successful initiatives that use big data in ECL models to increase discriminative power. Maintain unambiguous definitions so the same feature used in validation and production yields consistent results.

Interplay with PD and overall ECL

LGD and EAD predictions are inputs to the full ECL calculation. Coordination with the PD model effort is essential — for example, segmentation should be consistent across PD and LGD where reasonable. Where PD teams use ML approaches, cross-team knowledge-sharing is useful; consider techniques described in AI for PD modeling to harmonize modelling choices.

Explainability

Use SHAP values or partial dependence plots to explain drivers at portfolio and loan levels; explainability is critical for model validation and for Risk Committee Reports.

Practical use cases and scenarios

1. Retail secured book — improving LGD estimates

Scenario: a bank with 120,000 mortgage accounts seeks to refine LGD for loss provisioning. Approach: use ML to predict discounted recovery rates by combining property valuations, regional house-price indices and borrower arrears history. Outcome: decrease in out-of-sample LGD root-mean-square error (RMSE) by ~12–18% compared to the current regression model; provisioning volatility reduced.

2. Credit cards — forecasting EAD for revolvers

Scenario: a card issuer struggles to project committed line utilisation at default. Approach: train an LSTM or gradient boosting model with sequences of utilisation and payment behaviour, include macro stress multipliers for forward-looking scenarios. Outcome: improved coverage of tail exposure and more defensible stress adjustments in sensitivity testing used for regulatory reporting.

3. SME portfolio — dealing with limited defaults

Scenario: thin default history for small and medium enterprises. Approach: borrow strength via transfer learning, use pooled models with entity-level hierarchical features and incorporate external recovery datasets from industry partners or fintech sources highlighted in AI–FinTech for ECL prediction. Outcome: better stability across vintages and faster detection of regime shifts.

Impact on decisions, performance and profitability

Better LGD/EAD estimates translate into more accurate provisions, improved capital forecasting and clearer insight into product profitability. Example: for a corporate loan book with expected exposure £500m, a 5 percentage point reduction in average LGD (e.g., from 40% to 35%) reduces expected losses by £25m, which after tax materially increases retained earnings and affects pricing decisions.

Accounting impact on profitability

Changes in LGD/EAD feed directly into ECL and therefore P&L and retained earnings. When ML models change parameter estimates materially, the finance team must document the accounting rationale and run comparatives: quantify movement in Stage 1/2/3 provisions, show forecasted P&L impact and disclose methodology changes to auditors. Sensitivity Testing should be run to show the range of P&L outcomes under stress paths.

Capital planning and strategic decisions

More stable ECL projections improve capital allocation. A bank that reduces provisioning variability can free capital for lending or investment, improving return-on-equity. Use scenario outputs from ML models to run “what-if” analyses for pricing, limit setting and portfolio rebalancing.

Common mistakes and how to avoid them

• Relying on correlation without causal or forward-looking reasoning — ensure features are predictive of future losses, not only historically correlated.
• Ignoring data leakage — prevent use of post-default signals when training LGD/EAD models by strict event-time alignment.
• Poor documentation of ECL Methodology changes — always version and record changes and their expected accounting impact.
• Insufficient sensitivity testing — include systematic Sensitivity Testing across macro scenarios and model assumptions to show robustness.
• Weak model governance — embed thorough validation and regular backtesting procedures as part of Risk Model Governance to satisfy auditors.

Model validation should include benchmark comparisons to traditional approaches and stress checks such as PSI, calibration charts and case-level reviews. For modelling rules and formal procedures, follow the guidance in ECL modeling best practices.

Practical, actionable tips and checklists

Step-by-step implementation checklist

1. Define target variables precisely (discounting approach for LGD, horizon and drawdown assumptions for EAD).
2. Assemble a cross-functional team: modelling, data engineering, accounting, compliance, validation and business owners.
3. Build a reproducible data pipeline with time-stamped snapshots and lineage; include macro and alternative datasets.
4. Create baseline statistical models as benchmarks before deploying ML — compare using validation metrics and stability tests.
5. Perform sensitivity testing across at least three macro scenarios and document the P&L and capital impact.
6. Apply explainability tools (SHAP, LIME) and prepare simple narratives for the Risk Committee Reports.
7. Organize an independent model validation with backtesting, stress tests and a pre-deployment checklist.
8. Set up monitoring: PSI, population stability, prediction drift and regular reconciliation to observed outcomes.

Model validation and ongoing monitoring

Model Validation should include: holdout and time-series cross-validation, calibration plots, quantile accuracy for LGD, survival curve diagnostics for EAD timing, and scenario-level stress tests. For skills and team readiness, invest in training guided by the recommendations at AI/ML skills for ECL teams.

Reporting to the Risk Committee

Design concise Risk Committee Reports that include: model purpose, data sources, key drivers, validation results, sensitivity testing outcomes and the quantifiable accounting impact. Provide clear operational controls and a remediation plan for known limitations.

KPIs & success metrics

• Predictive accuracy (RMSE or MAE for LGD, C-index or time-dependent AUC for EAD timing)
• Calibration error (mean absolute calibration, calibration-in-the-large)
• Population Stability Index (PSI) and feature drift measures
• Provision variance reduction (year-on-year volatility of ECL)
• Coverage of stress scenarios in sensitivity testing
• Time to produce Risk Committee Reports and regulatory submissions
• Model runtime and deployability (retraining frequency capability)

FAQ

Reference pillar article

This article is part of a content cluster on technology and ECL. For broader context and strategic guidance on whether traditional methods are enough and how technology supports IFRS 9, see the pillar article: The Ultimate Guide: The role of technology in developing ECL calculations – are traditional methods enough, and how tech solutions support IFRS 9 requirements.

Looking ahead: adoption and the future of model-driven ECL

Expect continued convergence between machine learning and traditional econometric approaches. The future of AI in credit risk will see more hybrid models, automated feature engineering and tighter integration of alternative data. Keep an eye on vendor ecosystems and fintech partnerships that accelerate data acquisition and model deployment while remaining compliant.

Governance, team skills and vendor considerations

Risk Model Governance must cover lifecycle management, documentation, validation and monitoring — with clear roles for model owners, validators and business owners. When sourcing external models or tools, do due diligence on data lineage, versioning and reproducibility. Developing internal capability is essential; combine domain risk expertise with data science skills — for recommended curricula and team roles, review AI/ML skills for ECL teams.

When coordinating with product and finance teams, ensure the change management plan addresses the Accounting Impact on Profitability and that all material methodology changes are pre-approved by finance and disclosed appropriately.