AI & Machine Learning9 min read

Machine Learning for Individual Claims Reserving: State of Practice

By Jonas Mohamed Osman Abdelghafour

Published 16 September 2026

Where claim-level machine learning models genuinely improve reserving accuracy, and the validation and governance they require.

Executive answer

Claim-level machine learning models predict development or ultimate cost for each open claim from its own features rather than from a cohort average. Evidence from research and practice suggests genuine gains in segmentation, case-reserve adequacy monitoring and early identification of deteriorating claims, with more modest gains in the aggregate reserve itself.

Where the gains are real

The strongest use cases are operational: flagging claims likely to exceed their case reserve, prioritising claims-handling intervention, and producing segment-level views the triangle cannot support. Using the model as the primary aggregate reserve estimate is a larger step and demands considerably more validation evidence.

Known failure modes

Feature leakage from fields updated after the prediction date, non-stationarity when claims-handling practice changes, survivorship effects in training data drawn only from closed claims, and instability of predictions quarter on quarter. Each is detectable with disciplined temporal out-of-sample testing and easy to miss without it.

Governance considerations

A claim-level model used in financial reporting needs documented feature definitions, reproducible training data snapshots, monitoring of drift, explanation of material period-on-period movements, and a stated fallback to a traditional method. Reconciliation to a triangle-based control total remains good practice.

Conclusion

Machine learning has earned a place in reserving, most convincingly as an analytical and operational layer above a governed traditional estimate rather than as a replacement for it.

Primary sources

individual claims reservingmachine learninggradient boosting