Machine Learning for Insurance Reserving
Automatic detection of claims requiring closure
A supervised-learning study developed at SADA Assurances to support the prioritisation of open non-life insurance claim files likely to require closure.
Prioritise review without automating the claim decision.
The project focused on a practical reserving question: when many claim files remain open, can a model help identify those that deserve earlier review? The proposed role was triage and prioritisation, not autonomous closure.
Define the target, prepare historical data, compare models.
The workflow linked target definition, historical claims preparation, feature engineering and supervised-model comparison. Evaluation considered precision, recall, specificity, AUC and balanced accuracy so that different error patterns remained visible.
A decision-support case, assessed through more than one score.
The dissertation received 17/20. The case file is structured around operational interpretation: how a predictive signal could support a review queue, what its scope is, and why a good model score is not sufficient on its own.
What it demonstrates: connecting a business decision, a defined target and an interpretable evaluation framework.