Neural network reserving for bodily injury claims
A general insurer is replacing its chain-ladder reserving process for bodily injury with a neural network. The dataset has 18,000 claims, 60 development months and 1.2% of claims above £1m. The network predicts ultimate cost per claim and, when aggregated, is 3.8% lower than chain-ladder for the latest accident year. Backtesting used a random 20% holdout, and the model was trained after a COVID-related reporting delay. The network's error is small for small claims but it under-predicts the largest 50 claims by 18% on average. The reserving committee wants a single best estimate and a clear explanation of uncertainty. The model uses 45 features including injury type, age, solicitor, rehabilitation and payment delay. Two accident years have sparse data. The regulator requires reserves to be adequate with a high degree of confidence.
1. Assess whether the neural network is suitable as the sole basis for the best estimate.
Read the worked answer
The network is promising as a diagnostic or challenger, not yet as the sole best estimate. A random 20% holdout overstates out-of-time performance and mixes accident years; the COVID reporting delay means recent development patterns are not representative. The model is good for small claims but under-predicts the largest 50 by 18%, exactly where reserves are most sensitive. Aggregate under-reserving of 3.8% may be material. Sparse accident years and tail claims need explicit treatment: stratified validation by size and accident year, tail-focused loss functions or a separate large-claim model, and blending with chain-ladder or expert judgement. A single best estimate from the network would not satisfy the requirement for adequacy with high confidence. Use it to challenge assumptions, decompose the reserve and set ranges.
2. Design a validation plan focused on tail risk and uncertainty.
Read the worked answer
Validate out-of-time by accident year and development period; hold out entire accident years, not random claims. Measure aggregate and claim-level error, calibration by size band, tail quantiles and reserve adequacy. Backtest on pre-COVID and post-COVID periods; test reporting-delay sensitivity. Use bootstrap, ensemble or quantile outputs to produce a reserve distribution, and compare with chain-ladder ranges. Stress the largest claims and sparse years. Check stability of feature effects, especially solicitor and rehabilitation, which may change behaviour. Independent review should reproduce the data and model. Set monitoring for claim mix, settlement speed and large-claim emergence.
3. How would you present the choice to the reserving committee, including how to communicate uncertainty?
Read the worked answer
Explain that the committee is being asked to choose a best estimate, but the evidence supports a range and a blended approach. Present chain-ladder, network and blended estimates, with the 3.8% gap and the 18% tail under-prediction as key sensitivities. Recommend using the network for small claims where it is stable and retaining chain-ladder or expert judgement for the tail, or applying a tail loading. Disclose model risk and the COVID data limitation. A single number can be proposed only with a documented margin and committee sign-off; otherwise use the existing method plus the network as a challenger. Communication should state what would change the estimate: large-claim development, reporting delays or accident-year mix.