One-Shot Individual Claims Reserving

Battle Count: 60

This paper is squarely in the domain of actuarial science and insurance claims reserving. While it uses statistical and ML methods (regression, neural networks, transformers), the application is entirely focused on predicting insurance claim ultimates and reserves. There is no direct connection to financial markets, trading strategies, or asset pricing. The recursive forecasting structure and uncertainty quantification via bootstrap have tangential methodological parallels to quantitative finance, but the domain application is unrelated.

The core algorithm (Algorithm 3) is conceptually straightforward: recursive one-shot PtU forecasting using regression models. Linear regression implementation is very simple (Listing 2). However, the full pipeline involves: (1) data restructuring for individual claims with IBNR masking, (2) recursive iteration over development periods, (3) consistent cohort selection, (4) balance property enforcement for ML models, (5) bootstrap uncertainty quantification, and (6) IBNR prediction via cross-classified CL. Neural network and transformer implementations add complexity. The paper provides complete R code for all components.

The paper provides detailed R code listings (Listings 1-6) for all algorithms including linear regression, FNN, and transformer implementations. However, the datasets (accident insurance with 66,639 claims and liability insurance with 21,991 claims) are not explicitly stated as publicly available. The small-scale 5x5 triangle examples serve as proof of concept. Hyperparameters for neural networks are documented in Tables 11-12.