On the use of case estimate and transactional payment data in neural networks for individual loss reserving

Battle Count: 77

The paper is primarily focused on actuarial loss reserving in insurance, which is a distinct domain from quantitative trading. However, there are tangential connections: the use of LSTM for sequential/time-series prediction, the handling of skewed distributions, the train-test split methodology avoiding data leakage, and the general framework for comparing model architectures on complex financial data could inform quantitative finance applications. The risk management and capital allocation aspects have some overlap with portfolio risk assessment.

The models themselves (FNN and LSTM) are standard deep learning architectures implementable in PyTorch or TensorFlow. However, the full pipeline involves: (1) data simulation using R packages SynthETIC/SPLICE, (2) careful train-test splitting by finalisation time to avoid data leakage, (3) feature engineering for summary statistics, (4) log-transform and normalisation of targets, (5) non-parametric bias correction (Duan's smearing estimator), (6) grid search hyperparameter tuning across multiple configurations, (7) evaluation across 50 datasets with multiple metrics. The paper reports approximately 300 hours of compute for full tuning and training. The code is publicly available, reducing implementation burden.

All code and data are publicly available on Zenodo (https://zenodo.org/records/18005906) and GitHub (https://github.com/agi-lab/reserving-RNN). Datasets were simulated using open-source R packages SynthETIC and SPLICE. Modelling was conducted in Python. Detailed hyperparameter tables, training procedures, and evaluation scripts are provided. The paper specifies hardware used (Azure ML notebook, Nvidia RTX 3070 GPU, AMD Ryzen 5 5600X CPU) and approximate runtimes (24 hours for data creation, 300 hours for tuning/training, 24 hours for evaluation).