Simulation data and code for learning dividend and reinsurance policies in a multivariate renewal risk model
Description
This dataset contains simulation results and Python code supporting the study "Learning dividend and reinsurance policies from claims records in a multivariate renewal risk model" by Lihua Bai and Linyu Miao. The study develops entropy-regularized fitted Q-learning and actor-critic algorithms for joint dividend and proportional reinsurance decisions when claim size and interarrival distributions are unknown. The files include pathwise policy-evaluation outputs, learned control profiles, run-level summaries, experimental settings and random seeds, simulation and learning routines, and plotting scripts. The experiments cover compound Poisson benchmarks, a multivariate renewal model with common claims, distributional misspecification, and the effects of claim-age information, decision timing, claim frequency, and action approximation. The main renewal comparison includes ten independent training samples. All insurance examples are generated by numerical simulation; the collection contains no real policyholder records. Stored outcomes can be used to reproduce the reported figures without retraining. Simulation routines are provided to generate claim histories and repeat the additional learning experiments. See README_data.txt for instructions and FILE_INDEX.txt for individual file descriptions. Large trained-policy pickle files are not included and must be regenerated for calculations that require them.
Files
Steps to reproduce
1. Extract reproducibility.zip into a new directory. Read README.txt and docs/experiment_protocol.txt. Individual file descriptions are provided in docs/FILE_INDEX.txt. 2. Use Python 3.10.9 and install the recorded dependencies with: python -m pip install -r requirements.txt. 3. From the extracted package's root, run: python code/make_figures.py --part all. This creates eight figures and summary tables under generated/ without retraining or overwriting the delivered data. Add --format png to generate PNG figures only. 4. To inspect or repeat the additional training jobs, run: python code/run_experiments.py --dry-run, followed by python code/run_experiments.py --jobs 1. A selected group can be run with --only confirmation, age, timing, frequency, or continuous. New results are stored separately in generated/experiments/. 5. To repeat the fixed-target variance calculation, first regenerate its selected Q policy using: python code/numerics.py main --method q --seed 927301. Then run: python code/variance_check.py. Archived variance outcomes already suffice to reproduce the reported plot.
Institutions
- Nankai UniversityTianjin, Tianjin