resurv-replication-code

This repository contains the code to replicate the simulated case study of the manuscript A MACHINE LEARNING APPROACH BASED ON SURVIVAL ANALYSIS FOR IBNR FREQUENCIES IN NON-LIFE RESERVING.

The computations for obtaining the results in the manuscript were performed on the ERDA cloud (see Appendix Computational Details of the manuscript).

We do not share the private data on which we performed the real data case study.

The repository is organised as follows:

| resurv-replication-code
   |_ Fitting_Scoring
   |    |_ Scoring_results
   |    |_ Simulation_scripts
   |    |_ summarize_simulation_results.R
   |_ ReSurv_cv_results
   |_ cross_validation_scripts
   |    |_ simulation_0 ... simulation_5
   |         |_ bayes_deepsurv.R
   |         |_ bayes_w18.R
   |         |_ bayes_xgboost.R
   |_ slurm_scripts
        |_ simulation_fitting_scoring.sh
        |_ simulation_fitting_scoring_slurm.sh
        |_ slurm_job_cv.sh
        |_ slurm_job_init_cv.sh

Folders

Folder / File	Description
Fitting_Scoring	Orchestrates model fitting, scoring and post-processing.
Fitting_Scoring/Scoring_results	Stores intermediate scoring outputs generated during the simulations.
Fitting_Scoring/Simulation_scripts	Simulation drivers that fit and score chain ladder, Cox, XGBoost and neural network models for each scenario family.
Fitting_Scoring/summarize_simulation_results.R	Collects the simulation outputs and prepares the summary tables reported in the manuscript.
ReSurv_cv_results	Cross-validation results for every simulation scenario and model.
cross_validation_scripts	Hyper-parameter tuning scripts grouped per simulation (simulation_Alpha–simulation_Epsilon). Each folder contains the R scripts listed below.
slurm_scripts	Submission helpers for running the workflow on a Slurm cluster (examples listed below).

Scripts

Script	Location	Description
bayes_deepsurv.R	cross_validation_scripts/simulation_*	Tunes DeepSurv neural network hyperparameters for each simulation setting.
bayes_w18.R	cross_validation_scripts/simulation_*	Tunes W18 Cox model hyperparameters for each simulation setting.
bayes_xgboost.R	cross_validation_scripts/simulation_*	Tunes XGBoost hyperparameters for each simulation setting.
simulation_fitting_and_scoring_scenarios04_cox_xgboost_nn.R	Fitting_Scoring/Simulation_scripts	Fits and scores Cox, XGBoost, and neural network models for scenarios Alpha to Epsilon.
simulation_fitting_and_scoring_cl_methods_scenarios04.R	Fitting_Scoring/Simulation_scripts	Fits and scores the chain ladder benchmarks for scenarios Alpha to Epsilon.
simulation_fitting_and_scoring_zeta_cox_xgboost_nn.R	Fitting_Scoring/Simulation_scripts	Fits and scores Cox, XGBoost, and neural network models for scenario Zeta.
simulation_zeta_fitting_and_scoring_cl_methods.R	Fitting_Scoring/Simulation_scripts	Fits and scores the chain ladder benchmarks for scenario Zeta.
summarize_simulation_results.R	Fitting_Scoring	Aggregates scoring outputs and creates manuscript-ready summary tables.

Running on Slurm

The jobs were executed with the Slurm scheduler. To replicate our workflow:

1. Perform Bayesian hyper-parameter tuning for each dataset in each scenario. Example .sh files for simulation Alpha can be found in slurm_scripts. Submit with sbatch ~/slurm_job_init_cv.sh. The companion script slurm_job_cv.sh shows how to launch the subsequent cross-validation jobs. Both serve as templates for creating similar submission scripts for other simulations.
2. Fit and score the models. Example .sh files that fit and score chain ladder and ReSurv-based models are provided in slurm_scripts. Execute sbatch ~/simulation_fitting_scoring_slurm.sh (or adapt simulation_fitting_scoring.sh for interactive runs).
3. Save the final results in a LaTeX-friendly format. Run Fitting_Scoring/summarize_simulation_results.R after the simulations complete.

R Session Information

We refer to version 1.0.1 of ReSurv, which was used to obtain the paper's results. The initial release of ReSurv can be downloaded here.

# System info (private fields redacted)
Sys.info()
#>  sysname        "Linux"
#>  release        "5.14.0"
#>  version        "(kernel info omitted)"
#>  nodename       "(redacted)"
#>  machine        "x86_64"
#>  login          "(redacted)"
#>  user           "(redacted)"
#>  effective_user "(redacted)"

# R version and session info
sessionInfo()
#> R version 4.4.0 (2024-04-24)
#> Platform: x86_64-conda-linux-gnu
#> Running under: Ubuntu 24.04.2 LTS
#>
#> Matrix products: default
#> BLAS/LAPACK: (path omitted)
#>
#> locale:
#>  [1] LC_CTYPE=C.UTF-8       LC_NUMERIC=C
#>  [3] LC_TIME=C.UTF-8        LC_COLLATE=C.UTF-8
#>  [5] LC_MONETARY=C.UTF-8    LC_MESSAGES=C.UTF-8
#>  [7] LC_PAPER=C.UTF-8       LC_MEASUREMENT=C.UTF-8
#>
#> time zone: Europe/Copenhagen
#>
#> attached base packages:
#> [1] stats graphics grDevices utils datasets methods base
#>
#> loaded via a namespace (and not attached):
#> [1] compiler_4.4.0