Blitz

Blitz is a GPU-accelerated actuarial reserving sandbox written with AMD HIP. It parses Schedule P-style loss triangles, runs thousands of bootstrap chain-ladder simulations in parallel, and validates estimated IBNR reserve distributions against observed lower-triangle outcomes.

The project is built as a portfolio piece around practical actuarial modeling, GPU parallelism, and reproducible reserve validation.

What It Shows

• Conversion of 10x10 paid-loss triangles into compact binary arrays for HIP.
• A naive factor-resampling chain-ladder bootstrap with one GPU thread per bootstrap trial.
• An overdispersed Poisson bootstrap inspired by England-Verrall reserving, including residual resampling and phi diagnostics.
• Reserve distribution outputs for best estimate, P75, P95, P99, coefficient of variation, and a Cost-of-Capital risk margin illustration.
• Python validation scripts that compare bootstrap estimates against true lower-triangle outcomes and posted reserves.

Repository Layout

Path	Purpose
parse_triangles.py	Parses ppauto.csv and comauto.csv into GPU-ready binary triangles.
ibnr_bootstrap.hip	Naive factor-resampling chain-ladder bootstrap kernel.
ibnr_bootstrap_odp.hip	Overdispersed Poisson bootstrap kernel with residual resampling.
validate_ibnr.py	Validates naive bootstrap samples against realized lower-triangle outcomes.
validate_odp.py	Validates ODP bootstrap samples and reports reserve adequacy metrics.
reserve_charts.py	Produces Matplotlib/Seaborn charts for the README and portfolio documentation.
rng_test.hip	Small ROCm random-number sanity check.
ppauto.csv, comauto.csv	Source Schedule P auto-line triangle data used by the parser.
Makefile	Build and workflow shortcuts for the HIP executables.

Requirements

• AMD GPU with ROCm/HIP available on the host.
• hipcc and rocrand available to the compiler.
• Python 3.10+.
• Python packages from requirements.txt.

Install the Python dependency in a local environment:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Build

Build all HIP executables:

make all

If ROCm is installed outside your PATH, pass HIPCC explicitly:

make all HIPCC=/opt/rocm/bin/hipcc

Workflow

Generate the GPU input files from the source CSVs:

python parse_triangles.py

This writes:

• triangles.bin: float32 array with shape (146, 10, 10).
• premiums.bin: float32 array with shape (146, 10).
• companies.txt: company IDs and labels matching the triangle order.
• tri_meta.txt: generated metadata.

Run the naive bootstrap:

./ibnr_bootstrap
python validate_ibnr.py

Run the ODP bootstrap:

./ibnr_bootstrap_odp
python validate_odp.py

The HIP sources currently compile with N_COMPANIES=146, matching the included CSV sample after parse_triangles.py filters to complete 10x10 triangles.

Results

The charts below summarize the GPU workload, reserve calibration, portfolio reserve stack, and model uncertainty diagnostics.

Regenerate the chart PNGs after running the workflow above:

python reserve_charts.py --outdir docs/charts

To include a measured GPU throughput chart, let the module rerun both HIP bootstrap executables and parse their timings:

python reserve_charts.py --benchmark --outdir docs/charts

The generated chart set can include a GPU thread-grid heatmap, throughput bars, calibration lift, portfolio reserve stack, ODP reserve distributions, and uncertainty diagnostics.

Generated Artifacts

Generated binary inputs, bootstrap sample files, validation CSVs, compiled executables, generated chart images, Python caches, and local virtual environments are intentionally ignored by Git. Regenerate them locally with the workflow above.

Notes

This is an educational/portfolio project, not a production reserving system or financial recommendation engine. The kernels prioritize a clear demonstration of GPU parallelism and reserve-model validation over production-grade actuarial governance, data controls, or reporting workflows.