reserving-ml

Chain-ladder claims reserving on a run-off triangle — development factors, ultimates, IBNR, and a leave-one-diagonal-out backtest.

Open-source insurance-AI starter. A clean, dependency-free implementation of the chain-ladder method: from a cumulative claims run-off triangle it estimates volume-weighted development factors, squares the triangle, and computes ultimates and IBNR (incurred-but-not-reported) reserves — plus a simple backtest so you can sanity-check the projection.

⚠️ Starter / research code. Deterministic chain-ladder only — no tail factors, bootstrap, or Mack standard errors yet (good next steps).

Install

git clone https://github.com/Vaynork/reserving-ml
cd reserving-ml
pip install -e ".[dev]"

No runtime dependencies — pure standard library.

Quickstart

from reserving_ml import ChainLadder, development_factors, ibnr

# cumulative paid claims, ragged run-off triangle (row i observed up to development n-1-i)
triangle = [
    [1000, 1500, 1800, 1980],
    [1100, 1650, 1980],
    [1200, 1800],
    [1300],
]

development_factors(triangle)   # [1.5, 1.2, 1.1]
cl = ChainLadder(triangle).fit()
cl.ultimates()                  # projected ultimate per accident year
cl.ibnr()                       # {'per_accident_year': [...], 'total': ...}

Run the demo: python -m reserving_ml

What it computes

• development_factors — volume-weighted age-to-age factors.
• project — squares the triangle by applying the factors down the run-off.
• ultimates — projected ultimate cost per accident year.
• ibnr — ultimate minus the latest reported diagonal (the held reserve), per year and total.
• backtest_last_diagonal — drops the most recent diagonal, re-predicts it and reports MAE/MAPE.

License

Apache-2.0.