reservingmodels

Claims development and stochastic reserve estimation, from the triangle to the predictive distribution.

Overview

reservingmodels is the single place to reach for claims-reserving work. The deterministic engine — chain-ladder development, completion factors, the Bornhuetter–Ferguson / Benktander / Cape Cod methods, and Mack's analytic standard errors — is re-exported from actuarialpy, where those triangle primitives already live because the projection and pricing packages build on them. On top of them this package adds the stochastic layer: the over-dispersed-Poisson bootstrap of England & Verrall, which produces the full predictive distribution of the unpaid claims, plus residual diagnostics for checking the model before you trust the numbers.

A fitted BootstrapODP exposes .sample(size, rng), so it plugs straight into risksim as a portfolio component — reserve risk aggregated with prospective underwriting risk, with no dependency between the two packages.

Installation

pip install reservingmodels

Requires Python 3.10 or newer. Pulls in actuarialpy for the shared triangle primitives.

Quick start

import reservingmodels as rv

triangle = rv.datasets.taylor_ashe()          # a cumulative K x K triangle

# --- deterministic: the point estimate (re-exported chain ladder) ---------- #
cl = rv.ChainLadder.fit(triangle)
print(cl.project(triangle))                   # per-origin ultimate and IBNR
print(cl.mack_standard_errors(triangle))      # Mack (1993) analytic SE

# --- stochastic: the predictive distribution ------------------------------- #
boot = rv.BootstrapODP.fit(triangle)
dist = boot.reserve_distribution(size=100_000, rng=42)

print(f"reserve (point)   : {dist.point_reserve:,.0f}")
print(f"prediction error  : {dist.prediction_error():,.0f}")
print(f"75th / 99th pct   : {dist.quantile(0.75):,.0f} / {dist.quantile(0.99):,.0f}")
print(f"TVaR(99%)         : {dist.tvar(0.99):,.0f}")
print(dist.to_frame().round(0))               # per-origin + total exhibit

# --- reserve risk as a capital component ----------------------------------- #
import risksim as rs
port = rs.Portfolio([rs.PortfolioItem("reserve_risk", boot)])  # .sample seam
sim = port.simulate(100_000, rng=7)
print(f"portfolio TVaR    : {rs.metrics.tvar(sim.gross_losses, 0.99):,.0f}")

What's inside

• Triangles and deterministic methods — re-exported from actuarialpy: ChainLadder, completion_factors, apply_completion, develop_ultimate (chain-ladder, Bornhuetter–Ferguson, Benktander, Cape Cod), make_completion_triangle, and Mack's mack_standard_errors. These are aliases, not copies — rv.ChainLadder is actuarialpy.ChainLadder.
• Stochastic reserving — BootstrapODP, the semiparametric ODP bootstrap (estimation error by residual resampling, process error by a gamma draw), and ReserveDistribution for percentiles, VaR, TVaR, and a per-origin exhibit.
• Diagnostics — pearson_residuals, residual_summary, and calendar_year_effects for checking the ODP assumptions.
• The risksim seam — BootstrapODP.sample(size, rng) makes the reserve a portfolio component; risksim aggregates it with other risks.
• The Experience seam — reservingmodels.integrations.actuarialpy builds a triangle (and a fitted bootstrap) from a claims-listing actuarialpy.Experience.

The full API reference and worked examples live at openactuarial.org/reservingmodels.html.

The OpenActuarial ecosystem

reservingmodels shares the ecosystem conventions — tidy tables, explicit parameterizations, reproducible random-number handling — and composes across package seams:

Install everything at once with pip install openactuarial.

Development

git clone https://github.com/OpenActuarial/reservingmodels
cd reservingmodels
python -m pip install -e ".[dev]"
pytest
ruff check src tests

CI runs the same gate on Python 3.10–3.14 across Linux and Windows.

Versioning and stability

All ecosystem packages are pre-1.0: minor releases may change APIs, and every release is documented in CHANGELOG.md. Current per-package API stability is tracked at openactuarial.org/stability.html.

License

MIT — see LICENSE.