burncost

Python library for ratemaking-ready loss-experience preparation. Loss development (chain-ladder, Bornhuetter-Ferguson), trend estimation, and premium on-leveling via the parallelogram method. Designed to chain into a downstream rate model.

Installation

pip install burncost

From source:

git clone https://github.com/CosmikArt/burncost.git
cd burncost
pip install -e ".[dev]"

Quickstart

import matplotlib
matplotlib.use("Agg")  # headless-safe; remove if you want an interactive window

import numpy as np
import pandas as pd
from burncost import (
    LossTriangle,
    DevelopmentFactors,
    TrendEstimator,
    OnLevelPremium,
    BurningCostAnalysis,
)

# --- 1. Loss development ---------------------------------------------------
# Cumulative paid-loss triangle (AY x development month)
triangle_data = np.array([
    [1_000, 1_500, 1_800, 1_900],
    [1_100, 1_650, 1_980,   np.nan],
    [1_200, 1_800,   np.nan, np.nan],
    [1_300,   np.nan, np.nan, np.nan],
])
accident_years = [2022, 2023, 2024, 2025]
dev_periods    = [12, 24, 36, 48]

tri = LossTriangle(triangle_data, accident_years, dev_periods)
dev = DevelopmentFactors.from_triangle(tri)

ata   = dev.age_to_age()           # age-to-age factors per interval
atu   = dev.age_to_ultimate()      # cumulative to ultimate
sel   = dev.selected_factors()     # weighted-average selected factors
tail  = dev.tail_factor()          # tail beyond last observed period

# --- 2. Trend estimation ---------------------------------------------------
years     = np.array([2018, 2019, 2020, 2021, 2022, 2023, 2024, 2025])
severity  = np.array([4200, 4400, 4550, 4800, 5100, 5350, 5600, 5900])

trend = TrendEstimator()
trend.fit(years, severity, method="exponential")
annual_trend  = trend.trend_factor(from_year=2022, to_year=2026)
projected     = trend.predict(np.array([2026, 2027]))
trend.plot()

# --- 3. Premium on-leveling -----------------------------------------------
rate_changes = pd.DataFrame({
    "effective_date": pd.to_datetime(["2021-07-01", "2023-01-01", "2024-04-01"]),
    "change":        [0.05, 0.08, -0.03],
})

onlevel = OnLevelPremium()
onlevel.apply_rate_changes(rate_changes)
olf = onlevel.on_level_factor(policy_year=2022)

# --- 4. End-to-end burning cost pipeline -----------------------------------
bc = BurningCostAnalysis(
    triangle=tri,
    trend_estimator=trend,
    on_level=onlevel,
)
# bc.run() will wire development -> trending -> on-leveling into one table.

# --- 5. Diagnostics --------------------------------------------------------
from burncost.diagnostics import trend_fit_summary, development_stability
print(trend_fit_summary(trend))
print(development_stability(dev))

Modules

The five top-level classes are re-exported from the package root, so from burncost import X works for all of them. Internally each concern lives in its own module.

References

• Werner, G. & Modlin, C. Basic Ratemaking, CAS, Chapters 4-5 (trending and loss development).
• Mack, T. (1993). "Distribution-free Calculation of the Standard Error of Chain Ladder Reserve Estimates." ASTIN Bulletin, 23(2), 213-225.
• Bornhuetter, R. L. & Ferguson, R. E. (1972). "The Actuary and IBNR." Proceedings of the CAS, LIX, 181-195.
• Friedland, J. Estimating Unpaid Claims Using Basic Techniques. Casualty Actuarial Society.

Contributing

Run pytest before sending a PR.

Author

Isaac López

MIT License. See LICENSE.