insurance-conformal-risk 0.1.0

Conformal Risk Control for insurance pricing: distribution-free expected-loss guarantees

insurance-conformal-risk

Conformal Risk Control for UK insurance pricing. Distribution-free bounds on expected financial shortfall — not just coverage probability.

pip install insurance-conformal-risk

The problem

Your GBM predicts a pure premium of £450 for a motor policy. The actual claim comes in at £1,200. That's an underpricing event.

Standard conformal prediction (see insurance-conformal) tells you: "the claim will fall below your upper bound on 90% of policies." It says nothing about what happens in the other 10% — the shortfall could be £10 or £10,000.

Conformal Risk Control (CRC) controls the magnitude directly:

E[max(claim - upper_bound, 0) / premium] ≤ α

This bounds the expected underpricing shortfall as a fraction of premium income. With α = 0.05, you are guaranteeing that expected shortfall from underpriced policies is at most 5% of expected premium income — no parametric assumptions, finite-sample valid.

This is what actuaries actually want to know. Not "how often am I wrong?" but "how much does being wrong cost me?"

Background

This library implements Conformal Risk Control from Angelopoulos et al. (2024), ICLR, applied specifically to insurance pricing problems.

CRC extends split conformal prediction from coverage control to expected-loss control for any bounded monotone loss function. The algorithm is five lines of numpy. The value is in the correct insurance-specific loss functions, the finite-sample correction that naive implementations get wrong, and the workflow that maps onto how pricing teams actually work.

No Python package on PyPI implements general regression risk control with user-defined monotone losses. This is the first.

Three controllers

1. Premium Sufficiency (main use case)

Find the smallest loading factor such that expected shortfall is bounded:

from insurance_conformal_risk import PremiumSufficiencyController
import numpy as np

# y_cal: observed claims on held-out calibration set (n=1000-5000)
# premium_cal: model-predicted pure premiums for same policies
psc = PremiumSufficiencyController(alpha=0.05, B=5.0)
psc.calibrate(y_cal, premium_cal)

print(psc.lambda_hat_)  # e.g., 1.34 — load all premiums by 34%

# Apply to next year's book
result = psc.predict(premium_new)
# result["upper_bound"] = 1.34 * premium_new (risk-controlled bound)

The guarantee: E[max(claim - 1.34 × premium, 0) / premium] ≤ 0.05 on any exchangeable test set.

On setting B: B is the maximum possible normalised shortfall (max claim / min premium). For a policy limit of £50,000 and minimum premium of £200, B = 250. If you normalise by premium and your claims are bounded by the sum insured, B is well-defined. The default B=1 is only valid if claims never exceed premium — rarely true. Inspect your data.

2. Interval Width Control

Find the tightest prediction intervals that still keep expected width below a budget:

from insurance_conformal_risk import IntervalWidthController
import numpy as np

# widths_cal[i, j] = interval width for observation i at quantile level lambda_grid[j]
# (generate this by calling your conformal predictor at each lambda value)
lambda_grid = np.linspace(0.50, 0.995, 100)
controller = IntervalWidthController(width_target=800.0, scale=2000.0, lambda_grid=lambda_grid)
controller.calibrate_from_widths(widths_cal)

print(controller.lambda_hat_)  # e.g., 0.82 — use 82nd percentile intervals

3. Selective Underwriting

Accept only risks where expected loss on the accepted book is bounded:

from insurance_conformal_risk import SelectiveRiskController
import numpy as np

def large_claim_loss(y, scores):
    """Binary: 1 if claim exceeds £5,000."""
    return (y > 5000).astype(float)

src = SelectiveRiskController(alpha=0.08, loss_fn=large_claim_loss, xi_min=0.60)
src.calibrate(y_cal, scores_cal)
# src.threshold_: accept iff risk_score >= threshold

decisions = src.predict(scores_new)
# decisions["accept"]: True/False per policy

The guarantee: among accepted risks, E[large_claim_loss] ≤ 0.08, provided at least 60% of risks are accepted.

Integration with insurance-conformal

These two libraries work together. Use insurance-conformal to generate coverage-controlled intervals, then use insurance-conformal-risk to verify premium sufficiency:

from insurance_conformal import InsuranceConformalPredictor
from insurance_conformal_risk import PremiumSufficiencyController

# Step 1: standard conformal intervals (coverage control)
cp = InsuranceConformalPredictor(model=fitted_gbm, nonconformity="pearson_weighted")
cp.calibrate(X_cal, y_cal)
intervals_cal = cp.predict_interval(X_cal, alpha=0.10)

# Step 2: risk control on top of conformal upper bounds
psc = PremiumSufficiencyController(alpha=0.04, B=8.0)
psc.calibrate(y_cal, intervals_cal["upper"].to_numpy())

# The conformal upper bound is both coverage-controlled AND shortfall-controlled
intervals_new = cp.predict_interval(X_new, alpha=0.10)
bounds = psc.predict(intervals_new["upper"].to_numpy())

Regulatory framing

For Solvency II (Article 105) and Solvency UK model validation:

from insurance_conformal_risk.reporting import (
    premium_sufficiency_report,
    solvency_ii_model_error_note,
)

report = premium_sufficiency_report(
    lambda_hat=psc.lambda_hat_,
    alpha=psc.alpha,
    n_calibration=psc.n_calibration_,
    B=psc.B,
    corrected_risk=psc.risk_summary()["corrected_risk_at_lambda"],
    portfolio_gwp=45_000_000,  # £45m GWP
)

note = solvency_ii_model_error_note(psc.alpha, psc.lambda_hat_, psc.n_calibration_)
print(note)

Core algorithm

The CRC algorithm (Algorithm 1 of Angelopoulos et al. 2024):

1. Compute empirical risk: R̂(λ) = (1/n) Σ L_i(λ)
2. Apply finite-sample correction: (n/(n+1)) × R̂(λ) + B/(n+1) ≤ α
3. Find λ* = smallest λ satisfying the corrected inequality

The finite-sample correction is not optional. It accounts for the unseen test point. For n=500, B=5, the correction adds B/(n+1) ≈ 0.01 to the risk threshold — small but load-bearing for tight alpha values.

from insurance_conformal_risk import conformal_risk_calibration
import numpy as np

# losses[i, j] = loss for observation i at lambda_grid[j]
# Must be non-increasing in j (larger lambda = lower loss)
losses = np.random.rand(500, 200) * (1 - np.linspace(0, 1, 200))

lambdas = np.linspace(0, 2, 200)
lambda_hat, idx, risk_curve = conformal_risk_calibration(losses, lambdas, alpha=0.05, B=1.0)

Limitations (be explicit about these)

• Marginal guarantee only. CRC controls the average over the calibration distribution. A particular segment (young drivers, high-value properties) may have higher shortfall than alpha. Check shortfall_report() for segment diagnostics.
• Exchangeability required. Calibration and test data must be exchangeable. For insurance, this means same underwriting year, same distribution mix. Year-on-year deployment with changing book mix violates this. There are extensions for non-exchangeable data (arXiv:2310.01262) — not implemented here.
• B must be set correctly. Setting B too small produces invalid guarantees (the algorithm will still run; the guarantee will not hold). B is the maximum possible loss value. For the shortfall loss, this is max_claim / min_premium in your data. For unlimited policies, cap using a policy limit.
• Not a replacement for Solvency II SCR. CRC controls expected shortfall. Solvency II requires 99.5% VaR. These are different quantities. CRC bounds model error, not the full underwriting risk capital charge.

What this is not

This is not conformal prediction (coverage control) — that's insurance-conformal. Use standard conformal prediction when you want P(Y in interval) ≥ 1 - α. Use this library when you want E[financial_loss] ≤ α.

Installation

pip install insurance-conformal-risk

With scikit-learn for SelectiveRiskController integration:

pip install insurance-conformal-risk[sklearn]

References

• Angelopoulos, Bates, Fisch, Lei & Schuster (2024). Conformal Risk Control. ICLR 2024. arXiv:2208.02814
• Selective CRC: arXiv:2512.12844 (2025)
• Hong (2025). Conformal prediction of future insurance claims. arXiv:2503.03659

License

MIT. Built by Burning Cost.