From Explaining to Enabling: Actionable Transparency in Algorithmic Insurance Pricing

Agathe Fernandes Machado, Fei HuangarXiv 2026 · arXiv

TL;DR

This paper moves algorithmic transparency in insurance from "explaining the past" to "enabling the future." By combining Counterfactual Explanations with Conformal Prediction, the authors provide policyholders with actionable advice to lower their premiums that is also statistically reliable, ensuring that suggested changes aren't just artifacts of model uncertainty.

What Problem Does This Paper Address?

Machine Learning (ML) models now set insurance premiums with high precision, but they remain "black boxes" to the average consumer. Current Explainable AI (XAI) tools like SHAP or LIME tell you why your premium is high (e.g., "your age contributed +$200"), but they don't tell you how to lower it. Furthermore, even when models do suggest changes—like moving your property's contents to a higher floor—those suggestions might be "lucky guesses" by the model in regions of the data where it has very little experience.

The authors identify a Reliability Gap: a model might predict a lower premium for a specific change, but because the model is uncertain about that region of the feature space, the actual risk (and thus the actual premium) might not decrease at all. This is especially problematic for high-risk individuals who often live in the "data-sparse" outskirts of a model’s training set.

Methods

The authors formalize the requirements for "useful" recourse into three axioms:

  1. Validity: The model must actually predict a lower premium for the suggested change.
  2. Actionability: The suggestion must only involve things the user can actually change (e.g., "install a security system," not "be five years younger").
  3. Reliability: The actual outcome at the suggested change must be likely to meet the target, even accounting for model uncertainty.

To solve for Reliability, the authors integrate Conformal Prediction (CP). Unlike standard predictions that give a single number, CP generates a range (a prediction interval) that is guaranteed to contain the true value with a certain probability (e.g., 80%).

The core innovation is a CP-constrained search. Instead of just looking for the "closest" way to lower a premium, the algorithm filters for recommendations where the upper bound of the uncertainty interval is below the target premium. This forces the recommendation to move toward "denser" regions of the data where the model is more confident. They adapt four existing counterfactual techniques—WhatIf, NICE, MOC, and DiCE—to handle these new constraints in a regression (rather than classification) setting.

Datasets

The framework was evaluated on two distinct insurance contexts:

  • Flood Insurance (Severity Model): 57,832 real-world Florida flood claims from FEMA (2023–2024). This tested the model's ability to suggest property mitigations (like elevating a building) to reduce the "loss ratio."
  • Driver Telematics (Frequency Model): 98,058 synthetic policies based on real Canadian telematics data. This focused on driving behaviors (like rush-hour driving or extreme braking) to reduce claim frequency.

Key Insights and Ideas

  • The Cost of Certainty: Reliability isn't free. The authors prove (Proposition 3.4) that a reliable recommendation will almost always require a policyholder to change more features or make larger changes than a standard "valid" recommendation. To get a result the model is sure about, you have to move further into the "known" territory of the data.
  • Outliers are at Risk: High-risk policyholders are the most likely to receive unreliable advice because they exist in sparse regions of the feature space. Without the CP constraint, a model might give them "recourse" that is essentially a statistical hallucination.
  • Diagnostic Power: The method can reveal model flaws. For example, in the flood dataset, the model sometimes attributed less risk reduction to "elevating a building" than to "moving contents," which contradicts engineering knowledge. The counterfactual search surfaces these inconsistencies, acting as a "stress test" for the insurer's model.

Key Results

  • Reliability is Rare: In the telematics dataset, only 0.2% to 3.8% of standard "valid" recommendations actually met the 80% reliability threshold. This highlights how untrustworthy standard XAI can be for individual recourse.
  • Bigger (but Better) Changes: While reliable recommendations required changing more features (an average of 2.69 features vs. 1.40 for unconstrained ones), they were associated with much larger predicted risk reductions (61.4% vs. 25.7%).
  • Method Performance: WhatIf was the fastest method, making it suitable for large-scale deployment, while NICE provided the best balance of validity and reliability. MOC (a genetic algorithm) found the most plausible changes but struggled significantly with reliability.

Limitations and Future Work

  • The "Selection" Gap: The statistical guarantee of Conformal Prediction applies to the profile (the set of characteristics), but because the counterfactual is chosen via optimization, there is a slight statistical "leakage" that means the guarantee isn't 100% perfect for the specific individual.
  • Causality vs. Correlation: The paper assumes that the model's learned associations are causal. If a model says "driving at night is risky," it assumes that stopping night driving will lower risk. In reality, night driving might just be a proxy for a different, unobserved risk factor.
  • Differential Actionability: The authors note that a recommendation to "elevate your house" is actionable for a wealthy homeowner but may be impossible for a low-income policyholder. Future work is needed to incorporate the "cost" of changes into the search.
  • Human-Centric Design: How do you explain a "80% confidence upper bound" to a non-technical consumer without causing confusion or "algorithm aversion"? This remains an open design challenge.

Summary updated

Abstract

Machine learning (ML) systems increasingly make consequential decisions about individuals, including credit, hiring, and insurance. Research on ML-driven insurance pricing has asked how the pricing mechanism itself can be made accountable. Yet it remains unclear how an individual policyholder's algorithmically determined premium should be explained to them. We argue that a useful explanation must be valid, actionable, and reliable. A valid recommendation lowers the premium according to the model. An actionable recommendation changes only characteristics the policyholder can control. A reliable recommendation has an actual outcome that is likely to meet the target, even in sparsely observed regions of the feature space where the model is uncertain. We formalize these three requirements as axioms and develop a method that turns an algorithmic premium into recommendations that meet them. The method meets the third requirement with a Conformal Prediction constraint, which screens out recommendations that the model is too uncertain about and thereby bounds the rate of unreliable recommendations. We apply four Counterfactual Explanation techniques to 57,832 real Florida flood-insurance claims and 98,058 synthetic driver-telematics policies. We adapt these techniques for insurance applications, so that their recommendations change only controllable characteristics and resemble policyholders observed in the data. In our illustrations, the constrained recommendations for high-risk policyholders are both more reliable and associated with larger predicted risk reductions than the unconstrained ones. These results suggest that insurers can give policyholders explanations that are both actionable and trustworthy. The work contributes to research on algorithmic accountability and offers implications for insurers and regulators.
BibTeX
@misc{machado2026from,
  author = {Agathe Fernandes Machado and Fei Huang},
  title = {{From Explaining to Enabling: Actionable Transparency in Algorithmic Insurance Pricing}},
  year = {2026},
  eprint = {2610.34459},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.34459},
}
Figure