Abstract Machine‐learning models can provide accurate predictions in insurance pricing, but can also increase disparities between protected groups. Existing fairness‐aware pricing approaches typically target one fairness notion at a time, making it difficult to compare trade‐offs between predictive accuracy, group fairness, individual fairness, and counterfactual fairness. We propose a multi‐objective framework for fairness‐aware insurance pricing. The framework combines several fairness‐aware base models and uses the Non‐dominated Sorting Genetic Algorithm II (NSGA‐II) to approximate the Pareto front over four objectives. We then use the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) to select a compromise solution from the Pareto front. Using two motor insurance datasets, we find that Extreme Gradient Boosting (XGBoost) improves predictive accuracy relative to the generalized linear model (GLM), but it can worsen some fairness metrics. The proposed ensemble provides a balanced compromise across the considered objectives and offers a favorable aggregate accuracy–fairness compromise.
Abstract High‐emission private vehicles disproportionately contribute to urban air pollution and road accidents, both major public health threats and leading causes of death among young adults. Balancing individual driving behavior with collective environmental responsibility necessitates strategies that promote safer and cleaner driving. This paper presents a novel method for assessing crash risk using air pollutant emissions as exposure measures, encouraging reduced environmental impact and crash likelihood. Using over 1500 at‐fault crash claims from an insurance dataset merged with vehicle telematics data, we demonstrate that emission factor models can integrate pollutant‐based exposures into crash risk estimation. Results show that pollutant‐driven models perform comparably to traditional distance‐driven models when combined with behavioral telematics data. This approach offers an integrated framework for enhancing road safety and mitigating emissions by linking environmentally responsible driving practices with lower crash risk.
Abstract Government insurance premiums can diverge from derivative prices, creating pricing wedges relevant for subsidy design and market segmentation. Using Livestock Risk Protection (LRP) endorsements matched to Chicago Mercantile Exchange (CME) put options from 2005 to 2024, we examine how federal subsidy expansions changed public livestock insurance prices relative to market‐based protection. The net wedge equals the matched put premium minus the producer‐paid LRP premium and implementation costs. Average wedges are near zero overall but shift from negative before the 2019–2020 subsidy expansion to positive after the 2020 tiered subsidy schedule. Gross LRP premiums generally remain above matched put premiums, and a pre‐expansion subsidy‐rate counterfactual reverses the post‐expansion wedges, showing that statutory subsidies drive the change. Out‐of‐sample predictions based solely on lagged information identify matched contracts with more favorable realized wedges.
Although temperature dynamics exhibit pronounced long-memory behavior, most existing temperature models and weather derivative valuation frameworks neglect such persistence, leading to biased forecasts and systematic mispricing. We propose the generalized fractional Ornstein-Uhlenbeck (gfOU) process that parsimoniously incorporates time-varying trends and seasonality while capturing both short- and long-range dependence. Under the stationary fOU process, we derive a tractable closed-form autocovariance function, quantify the implications of misspecification, and obtain weather contract prices under the risk-neutral valuation framework. Incorporating long memory yields economically material improvements in forecast accuracy, insurers' profitability, and reserve adequacy. Empirical results reveal substantial spatial and temporal heterogeneity in temperature persistence across the continental United States. The predictive gains of the gfOU model primarily reflect its ability to exploit long-memory dynamics embedded in historical temperature realizations.
In insurance markets, claim costs are highly variable, heavy-tailed, and difficult to predict. At the same time, policyholder retention and lapse behavior (customer churn) are critical determinants of long-term profitability and solvency. Most existing models in the literature treat claim costs and lapses as independent, overlooking potential latent associations that arise from adverse selection and unobserved heterogeneity. In this article, we introduce a joint modeling framework that simultaneously captures individual-level claim costs and churn behavior, using multivariate Tweedie regression with shared random effects. This framework integrates claim cost dynamics with lapse risk, allowing insurers to more accurately predict costs, classify policyholder profitability, and design retention or pricing strategies. Applying our approach to data from the Wisconsin Local Government Property Insurance Fund, we demonstrate that accounting for dependence between claim risk and lapse risk improves out-of-sample prediction and yields actionable insights for customer valuation and management.
This paper provides a model for allocating capital to different insurance lines with varying development periods for a value-maximizing insurance company. In our model, the company makes capitalization and exposure decisions considering its capital level and its relevant loss history. As in simpler settings, the optimal portfolio can be characterized via risk-adjusted return ratios, although the model attaches different valuation weights to cash flows with different tenors. Numerical results show that our approach yields substantively different guidance relative to that obtained from conventional capital allocation approaches, particularly on the relative valuation of long-tailed versus short-tailed liabilities, since long-tailed lines can provide a source of short-term financing. We discuss robustness and implications of our findings.
In many low- and middle-income countries, social insurance provides basic pension benefits with limited cover for illness and care costs, while private insurance markets are underdeveloped. Using an online survey of retirement portfolio choices in urban China, we examined the stated demand for longevity, critical illness, and long-term care (LTC) insurance under realistic financial constraints. The most preferred portfolios had 50% cover for expected out-of-pocket critical illness and LTC costs, and a monthly annuity of around 20% of average urban disposable income. Access to critical illness and LTC insurance increased annuity demand in some cases, with effects varying by wealth and level of cover. Product knowledge, higher financial competence, stronger bequest motives, and lower risk tolerance were linked to higher demand for critical illness and LTC cover but lower demand for annuities. Results inform the development of retirement insurance markets in countries with ageing populations and social and private insurance gaps.
The impact of reputation of insurers on its market performance has long been a significant issue. Using data from China's life insurance industry from 2011 to 2021, this study combines ChatGPT with active learning to construct the perceived reputations of Chinese insurers. We find that reputation is negatively associated with surrender volume. Further results show that the negative reputational effect on surrender is more salient in positive sentiment, while the reputational effect with negative sentiment is mainly realized through the negative events of lawsuits. A positive reputation matters mainly in financial soundness, customer service, and corporate social responsibility, while managing negative reputation risks is particularly important for corporate governance and government relations. Finally, the paper finds that policyholders prioritize insurers' reputation when inflation, unemployment, and interest rates are low or when the term spread is high. Policyholders are also more sensitive to larger insurers' reputation in a more dynamic market.
Econometric studies of insurance markets have analyzed the Positive Correlation Property to test for the presence of asymmetric information. Car-insurance studies frequently compare policies purchasing Mandatory Third-Party Liability alone with policies that purchase additional coverage and use the presence of a liability claim as a measure of risk. Using data from the Italian market, we show that this approach can yield conflicting results from the same dataset. First, different types of additional coverage can yield different results; second, the presence of a claim is a poor or even misleading indicator of the cost to the insurance company. A test using an unused observable is similarly problematic. Our findings highlight that selection may operate along dimensions of risk not captured by claim probabilities, complicating the interpretation of standard empirical tests and the distinction between adverse and advantageous selection.
Using the universe of transaction-level data in the U.S. corporate bond market around uninformative downgrades, we find an abnormal increase in trading volume, abnormal bond returns, and a subsequent reversal. On the contrary, we do not find a reversal for abnormal bond returns associated with informative rating actions. We then focus on the largest, domestic, institutional investor from our sample, and match individual investment transactions to firm characteristics. We document an association between restatements and abnormal trading on uninformative news. These results provide supportive evidence that some institutional investors perceive uninformative downgrades as informative, hence leading to short-lived mispricing episodes.
Health insurance mitigates health risks by covering the cost of treatment, but it also reduces the financial risk associated with those costs. In a multivariate preference framework, this study analyzes the effects of changes in wealth, health status, insurance premiums, treatment costs, and the severity of health loss on the demand for treatment and health insurance in the face of health loss. The conditions under which demand for treatment and insurance increases depend on the attitude toward the correlation between wealth and health, namely, correlation loving or correlation aversion, as well as the attitude toward wealth and health risks. These findings have implications for government subsidization policies related to treatment and health insurance and health promotion policies.
This paper explores whether insurance against aggregate risk is effective in economies with heterogeneous agents. While aggregate shocks are typically viewed as uninsurable, we show that insurance becomes effective when agents differ in productivity or initial wealth. We develop a two-period general equilibrium model with a CES production technology and Greenwood-Hercowitz-Huffman preferences. Insurance is effective only when heterogeneity generates asymmetric income responses, in which case state-contingent claims restore Pareto efficiency. Quantitative results show that welfare gains arise only with heterogeneity and are larger for agents further from the average. Moreover, welfare gains increase with dispersion, while the effect of average exposure is non-monotonic due to general-equilibrium price adjustments.
We examine risk attention as an overlooked driver of insurance uptake. We leverage the COVID-19 pandemic as a natural experiment-when all COVID-related expenses were covered by public funds in China and public attention shifted to the only major under-covered source of catastrophic medical spending: critical illnesses (CI). Using a unique private CI insurance dataset with both unbound insurance applications and bound contracts, we document a rise in bound insurance contracts and a disproportionately larger increase in unbound applications, particularly among lower-risk individuals. This surge reflects the increased interest and awareness of CI risks and insurance, as well as realized CI insurance uptake. We find robust evidence that the surge is driven by heightened attention to previously overlooked CI risks, and rule out competing explanations.
This paper introduces an empirical framework to evaluate the welfare implications of fair and accountable insurance pricing by modeling the complete pricing process, including demand and price optimization. Moving beyond traditional cost modeling, we analyze both discrimination-related fairness criteria and broader regulatory constraints, such as price optimization bans, that constrain insurers' pricing behavior. Using two French auto insurance datasets with gender as the protected attribute, we provide the first systematic empirical evaluation of the welfare effects of these pricing constraints from the perspectives of both consumers and firms. Our findings reveal: (1) a fundamental tension between fairness in prices and markups; (2) price equalization policies eliminate price gaps but amplify markup disparities; (3) accountability constraints reduce markup gaps while imposing profit losses; and (4) price optimization bans exhibit market-contingent welfare effects that depend critically on underlying market characteristics.
Using data from over 56,000 simulated auto races worldwide, we analyze risk-taking at the margins, consistent with reference-dependent preferences. We show that participants' risk-taking changes when a desired intermittent outcome is presented, sometimes at the expense of a more favorable expected end state. Specifically, we find that intermediate kinks in the utility function induce players to take less (more) risk given opportunities to increase (lose) temporal status, providing important intuition regarding the incentives for risk-taking at the margin of wealth kinks (e.g., retirement age, family changes, etc.). Risk aversion strengthens at kinks, but risk-taking increases with more individual investment.
Insurance fraud detection remains a challenging task due to severe data imbalance, evolving fraudulent behaviors, and the high false-negative rates exhibited by several state-of-the-art machine learning models. Traditional approaches often struggle to generalize real-world data and capture complex, non-linear feature interactions in insurance claims. This study aims to improve fraud detection performance by leveraging recent advances in deep learning. A comprehensive comparison between traditional machine learning models and deep learning techniques is performed on two distinct datasets using resampling strategies. The study proposes three convolutional neural network-based architectures to improve detection accuracy. Furthermore, a hybrid machine learning deep learning (ML-DL) framework is introduced to more effectively leverage discriminative features. Experimental results demonstrate that deep learning models would vary on each dataset due to the presence of variations in data characteristics, while the proposed hybrid ML-DL model achieves the best overall performance, highlighting its effectiveness in improving fraud prediction accuracy.
Registered Index-Linked Annuities (RILAs) have quickly become one of the most popular retirement savings and investment vehicles in the United States. Researchers have analyzed their ability to help investors accumulate wealth—and have praised them for their relatively low cost and transparency—but have not yet considered whether RILAs can be a suitable component of retirement planning during the decumulation phase as well. This study aims to fill that gap by embedding RILAs in a lifecycle utility framework that models an investor's optimal decision-making post-retirement. In that context, I find that RILAs are essentially a like-for-like substitute for traditional mutual funds, in terms of both the total utility provided to the retiree and his optimal consumption and annuitization decisions.
This study examines how InsurTech-enabled information provision, specifically the disclosure of claimant information previously unavailable in conventional insurance, influences individuals' insurance uptake. We leverage Mutual Aid (MA) platforms as a natural context to examine how socially framed loss information, peer influence, and salience shape insurance decisions, offering broader insights into the role of InsurTech within traditional insurance markets. Using unique user-level data from a leading MA platform in China and a dynamic discrete choice framework, we show that engagement with MA activities significantly increases the likelihood of subsequent private insurance uptake. Our counterfactual simulations highlight the critical role of strategically structured information in promoting insurance adoption. Overall, the study demonstrates how a digitally mediated and publicly observable risk-sharing environment can shape consumer behavior and provides new perspectives on the evolving intersection between InsurTech and traditional insurance markets.
In high-risk environments, traditional indemnity insurance is often unaffordable or ineffective, despite its well-known optimality under expected utility. This paper compares excess-of-loss indemnity insurance with parametric insurance within a common mean-variance framework, allowing for fixed costs, heterogeneous premium loadings, and binding budget constraints. We show that, once these realistic frictions are introduced, parametric insurance can yield higher welfare for risk-averse individuals, even under the same utility objective. The welfare advantage arises precisely when indemnity insurance becomes impractical, and disappears once both contracts are unconstrained. Our results help reconcile classical insurance theory with the growing use of parametric risk transfer in high-risk settings.
As the origin of modern commercial insurance, ship insurance underpins global maritime supply chain stability. Yet shipping modernization and AI advances expose three flaws in traditional risk profiling: misalignment with frequency‐severity pricing, inadequate for accommodating to complex risk factor system, and lack of data stream adjustment mechanisms. To address these, we propose POM principles ( Personalized risk portrait , Omnispective risk factors , and Maneuverable calibration ) and an AI framework with three cores: (1) extensible “retrospective + prospective” risk factors; (2) independent AI modules for premium rate/insurance amount prediction; (3) data steam calibration on historical data. Validated via 15,007 records (15% of China's 2016–2021 registered ships) using random forest regression, it outperforms traditional generalized linear models and mainstream machine learning models in accuracy and risk differentiation. This pioneers intelligent ship insurance profiling, fills gaps in individualized pricing, and offers insights for sectors like aviation insurance, sharing its “premium rate × insurance amount” logic.
