acceptodds
Under review as a conference paper at ICLR 2027

-Fair Insurance Pricing: A Fairness Continuum

Abstract

Fairness in insurance pricing remains a long-standing and deeply debated puzzle. On one hand, insurers, driven by profitability considerations, set premiums that differentiate across individual risks to achieve actuarial fairness. On the other hand, insurance serves a critical societal function by pooling risks across a population, motivating cross-subsidization among groups to promote solidarity fairness. The tension between these two competing notions of fairness makes insurance pricing inherently complex, particularly in modern settings where granular data allow for increasingly fine risk differentiation and regulators face growing pressure to protect vulnerable groups. To address this challenge, we propose an -Fair Individual Solvent Premium (-FISP) framework for insurance pricing that explicitly captures the trade-off between actuarial and solidarity fairness while guaranteeing solvency, a fundamental requirement in insurance operations. We formulate the pricing problem as a constrained optimization task, where actuarially fair premiums are adjusted subject to budget constraints on cross-subsidization within each risk class. This formulation naturally yields a family of solutions parameterized by , tracing a continuum between purely actuarial and purely solidarity-based pricing and enabling decision-makers to select an operating point along this fairness spectrum. We derive various theoretical results for the proposed framework. Numerical experiments show that -FISP is computationally tractable and aligns well with the U.S. regulatory regimes featuring heterogeneous state-level fairness requirements.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
insurance pricing, fairness

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_xEPROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes α\alpha-Fair Individual Solvent Premium (α\alpha-FISP), a framework for insurance pricing that balances two competing notions of fairness:

  • Actuarial fairness: Premiums reflect individuals' expected insurance losses.
  • Solidarity fairness: Individuals with the same non-sensitive characteristics pay identical premiums, regardless of their sensitive attributes.

The authors introduce a fairness parameter α∈[0,1]\alpha\in[0,1] to control the trade-off. They formulate premium determination as a constrained optimization problem that minimizes prediction error while enforcing fairness and solvency requirements. The paper establishes existence and uniqueness of the optimal premiums, generalization guarantees and boundedness.

Critical strengths

This paper provides an interpretable fairness trade-off and a mathematically tractable formulation.

Critical weaknesses

There is gap between the theory and numerical results. The finite-sample theory assumes that the empirical estimator belongs to a globally α\alpha-fair hypothesis class HαH_\alpha. However, the numerical algorithm uses a neural-network implementation based on Lagrangian optimization, which enforces constraints approximately. Furthermore, the theoretical assumptions can be difficult to establish for practical neural-network models, which limit the scope of contributions.

Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer xEPR

We thank the reviewer for the constructive feedback. We carefully address the concerns and questions raised as follows:

Thank you for highlighting this distinction. The finite-sample results characterize a specified constrained estimator, rather than establish convergence or global feasibility of an arbitrary neural-network optimizer. Theorem 4.2 assumes empirical risk minimization within the globally fair class HαH_\alpha, with calibrated relaxation of the empirical solvency constraints. The Lagrangian implementation approximates this problem, and its reported test-time violation rates assess practical compliance rather than certify the exact guarantees of Corollary 4.4 . We will state this distinction explicitly alongside both the theoretical results and the implementation description.

Nevertheless, the assumptions are not inherently incompatible with neural-network models. Remark C.1 identifies bounded ReLU-network classes satisfying the required hypothesis-class covering-number condition when architecture, parameter magnitudes, and the input domain are controlled. Thus, this component of the analysis applies to concrete neural-network classes, not only abstract function spaces. However, accurate plug-in estimation and inclusion of the population optimizer in the hypothesis class are additional substantive assumptions; we do not claim that using a neural network automatically establishes them.

The contribution should therefore be understood at three distinct levels: the population results characterize the constrained pricing problem independently of a neural-network parameterization; the finite-sample analysis identifies sufficient conditions for statistical guarantees; and the experiments investigate an approximate numerical implementation. In particular, existence, boundedness, and risk-order preservation do not depend on the stronger finite-sample assumptions. We will make these scopes explicit and describe the experiments as illustrating the framework’s behavior, rather than verifying that every assumption of the finite-sample theory holds for the trained networks.

We have added a remark following Corollary 4.4 (in red) showing that global α\alpha-fairness can be enforced by a simple neural-network output transformation. This demonstrates that the globally fair hypothesis class is constructively implementable, rather than requiring exhaustive verification over the input space. We explicitly distinguish this alternative parameterization from the approximate Lagrangian implementation used in the experiments and clarify that it does not, by itself, establish the remaining assumptions of the finite-sample results.

I appreciate the authors’ response, which partially addresses my concerns. However, I believe my original score remains appropriate and have therefore decided to maintain it.

Reviewer_B9rBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a tension in insurance pricing: charging people according to their expected losses versus charging the same premium across sensitive groups with the same non-sensitive characteristics. The authors control this trade-off through α\alpha, formulate a constrained optimization problem and evaluate it on simulated data and a health insurance dataset.

Critical strengths
  • The premium-ratio constraint in Definition 3.6 gives α\alpha a direct interpretation: it specifies how much prices may differ within a risk class.
  • Propositions 4.5 and 4.7(iii) establish useful properties of the population solution: premiums stay within the range of expected losses and preserve the ordering of risks.
Critical weaknesses
  • I’m not sure what justifies calling the conditional expectations in Definitions 3.1–3.2 “fair” premiums. My concern goes beyond terminology. I’m not convinced the paper makes a clear contribution to the broader ML community, and in its current form it seems better suited to an insurance or actuarial science venue than to ICLR.
  • Section 5.3 uses MSE–α\alpha curves to guide α\alpha selection, but this leaves open how to decide which differences in expected loss should justify different premiums.
  • Definition 3.4 compares only identical xx values; I could not tell from Section 5.1 how these constraints are enforced when inputs are continuous or high-dimensional.
Questions for authors
  1. I’m new to insurance pricing: would charging male smokers more than comparable female smokers be unfair under this framework if their expected loss is higher? How does this differ from charging more simply because someone is male?

  2. If I understand Definition 3.4 correctly, groups are compared only at the same xx. How does this apply to continuous variables like income, where exact matches may not exist?

1 comment
AuthorsOpenReview ↗

Response to Reviewer B9r8

We thank the reviewer for the constructive feedback. We carefully address the concerns and questions raised as follows:

“Actuarial fairness” and “solidarity fairness” represent different principles of allocating expected losses, not a claim that conditional expectations are intrinsically fair. The former matches contributions to group-specific expected costs; the latter pools sensitive-group differences and requires equal contributions within a risk class. Definitions 3.1–3.2 formalize these competing principles without declaring either universally preferable. Our contribution is not these benchmarks, but a constrained learning framework connecting risk differentiation, cross-subsidization, and conditional expected-loss coverage. We establish well-posedness, boundedness and risk-order preservation of the population solution, and finite-sample guarantees separating statistical error from fairness-induced prediction cost. The relevance to ML lies in studying how fairness restrictions interact with statistical learning and operational feasibility when predictions determine financial allocations—not merely applying an existing predictor to insurance data.

The MSE–α\alpha curves inform the consequences of a fairness choice; they do not determine its ethical justification. Which attributes define risk classes and how much sensitive-group differentiation is acceptable require context-specific judgments. Conditional on those choices, the curves quantify the predictive cost of stronger pooling and support comparison among acceptable alternatives. Higher predictive accuracy does not itself justify differential premiums. We revised Section 5.3 (in red) to clarify that the curves provide decision support rather than identify an ethically optimal α\alpha, consistent with the scope stated in our Ethics Statement.

Definition 3.4 imposes a restriction on the pricing function, rather than requiring matched observations in the dataset. For example, at a given income and fixed values of the other non-sensitive attributes, the fitted function can be evaluated with each sensitive-group label. The resulting premiums are then compared through the fairness inequalities, even if the dataset contains no two policyholders with exactly that feature vector. These comparisons require model evaluations, not additional observed policyholders or loss outcomes at the same input. Continuous covariates therefore do not make the requirement vacuous. However, its scope is deliberately conditional: it controls differences across sensitive groups at the same input, not differences between nearby but unequal incomes. Where broader comparisons are desired, the framework can instead be applied to predefined coarser classes, such as income bands, with fairness and solvency interpreted at that chosen resolution. We expanded the implementation description in Section 5.1 (in red) to make the constraint evaluations explicit.

Our framework allows controlled risk-related premium differences, but does not reverse the underlying risk ordering to achieve solidarity. Specifically, Proposition 4.7(iii) establishes that, under the stated assumptions, the population-optimal premiums satisfy μmale(x)≥μfemale(x)  ⟹  f∗(x,male)≥f∗(x,female)\mu_{\text{male}}(x) \ge \mu_{\text{female}}(x) \implies f^* (x, \text{male}) \ge f^* (x, \text{female}) Therefore, when male smokers have higher expected losses than otherwise comparable female smokers, the fairness adjustment may narrow or eliminate the premium difference, but cannot reverse it and charge the lower-risk group more. This addresses the concern that promoting solidarity could introduce another inequity by reversing risk-based premium rankings. Moreover, the same result implies that groups with equal conditional expected losses receive equal optimal premiums. An additional surcharge merely for belonging to a particular group, despite equal expected losses, is therefore not supported by the optimization. The role of α\alpha is to regulate the permissible extent of differentiation: a higher expected loss does not exempt a premium difference from the proportional fairness requirement, and at α=1\alpha=1, both groups pay the same pooled expected loss. Thus, the framework combines controlled risk sharing with preservation of risk ordering, rather than treating all premium differences as unfair or allowing arbitrary group-based surcharges.

Reviewer_zhkrOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper "alpha-Fair Insurance Pricing: A Fairness Continuum" proposes an insurance pricing model that seeks a compromise between actuarial fairness (setting a price that is the closest to the real cost of the individual) and solidarity fairness (allocating the same treatement for any given individual, independently from its sensitive group). the α\alpha hyper-parameter allows to set the preference between these two objectives.

Critical strengths
  • Interesting problem
  • The theoretical analysis looks well sounded
Critical weaknesses
  • The paper lacks pedagogy. It lacks background to well understand the objectives and the setting of the presented work. From my point of view, the background section should have described the differences between different objectives on the targetted group fairness. For instances, fairness through unawareness and the DFIP method that is considered as a baseline for experiments (this method looks to target a different objective and from my point of view is maybe not fully comparable with the proposed approach, see below). Next, in many places notations are not defined (e.g., P_L and C_L) or the considered constraints are clearly not sufficiently visible (e.g., the constraint P_L\leq 1/(1+\alpha) considered in figure 2 is only given in a large proposition from section 4, very difficult to be interpreted in this section 5). More generally, not enough interpretations, justifications are given throughout all the paper.

  • I feel that the considered baseline DFIP is not fully comparable to the proposed method FISP, as they target to different solidarity fairness definitions : FISP targets sensitive unawareness: when α→1\alpha \rightarrow 1, we have f(x)=f(x,d)f(x)=f(x,d) for every (x,d)(x,d), thus f(x)=E[Y∣X=x]=∑dμ(x,d) P(d∣x)f(x)=E[Y\mid X=x]=\sum_d \mu(x,d)\,P(d\mid x) (since the mse is minimized under the joint proba) which suffers from proxy discrimination through P(D∣X)P(D|X). DFIP aims to remove that bias, replacing P(D∣X)P(D|X) by P(D)P(D) by targeting f(x)=E[Y∣do(X=x)]=∑dμ(x,d) P(d)f(x)=E[Y\mid do(X=x)]=\sum_d \mu(x,d)\,P(d). Thus, the comparison is not fully fair, as both methods do not target the same thing. The experiments should measure different fairness metrics to be useful. Also, I am not fully convinced by this definition of solidarity fairness: if features X are strongly correlated with D, D strongly impacts the prediction even with α→1\alpha \rightarrow 1. This looks unfair to me.

  • Most approaches consider that D is not available at test time (following legislation). DFIP for instance uses it only for training. Could this usage of D for prediction be justified in some way ?

  • Considering fairness disparities only for similar x looks limitative for real worl x vectors with many variables (that can be continous by the way). Also, in real settings, how do authors estimate p_{d|x} which is needed for the solvability constraint ? for high-dimentional d and or x, this looks untractable.

Questions for authors

See weaknesses above for the main questions. Other questions / remarks :

  • It would have been useful to given a pareto front considering mse and and solidarity fairness to analyze the obtained tradeoff.

  • From my point of view it is somewhat misleading to talk about actuarial fairness, while the criterion only sets a lower bound on the price, not any upper one. Thus it is not any matter of fairness, but only a profit preservation guarantee.

  • In the experiments, it would be useful to know the number of similar samples for x. Or any statistics that inform about the possible identification of p_d|x. The number of different d for any x or something like that...

3 comments
AuthorsOpenReview ↗

Response to Reviewer zhkr

We thank the reviewer for the constructive feedback. We carefully address the concerns and questions raised as follows:

We have modified the manuscript and highlighted the modifications in red to address the concerns regarding pedagogy (Section 2.3) and notation (Section 5.1). Please see the revised paper for details.

We agree that predictive accuracy and solvency alone do not establish that one method is fairer than another. However, methods targeting different fairness objectives can still be meaningfully compared in terms of their prediction error and conditional expected-loss coverage. This is the purpose of our comparison with DFIP, rather than establishing superiority under a shared fairness definition. Within our framework, α explicitly specifies the permitted within-class premium disparity, and the reported fairness-violation rates assess adherence to that requirement. DFIP assigns identical premiums across sensitive groups within each risk class by construction; we do not claim to improve upon its protection against proxy discrimination. Marginal differences in average premiums would assess a separate parity objective, rather than directly evaluate our conditional risk-sharing criterion. We will therefore clarify the scope of the comparison and revise its interpretation: the experiments illustrate the predictive and financial consequences of different fairness objectives, not a ranking of those objectives.

Our framework separates the estimation of group-specific risks from the subsequent fairness-constrained premium adjustment. Consequently, it is not tied to a particular estimation method or to direct observation of sensitive attributes. Any method that estimates the required group-specific expected losses and conditional group probabilities—whether sensitive attributes are noisy, privatized, or unobserved—can supply inputs to the pricing optimization, provided these quantities are identifiable from the available information. For example, existing work develops methods for estimating group-specific expected losses from privatized sensitive attributes. These estimates can be incorporated into a plug-in implementation without changing the underlying fairness and solvency formulation.

Our fairness requirement compares group-specific premiums while holding non-sensitive attributes fixed; it does not require finding observed policyholders with identical feature vectors. A fitted pricing model can evaluate premiums across sensitive groups at a given input, so continuous covariates do not make the comparison undefined or vacuous. This reflects our objective of controlling sensitive-group differentiation within an adopted risk classification. Similarly, estimating pd∣xp_{d\mid x} does not require calculating empirical frequencies separately for every distinct xx. These probabilities can be learned through a probabilistic classifier trained across observations, such as a neural network with softmax outputs and cross-entropy loss. The framework can also leverage established methods for high-dimensional estimation. Moreover, although our implementation uses the finest risk classification—treating each distinct feature vector as a risk class—this granularity is not essential. Practitioners can instead define coarser, actuarially meaningful classes. Both group probabilities and expected losses would then be estimated conditional on the chosen classes. This provides an additional way to balance risk differentiation against statistical stability through greater pooling. Fairness and solvency would consequently apply at that class resolution, rather than separately to every original feature vector.

We would like to clarify that actuarial fairness is not defined by the solvency inequality. Definition 3.1 defines the actuarially fair premium as the conditional expected loss. Individual solvency is a separate requirement ensuring that each risk class’s average premium covers its expected loss; it does not require every policyholder’s premium to exceed their own expected loss. Importantly, the framework does not maximize profits or merely select any premium satisfying this lower bound. Its squared-loss objective is equivalent to minimizing squared deviations from actuarially fair premiums, penalizing both overpricing and underpricing relative to those benchmarks. Moreover, Proposition 4.5 establishes a two-sided bound on the population-optimal premiums. Thus, the optimal premium cannot exceed the highest actuarially fair premium within its risk class. The absence of an explicit upper-bound constraint therefore does not imply unrestricted pricing at the optimum. We do not claim that every adjusted premium remains actuarially fair: departures from individual expected losses are precisely what allow cross-subsidization toward solidarity. A positive expected surplus is permitted, but it is not the quantity being maximized, nor does the solvency condition guarantee realized profitability.

I thank the authors for their response. The clarification on actuarial fairness shows that premiums are bounded at the optimum (Prop. 4.5). It also confirms that the actuarial side of the continuum is not enforced as a fairness requirement: it is only the regression target, and the sole hard constraint is class-level solvency, i.e. a financial guarantee rather than a fairness one. My main concerns remain.

Comparison with DFIP: when α→1\alpha \to 1, both methods yield premiums independent of dd and differ only in the weighting of groups: ∑dμ(x,d)P(d∣x)\sum_d \mu(x,d)P(d\mid x) for FISP vs. ∑dμ(x,d)P(d)\sum_d \mu(x,d)P(d) for DFIP. The first is MSE-optimal among functions of xx and exactly covers class-level expected losses, so FISP wins on both reported metrics by construction. These metrics only measure the cost of DFIP’s correction of proxy discrimination, not what it gains. A meaningful comparison would require a proxy-discrimination metric (e.g., the deviation from ∑dμ(x,d)P(d)\sum_d \mu(x,d)P(d)) and a Pareto front over α\alpha.

Proxy discrimination: the authors acknowledge that FISP does not address it, but this limitation of their solidarity definition should be explicitly discussed in the paper.

Use of DD at pricing time: the response addresses the estimation of μ\mu and P(d∣x)P(d\mid x), but for α<1\alpha<1 the premium f(x,d)f(x,d) still requires the policyholder’s sensitive attribute at pricing time, and its legal basis (e.g., the EU ban on gender-based insurance pricing) is not discussed.

Since the experimental comparison is uninformative by construction and the scope of the fairness notion (proxy discrimination, use of DD at pricing time) remains unaddressed, I keep my score.

AuthorsOpenReview ↗

Response to Reviewer zhkr

We thank the reviewer for the continued discussion. We agree that the distinction between evaluating a pricing framework and establishing superiority across fairness objectives is important. Our revised manuscript incorporates this distinction and now explicitly addresses the pricing-time deployment limitation. We clarify the relevant passages below.

Role of actuarial fairness: We agree that class-level solvency is a financial feasibility condition and does not, by itself, establish actuarial fairness. Our actuarial allocation principle is instead represented through the objective: premiums should remain as close as possible, in expected squared deviation, to their group-specific expected losses, subject to the prescribed restrictions on premium differentiation. Thus, being a regression target and representing an allocation principle are not mutually exclusive. The proportional constraints regulate the departure from actuarial differentiation toward within-class solidarity, while solvency ensures conditional expected-loss coverage. The “continuum” refers to the resulting optimal pricing rules as α\alpha varies, not to simultaneous enforcement of both endpoint principles as hard constraints.

Purpose of the DFIP comparison: We agree that, at the population endpoint α=1\alpha=1, MSE optimality and exact conditional expected-loss coverage follow analytically. We do not present these properties as an empirical discovery or as evidence that our fairness objective is superior. The experiments quantify the magnitudes of prediction errors and solvency margins in the fitted models and examine practical constraint satisfaction across fairness budgets. An analytically expected ordering does not make those magnitudes or implementation outcomes predetermined. Importantly, this interpretation is already explicit in revised Section 5.3, lines 464-470, which describes DFIP as a reference for an alternative fairness objective, and lines 497-500, which explain that the solvency results characterize financial implications rather than establish fairness superiority. We acknowledge that the reported metrics do not quantify DFIP's gains in proxy-discrimination protection. A proxy-discrimination metric and corresponding Pareto analysis would support a broader, joint evaluation of the two objectives; our experiments make the narrower claim of illustrating their predictive and financial consequences. In particular, the MSE- α\alpha curves are not presented as a Pareto frontier for proxy discrimination.

Scope of solidarity and proxy discrimination: This limitation is explicitly discussed in the revised paper. Section 2.2, lines 124-134, explains that unawareness may retain proxy effects and distinguishes its conditional weighting from DFIP's common reference composition. Lines 134-140 then clarify our objective, specifically stating in lines 138-139 that it regulates premium differences within risk classes "rather than to remove proxy effects from risk classification." The Ethics Statement, lines 559-560, further notes that designating variables as non-sensitive does not prevent them from acting as proxies. Thus, solidarity here is conditional on the adopted risk classification; we do not claim that increasing α\alpha eliminates proxy discrimination or provides universal nondiscrimination.

Use of DD at pricing time: We agree that our previous discussion of estimation did not fully answer this question. We have now added an explicit deployment-scope paragraph at the end of Section 2.2, lines 141-147, highlighted in blue. Lines 141-145 state that deploying group-dependent premiums requires the sensitive attribute to be available and its use in pricing to be permissible, and that estimating group-specific quantities from privatized or indirectly observed attributes does not remove this prediction-time requirement. Where such differentiation is prohibited, the group-dependent solutions are not directly deployable. Lines 145-147 clarify that the α=1\alpha=1 endpoint does not require the policyholder's sensitive attribute at prediction, while explicitly acknowledging that proxy effects may remain. This complements the Ethics Statement, lines 564-566, which describes the experiments as methodological illustrations rather than authorization for operational pricing and conditions deployment on applicable restrictions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.