betaFeedback

WorksComputer Science

Yielding Fairness: Coupled Learning and Fairness for Value-Weighted Ranking

Abstract

Risk screening and recommendation systems pass or block accounts that generate both value and loss. Under a blocked-value budget, a cap on expected value forgone through blocking, the optimal ranking orders accounts by expected value per unit expected loss, a ratio of conditional expectations. This rule can conflict with accepted harm or net-harm criteria, blocking accounts that another criterion would pass at the same budget. We call these cases account-level objections and define group complaint disparity as the difference between groups' normalized realized objections. Weighted binary cross-entropy (W-BCE) implements the value-per-loss order with one score. Objection penalties modify the ranking, while group multipliers scale those penalties differently by group, tracing tradeoffs between loss capture and complaint disparity. This preserves the value-budgeted ranking structure and makes the cost of reducing contestable blocks explicit. Across public recommendation, insurance, credit, and mortgage datasets, we compare undamped W-BCE with ordinary BCE using yield Gini, an area summary of loss capture versus blocked value. Logarithmic weight damping is evaluated separately. Held-out evaluations measure capture-disparity tradeoffs. Validation-selected group-aware complaint penalties improve complaint closure in some cohorts. A corrected comparison normalizes and enlarges candidate policy banks using validation-only price sweeps and retains all ten applicable sources. The source-level results do not establish a statistically significant average advantage over group-blind pricing at the fixed minimum retained loss capture. We also compare with blocked-count, blocked-value, and protected-loss parity postprocessors, which do not directly constrain realized objections, and measure objections in candidate-count and loss units.

CCS Concepts

• Computing methodologies → Machine learning.

Keywords

learning to rank, counterfactual ranking, fairness of exposure, ratio-of-expectations, constrained ranking, yield optimization

Introduction

Risk screening and recommendation systems rank candidates whose outcomes carry both value and loss. We study a binary pass/block decision, calling each candidate an account. For account ii, passing creates nonnegative value aia_i and nonnegative cost or loss bib_i; blocking prevents both. In insurance, aia_i can be exposure and bib_i paid claims; in lending and mortgages, they can be return or principal and credit loss; in web recommendation, they can be engagement and negative feedback. If these realized outcomes were known before the decision, a value-budgeted rule would sort accounts by decreasing realized value per unit loss, ai/bia_i/b_i, pass accounts to the left of a vertical cut, and block the low-score accounts to its right. Moving that cut left blocks more accounts and plots blocked value against loss prevented, as in Figure 1. We take ai/bi=+∞a_i/b_i = +\infty when bi=0<aib_i = 0 < a_i; candidates with ai=bi=0a_i = b_i = 0 contribute no value or loss to this objective. Figure 1 labels groups 0 and 1 as G0 and G1.

Figure 1. Value-weighted yield ranking on realized accounts. Each bar is an account with realized value aia_i and loss bib_i, and bars are sorted by realized value per unit loss ai/bia_i/b_i, from high to low. A vertical rank cut passes accounts to its left and blocks accounts to its right; equivalently, it blocks accounts at or below the cut value on the realized value-per-loss scale. Moving the cut from right to left blocks more accounts. The curves show, for each possible cut, the blocked value and loss prevented by the blocked suffix. The group-1 orange outline shows how lower-value accounts can be blocked more often under a common value-per-loss rule; the inset reports the group block counts for the shown cut, with G0 and G1 denoting groups 0 and 1. Markers 1 and 2 identify the similar-loss, different-value pair used in Figure 2.

In realistic systems, the ranking is over expected outcomes. At decision time, the system observes covariates XX and chooses pass or block; AA and BB are the random value and loss outcomes to be learned. AA and BB may both be positive, and their conditional means may depend on XX; when an account is blocked, the value and loss that passing it would have produced are not observed. Let π(X)∈[0,1]\pi(X) \in[0, 1] be the probability of passing: π=1\pi= 1 blocks, π=0\pi= 0 passes, and intermediate values randomize the decision. We call E[πB]\mathbb{E}[\pi B] loss capture, the expected loss prevented.

For some business operating point, say a blocked-value budget β\beta, an upper bound on E[πA]\mathbb{E}[\pi A], the population objective is

max⁡πE[π(X)B]subject toE[π(X)A]≤β.(1)\max_{\pi} \mathbb{E}[\pi(X)B] \quad\text{subject to} \quad\mathbb{E}[\pi(X)A] \le\beta. \tag*{(1)}

The standard budgeted-ranking ordering argument [5, 7, 20, 27] says that a solution ranks by the conditional value per unit loss,

Φ(x)=E[A∣X=x]E[B∣X=x],(2)\Phi(x) = \frac{\mathbb{E}[A \mid X = x]}{\mathbb{E}[B \mid X = x]}, \tag*{(2)}

and blocks accounts with the lowest values of this value-per-loss ratio. We set Φ=+∞\Phi= +\infty when mean loss is zero and mean value is positive; when both means vanish, this objective is indifferent to the decision. This ranking is followed by setting a threshold based on business considerations.

The ratio ranking is population optimal, with ties allowing multiple optimal policies. Different model forms can estimate this order. A ranking loss can target this order directly, but so can weighted binary cross-entropy (W-BCE), a pointwise classifier trained with value and loss as example weights. These and other choices have different implications for business operations, model training, and inference.

Our fairness question concerns conflicts between specified decision norms, within the broader normative questions studied in algorithmic fairness [2]. A value-per-loss rule can disagree with an accepted harm or net-harm rule at the same blocked-value budget. Because value appears in the score, two accounts or groups with similar loss can be ordered differently and blocked at different rates under the same value-per-loss rule. If harm or net-harm criteria are accepted as admissible review standards, then a block those criteria would have passed is a contestable value-driven block, not merely a different outcome rate. Figure 1 illustrates this effect with realized accounts and the induced yield curve. Ordinary pass/block, blocked-value, or protected-loss parity checks do not ask whether the blocked accounts had such objections, so we study the interaction between value-weighted ranking, model form, and group-fairness criteria.

Contributions. Our contributions are:

  1. Operational rationale for W-BCE. Building on classical budgeted allocation and example-dependent cost-sensitive learning [7, 26, 27], we show why W-BCE is useful as an operational model form: it recovers the value-per-loss order at the population optimum, trains on unpreprocessed value and loss outcomes as a standard pointwise classifier with example weights [30, 31], uses one bounded score with no inference-time division, and exposes which outcome transformations preserve the order.

  1. Common-budget objections and group disparity. Fair classification and fair-ranking commonly measure decision rates, label-conditioned rates, or exposure [1, 17, 33, 34]. We instead start from account-level objections: when an accepted alternative criterion would pass a candidate at the same blocked-value budget, a value-per-loss block of that candidate is objectionable under that criterion. We then define protected-group disparity by aggregating these realized objections across groups. This explains why ordinary block, blocked-value, or protected-loss gaps can miss the complaint studied here.

  1. Account and group modifications. We modify W-BCE at the two levels induced by the objection definition. At the account level, objection penalties change the ranking target by subtracting the penalty from protected loss. At the group level, group multipliers characterize supported capture–complaint frontier points under the regularity and dual-attainment conditions of Proposition 4.1. These modifications preserve the value-budgeted ranking structure while making the tradeoff between loss capture and objections explicit, using the same constrained-classification and fair-ranking tools as prior thresholding and reduction methods [1, 17, 33].

  1. Public-data evidence. Across randomized recommendation, insurance, mortgage, lending, and public group-bearing datasets, we compare W-BCE with ordinary BCE and compare group-aware complaint multipliers with group-blind prices and block-, value-, and loss-parity postprocessors. The results compare fixed raw W-BCE with ordinary BCE, expose when existing fairness targets leave complaint disparities, and report cohort-level gains alongside the absence of a statistically significant average group-aware advantage under normalized, enlarged price banks.

Paper outline. In Section 2, we analyze the value-per-loss target, label-only limitations, and the W-BCE reduction. In Section 3, we define norm objections and the objection-penalized ranking rule. In Section 4, we aggregate account-level objections into protected-group disparity and give the corresponding group-aware frontier. In Section 5, we compare W-BCE with ordinary BCE and group-aware complaint-disparity control with group-blind and parity baselines. In Section 6, we discuss related work, limitations, and conclusions.

Yield-Model Form

In Section 1, we introduced the value-versus-loss ranking problem and its value-per-loss target. Here we justify W-BCE as an operational way to estimate that order: W-BCE uses one weighted pointwise classifier, trains on realized value and loss outcomes without preprocessing, and avoids inference-time division by an estimated denominator. In Sections 3 and 4, we reuse the value-budgeted ranking form for account-level objections and protected-group complaint disparities.

We use a simple single-output classifier trained on nn accounts (xi,ai,bi)(x_i,a_i,b_i). Observation ii uses soft label qiq_i and weight wiw_i,

qi=biai+bi,wi=ai+bi,(3)q_i = \frac{b_i}{a_i+b_i}, \qquad w_i = a_i+b_i, \tag*{(3)}

where ai,bi≥0a_i,b_i \ge0 are realized value and loss. Observations with ai+bi=0a_i+b_i=0 receive zero weight and are omitted from W-BCE training because the metric is indifferent to them. Thus bi=0<aib_i = 0 < a_i gives label 0, while ai=0<bia_i = 0 < b_i gives label 1. Let W=A+BW = A + B and Q=B/(A+B)Q = B/(A+B) when W>0W > 0, with Q=0Q = 0 when W=0W = 0. The classifier emits y^θ(x)=σ(sθ(x))∈(0,1)\hat{y}_\theta(\mathbf{x}) = \sigma(s_\theta(\mathbf{x})) \in(0, 1), where σ\sigma is the logistic sigmoid, sθs_\theta is the learned logit, and θ\theta denotes model parameters. The empirical W-BCE loss is

L^W−BCE=−1n∑i=1n[bilog⁡y^θ(xi)+ailog⁡{1−y^θ(xi)}].(4)\widehat{\mathcal{L}}_{\mathrm{W-BCE}} = -\frac{1}{n}\sum_{i=1}^{n}\left[b_i\log\hat{y}_\theta(\mathbf{x}_i) + a_i\log\{1-\hat{y}_\theta(\mathbf{x}_i)\}\right]. \tag*{(4)}

At inference the system ranks by y^θ(x)\hat{y}_\theta(\mathbf{x}) and applies the selected threshold.

Yield-order estimation necessity. For any blocked-value budget β∈[0,E[A]]\beta\in[0, E[A]], the classical fractional-knapsack and Neyman–Pearson ordering argument [5, 7, 20, 27] shows that an optimizer of Eq. (1) blocks accounts with the lowest values of the value-per-loss ratio in Eq. (2), with randomization on ties if necessary.

We can consider several ways to estimate this order. W-BCE does not use a ranking loss, but it is known from weighted proper-loss theory [30, 31] that its label-plus-weight construction in Eq. (3) elicits a score monotone decreasing in Φ\Phi, as stated in Theorem 2.1. We also consider alternative label-only transformations that assign each account one scalar relevance label and then optimize a pointwise, pairwise, or listwise surrogate [3, 36]. Our impossibility claim concerns methods whose optimal order is a monotone function of the conditional mean of that scalar label, as defined below.

Theorem 2.1 (Weighted proper-loss consistency). Optimize predictions over [0,1][0, 1] with proper boundary extensions. For any strictly proper binary loss ℓ\ell, at every covariate value with 0<E[A+B∣X=x]<∞0 < E[A+B \mid X=x] < \infty, the population minimizer of

E[W{Qℓ(1,y^)+(1−Q)ℓ(0,y^)}∣X=x]E\left[W\{Q\ell(1,\hat{y})+(1-Q)\ell(0,\hat{y})\}\mid X=x\right]

is

y^∗(x)=E[B∣X=x]E[A+B∣X=x]=11+Φ(x).\hat{y}^{*}(x)=\frac{E[B\mid X=x]}{E[A+B\mid X=x]}=\frac{1}{1+\Phi(x)}.

Here the last expression equals one when Φ(x)=0\Phi(x)=0. Consequently, descending classifier scores induce the same blocking order as ascending Φ\Phi; where the conditional total stake is zero, the loss and decision are both indifferent. Finite-logit BCE approaches endpoint optima as infima; Appendix C details these boundary cases.

Proof. Conditioning on X=xX=x, the identities WQ=BWQ=B and W(1−Q)=AW(1-Q)=A reduce the risk to E[A+B∣x]{qˉℓ(1,y^)+(1−qˉ)ℓ(0,y^)}E[A+B\mid x]\{\bar{q}\ell(1,\hat{y})+(1-\bar{q})\ell(0,\hat{y})\}, where qˉ=E[B∣x]/E[A+B∣x]\bar{q}=E[B\mid x]/E[A+B\mid x]. Strict properness gives the unique optimum qˉ=1/(1+Φ)\bar{q}=1/(1+\Phi), which decreases strictly with Φ\Phi. □

Define a method as label-only when its population-optimal order is a strictly monotone transform of E[g(B,A)∣X=x]E[g(B,A)\mid X=x] for one scalar label transformation gg. Such methods do not rank correctly in general.

Proposition 2.2 (Label-only limitation). No label-only reduction recovers the yield order universally. Specifically, there do not exist a measurable scalar label g(B,A)g(B,A) and a fixed strictly monotone function HH such that

E[g(B,A)∣X=x]=H(E[A∣X=x]E[B∣X=x])E[g(B,A)\mid X=x]=H\left(\frac{E[A\mid X=x]}{E[B\mid X=x]}\right)

for every conditional distribution of (B,A)(B,A) with finite means and positive mean loss.

Proof. Consider a degenerate distribution concentrated at (B0,A0)(B_0,A_0) with B0>0B_0>0. The assumed identity requires g(B0,A0)=H(A0/B0)g(B_0,A_0)=H(A_0/B_0), so any universally valid scalar label must depend on an observation only through A/BA/B. We write δ(b,a)\delta_{(b,a)} for the distribution assigning probability one to (B,A)=(b,a)(B,A)=(b,a). Now let P1=12δ(1,1)+12δ(2,1)P_1=\frac{1}{2}\delta_{(1,1)}+\frac{1}{2}\delta_{(2,1)} and P2=12δ(1,1)+12δ(6,3)P_2=\frac{1}{2}\delta_{(1,1)}+\frac{1}{2}\delta_{(6,3)}. Under both distributions, A/BA/B equals 1 or 1/21/2 with equal probability, so EP1[g]=EP2[g]E_{P_1}[g]=E_{P_2}[g]. Yet EP1[A]/EP1[B]=2/3E_{P_1}[A]/E_{P_1}[B]=2/3 while EP2[A]/EP2[B]=4/7E_{P_2}[A]/E_{P_2}[B]=4/7. Strict monotonicity of HH requires different expected labels, giving a contradiction. □

This explains common failures: predicting BB ignores blocked value; observed-ratio regression targets E[A/B∣x]E[A/B\mid x] and depends on a zero-denominator convention; unweighted regression on B/(A+B)B/(A+B) learns another expectation of a ratio. When B>0B>0 almost surely and the inverse moments exist, observed-ratio regression has the exact discrepancy, where Cov⁡\operatorname{Cov} denotes conditional covariance,

E[A/B∣x]−E[A∣x]E[B∣x]=E[A∣x](E[1/B∣x]−1E[B∣x])+Cov⁡(A,1/B∣x),(5)E[A/B\mid x]-\frac{E[A\mid x]}{E[B\mid x]}=E[A\mid x]\left(E[1/B\mid x]-\frac{1}{E[B\mid x]}\right)+\operatorname{Cov}(A,1/B\mid x), \tag*{(5)}

so the Jensen term is driven by denominator variation, with a covariance term that can reinforce or offset it.

The instance weight is what turns the fractional label into the ratio of conditional expectations. Without the weight, fractional regression learns

qfrac(x)=E[Q∣X=x],q_{\mathrm{frac}}(x)=E[Q\mid X=x],

an expectation of an observed fraction. W-BCE instead learns

qW(x)=E[WQ∣x]E[W∣x]=E[B∣x]E[B+A∣x]=11+Φ(x),q_W(x)=\frac{E[WQ\mid x]}{E[W\mid x]}=\frac{E[B\mid x]}{E[B+A\mid x]} =\frac{1}{1+\Phi(x)},

which preserves the yield order.

Training and serving simplicity. The preceding results rule out label-only reductions, so the operational contrast is between explicit rankers and W-BCE. An explicit ranker must construct scores or comparisons from the value-per-loss target. A plug-in ranker does this by estimating E[B∣x]E[B\mid x] and E[A∣x]E[A\mid x] separately and dividing at inference; pairwise or listwise rankers must likewise encode the target order in their training comparisons. W-BCE recovers the same population order with one weighted pointwise classifier, avoids inference-time division by an estimated denominator, and gives thresholds the population interpretation

y^∗(x)=11+Φ(x)⟺Φ(x)=1−y^∗(x)y^∗(x).\hat{y}^{*}(x)=\frac{1}{1+\Phi(x)} \quad\Longleftrightarrow\quad \Phi(x)=\frac{1-\hat{y}^{*}(x)}{\hat{y}^{*}(x)}.

Preprocessing effects on ranking. Ratio-based implementations, including plain W-BCE, raise preprocessing questions. In finite samples, we may clip, cap, or damp the soft label, the weight, or the implied numerator and denominator. The next proposition, proved in Appendix B, shows that preprocessing makes W-BCE optimize a transformed numerator and denominator. Only a narrow linear class preserves the value-per-loss order for every distribution.

Proposition 2.3 (Preprocessing changes the ranking target). Let Q′∈[0,1]Q' \in[0,1] and W′>0W' > 0 be finite measurable transformations on the full outcome cone A,B≥0A,B \ge0, A+B>0A+B>0. Assume finite conditional transformed means and positive transformed denominator. After replacing (Q,W)(Q,W) by (Q′,W′)(Q',W'), W-BCE ranks by

E[W′Q′∣X=x]E[W′(1−Q′)∣X=x].\frac{E[W'Q' \mid X=x]}{E[W'(1-Q') \mid X=x]}.

On all finite-support distributions on that cone with positive outcome means and positive transformed denominator, this order equals a fixed strictly decreasing function of Φ\Phi for every such distribution if and only if the transformed outcome sides satisfy

W′Q′=p1B+p2A,W′(1−Q′)=r1B+r2A,W'Q'=p_1B+p_2A,\qquad W'(1-Q')=r_1B+r_2A,

for fixed nonnegative constants p1,p2,r1,r2p_1,p_2,r_1,r_2 with p1r2−p2r1>0p_1r_2-p_2r_1>0. In particular, nonlinear weight damping is not equivalent to raw W-BCE.

Proof sketch. Put U=W′Q′U=W'Q' and V=W′(1−Q′)V=W'(1-Q'), and let HH be the fixed strictly decreasing ratio function in the hypothesis. For each t>0t>0, two-point mixtures with E[A−tB]=0E[A-tB]=0 must also have E[U−H(t)V]=0E[U-H(t)V]=0. Varying mixture endpoints gives U−H(t)V=ct(A−tB)U-H(t)V=c_t(A-tB) pointwise for some scalar ctc_t. Two distinct values of tt express U,VU,V linearly in A,BA,B. Nonnegativity on the outcome cone forces nonnegative coefficients, and strict decrease of HH forces the positive determinant. Conversely, the ratio is (p1+p2Φ)/(r1+r2Φ)(p_1+p_2\Phi)/(r_1+r_2\Phi), whose derivative has the required negative sign. Appendix B gives the full argument. □

Thus clipping, logarithms, square roots, and other nonlinear damping are finite-sample bias–robustness choices, not equivalent implementations of W-BCE.

Account-Level Objections

We return to the fairness issue illustrated in Figure 1. An account blocked by the value-per-loss rule may be passed by another specified norm criterion at the same blocked-value budget. We call this an account-level objection. Once we accept an alternative criterion at the same blocked-value budget, a value-rule block of a candidate that criterion would pass is a contestable block under the comparison-based logic of constrained fair classification, cost-of-fairness analysis, reductions, and fair ranking [1, 6, 17, 26, 33, 34].

We do not choose the admissible criteria; we take them as specified alternatives. Having motivated W-BCE in § 2, we add an objection penalty at the account level. Appendix D gives the validation procedure, and § 5 reports the empirical frontier. In § 4, we discuss protected-group complaint disparity. Figure 2 illustrates how objection penalties change the baseline order for accounts 1 and 2 marked in Figure 1.

Figure 2. Objection penalties for accounts 1 and 2 in Figure 1. Subtracting penalties from loss changes the value-per-loss scores and the account order. Account 1 still passes; account 2 moves from block to pass. Dashed marks show the original loss on the priced bars.

Norm-criterion objections. Let X\mathcal{X} be the account covariate space. A policy is a block/pass rule

π:X→[0,1],\pi:\mathcal{X}\to[0,1],

where π(x)=1\pi(x)=1 blocks account xx, π(x)=0\pi(x)=0 passes the account, and intermediate values represent threshold randomization. To devise a policy, we need a score and a threshold.

Let λ0≥0\lambda_0\ge0 convert value units into loss units for the net-harm criterion. In this work we consider the following score functions:

sratio(x)=E[A∣X=x]E[B∣X=x],s_{\mathrm{ratio}}(x)=\frac{E[A\mid X=x]}{E[B\mid X=x]},
sharm(x)=E[B∣X=x],s_{\mathrm{harm}}(x)=E[B\mid X=x],
snet(x)=E[B∣X=x]−λ0E[A∣X=x].s_{\mathrm{net}}(x)=E[B\mid X=x]-\lambda_0E[A\mid X=x].

The first is the reference yield criterion and minimizes the yield curve by blocking low scores (Section 2). The second ranks by expected harm. The third ranks by expected harm net of λ0\lambda_0 times expected value. Insurance and credit experiments use λ0=1\lambda_0=1 in their recorded units; the video task uses λ0≈1.6424×10−4\lambda_0\approx1.6424\times10^{-4} loss units per second of play time. These are not the only possible scores reflecting norms, but they convey the point that conflicting norms can exist. In the remainder, the norm set is C={sharm,snet}C=\{s_{\mathrm{harm}},s_{\mathrm{net}}\}.

To ask whether objection norm criterion s∈Cs\in C would pass an account, let τs(β)\tau_s(\beta) be the score threshold chosen to satisfy blocked-value budget β\beta. We use the blocking rule:

πsβ(x)=1(s(x)≥τs(β)),E[πsβA]≤β,\pi_s^\beta(x)=1(s(x)\ge\tau_s(\beta)),\qquad E[\pi_s^\beta A]\le\beta,

(with randomization on ties, if needed). πsβ(x)=0\pi_s^\beta(x)=0 therefore means norm criterion ss would pass xx at the common operating point. For any policy π\pi, the account-level objection generated by norm criterion ss is

Os(x;π)=π(x){1−πsβ(x)}.O_s(x;\pi)=\pi(x)\{1-\pi_s^\beta(x)\}.

It is one exactly when π\pi blocks the account but norm criterion ss would pass. We define the all-criteria count objection as

OC(x;π)=π(x)(∃s∈C:πsβ(x)=0).O_C(x;\pi)=\pi(x)(\exists s\in C:\pi_s^\beta(x)=0).

Objections can be penalized in three measurement units. Using w=1w=1 counts accounts, w=Aw=A measures blocked value, and w=Bw=B measures protected loss. The count version uses OCO_C directly; the blocked-value and protected-loss versions multiply the same objection event by the corresponding measurement unit.

Objection-penalized ranking. The objections enter directly on the W-BCE loss side. Let λs≥0\lambda_s\ge0 weight blocking an account that norm criterion s∈Cs\in C would pass. Write λ=(λs)s∈C\lambda=(\lambda_s)_{s\in C} for the penalty vector. For count objections, we penalize

mλ(x)=∑s∈Cλs1(πsβ(x)=0).m_\lambda(x)=\sum_{s\in C}\lambda_s1(\pi_s^\beta(x)=0).

For value or loss units, multiply each indicator by the corresponding conditional mean given XX; empirical policies use fixed auxiliary predictions. This keeps mλ(x)m_\lambda(x) deterministic at decision time. The objection-penalized policy maximizes

max⁡0≤π≤1E[π{B−mλ(X)}]s.t.E[πA]≤β.(6)\max_{0\le\pi\le1}E[\pi\{B-m_\lambda(X)\}] \quad\text{s.t.}\quad E[\pi A]\le\beta. \tag*{(6)}

Thus λs=0\lambda_s=0 recovers the ordinary yield rule, and increasing λs\lambda_s penalizes a contested block under criterion ss.

Proposition 3.1 (Objection-penalized order). An optimizer of Eq. (6) thresholds

Ψλ(x)=E[A∣X=x]E[B∣X=x]−mλ(x),(7)\Psi_\lambda(x)=\frac{E[A\mid X=x]}{E[B\mid X=x]-m_\lambda(x)}, \tag*{(7)}

from below, with randomization on ties; nonpositive denominators are never blocked. If the best positive part of E[B∣X=x]−mλ(x)E[B\mid X=x]-m_\lambda(x) spends less than β\beta, the optimum under-spends the budget.

Proof. The Lagrangian with multiplier μ≥0\mu\ge0 for the blocked-value budget is

E[π{B−mλ(X)−μA}]+μβ.E[\pi\{B-m_{\lambda}(X)-\mu A\}]+\mu\beta.

After conditioning on X=xX=x, pointwise maximization blocks exactly when

E[B∣X=x]−mλ(x)>μE[A∣X=x],E[B \mid X=x]-m_{\lambda}(x)>\mu E[A \mid X=x],

which is equivalent to thresholding Eq. (7) from below on accounts with positive penalized loss. If the objective coefficient is nonpositive outside a set whose value is below β\beta, blocking additional accounts only lowers Eq. (6), so the budget ceiling is slack.

The empirical account implementation keeps its fitted predictors unchanged and applies this penalty as a postprocessor. The population rule does not require training with negative loss weights or labels outside [0,1][0,1].

Group-Level Disparity

In § 3, we defined when a blocked candidate has an account-level objection. We now define group-level disparity by aggregating those realized objections within groups and comparing the group aggregates. For simplicity, we write the section for two groups, g∈{0,1}g \in\{0,1\}, and control disparity in either direction through a scalar gap. Let GG be the random group label. With more groups, or when only selected group comparisons are relevant, the same construction applies to a vector of chosen group-disparity constraints. We first define the group complaint disparity, then give the constrained allocation rule; Appendix § 14 gives an algebraic denominator decomposition.

Group complaint disparity. In § 3, we defined the all-criteria account objection OC(x;π)O_C(x;\pi). For the group level, allow π:X×{0,1}→[0,1]\pi:\mathcal{X}\times\{0,1\}\to[0,1] to depend on covariates and group. We use

OC(x,g;π)=π(x,g)1(∃s∈C:πsβ(x)=0).(8)O_C(x,g;\pi)=\pi(x,g)1(\exists s\in C:\pi_s^{\beta}(x)=0). \tag*{(8)}

Let U1=1U_1=1, UA=AU_A=A, and UB=BU_B=B. For each reported unit w∈{1,A,B}w\in\{1,A,B\}, assume Pr⁡(G=g)>0\Pr(G=g)>0 and 0<E[Uw∣G=g]<∞0<E[U_w\mid G=g]<\infty, and let

Dgw(π)=E[OC(X,G;π)Uw∣G=g]E[Uw∣G=g].(9)D_g^w(\pi)=\frac{E[O_C(X,G;\pi)U_w\mid G=g]}{E[U_w\mid G=g]}. \tag*{(9)}

For two groups, we define the group complaint disparity as

ΔDw(π)=D1w(π)−D0w(π).(10)\Delta_D^w(\pi)=D_1^w(\pi)-D_0^w(\pi). \tag*{(10)}

Here Dgw(π)D_g^w(\pi) is the normalized realized objection imposed on group gg, per unit of that group’s own count, value, or loss. The denominator is the group’s total amount in unit ww, not only the accounts with objections; this keeps a group’s association with account-level objections inside the group disparity being measured. In the running example, the lower-revenue group can have larger DgAD_g^A because the value-per-loss rule blocks more of its objection-bearing value despite similar expected harm.

Figure 5 in Appendix § 14 visualizes this denominator choice by regrouping the account table from Figure 1. The left panel shows first the bars by groups; the right panel aggregates the same value masses into each group’s total value denominator, objection-bearing value, and blocked objection-bearing value. A denominator restricted to objection-bearing value would remove the group association captured by DgAD_g^A.

Complaint-disparity control. Given ΔDw\Delta_D^w, group complaint-disparity control is a constrained allocation problem. With two groups (for simplicity), we maximize prevented loss subject to the blocked-value budget and a complaint-disparity target with tolerance ϵ≥0\epsilon\ge0,

max⁡0≤π≤1E[πB]s.t.E[πA]≤β,∣ΔDw(π)∣≤ϵ.(11)\max_{0\le\pi\le1} E[\pi B]\quad\text{s.t.}\quad E[\pi A]\le\beta,\qquad\lvert\Delta_D^w(\pi)\rvert\le\epsilon. \tag*{(11)}

Sweeping the complaint tolerance or its multiplier gives the capture–disparity frontier. With more than two groups, the scalar complaint multiplier becomes a vector of group coefficients and the same Lagrange form applies.

For W-BCE implementation, denote the indicator in Eq. (8) by qC(x)q_C(x), so

OC(X,G;π)=π(X,G)qC(X).(12)O_C(X,G;\pi)=\pi(X,G)q_C(X). \tag*{(12)}

Let

uw(x,g)={1,w=1,E[A∣X=x,G=g],w=A,E[B∣X=x,G=g],w=B,u_w(x,g)= \begin{cases} 1, & w=1,\\ E[A\mid X=x,G=g], & w=A,\\ E[B\mid X=x,G=g], & w=B, \end{cases}

and use validation rows to estimate each group denominator.

Let Pg=Pr⁡(G=g)>0P_g=\Pr(G=g)>0. Attach nonnegative multipliers α+\alpha_+ and α−\alpha_- to ΔDw≤ϵ\Delta_D^w\le\epsilon and −ΔDw≤ϵ-\Delta_D^w\le\epsilon, respectively, and set ν=α+−α−\nu=\alpha_+-\alpha_-. The conditional coefficients use κ1=ν/P1\kappa_1=\nu/P_1 and κ0=−ν/P0\kappa_0=-\nu/P_0, because the capture objective averages groups with weights PgP_g while each complaint rate conditions on its own group. Write κ=(κ0,κ1)\kappa=(\kappa_0,\kappa_1) for this group coefficient vector. Let μ≥0\mu\ge0 be the blocked-value budget multiplier.

Proposition 4.1 (Optimal complaint-disparity postprocessor). Fix a measurement unit ww and plug-in estimates of E[A∣X,G]E[A\mid X,G], E[B∣X,G]E[B\mid X,G], and qCq_C. Let Ugw=E[uw(X,g)∣G=g]>0U_g^w=E[u_w(X,g)\mid G=g]>0 be the

group denominator. Define ag(x)=E[A∣X=x,G=g]a_g(x)=E[A\mid X=x,G=g] and bg(x)=E[B∣X=x,G=g]b_g(x)=E[B\mid X=x,G=g]. Assume integrable outcome and complaint coefficients, finite positive group denominators, no duality gap, and attainment by finite multipliers. At a capture-maximizing supported point on the linear blocked-value and complaint-disparity frontier, an optimal randomized binary policy has the form

πμ,κ(x,g)=1(bg(x)−μag(x)−κgqC(x)uw(x,g)/Ugw>0),(13)\pi_{\mu,\kappa}(x,g)=\mathbf{1}\left(b_g(x)-\mu a_g(x)-\kappa_g q_C(x)u_w(x,g)/U_g^w>0\right), \tag*{(13)}

with randomization on ties.

Proof. In the finite empirical problem, the randomized allocation set is a polytope; Appendix C gives the population conditions. The objective and constraints in Eq. (11) are linear in π\pi. Attach a multiplier μ\mu to the blocked-value budget and group multipliers to the complaint constraints. After conditioning on (X,G)=(x,g)(X,G)=(x,g), the Lagrangian is pointwise linear in π(x,g)\pi(x,g) with coefficient bg(x)−μag(x)−κgqC(x)uw(x,g)/Ugwb_g(x)-\mu a_g(x)-\kappa_gq_C(x)u_w(x,g)/U_g^w, giving Eq. (13); ties may be randomized to hit exact constraints. □

Proposition 4.1 gives a tilted value/loss rule. The multiplier μ\mu sets the blocked-value operating point, while a signed group-penalty contrast tilts decisions against ΔDw\Delta_D^w. Because Eq. (11) constrains ∣ΔDw∣\lvert\Delta_D^w\rvert, the active side fixes the sign: if D1w>D0wD_1^w>D_0^w, the tilt penalizes group-1 objections relative to group 0; if D1w<D0wD_1^w<D_0^w, it reverses. Sweeping that contrast and solving for μ\mu on validation data gives the capture–complaint frontier.

At a required loss-capture level, exact complaint parity may be infeasible. Let π0\pi_0 be a budget-feasible reference policy. Fix a minimum capture requirement E[πB]≥ηE[\pi B]\ge\eta, for example η=0.9E[π0B]\eta=0.9E[\pi_0B] for 90% of the baseline policy’s capture. Among all policies that satisfy this capture requirement and the value budget, collect the attainable complaint gaps Δw(π)=D1w(π)−D0w(π)\Delta^w(\pi)=D_1^w(\pi)-D_0^w(\pi). Proposition 4.2 shows that these attainable gap values form an interval. Exact parity is feasible if and only if that interval contains zero; otherwise its closest endpoint gives the unavoidable residual gap.

Proposition 4.2 (Frontier feasibility). Fix an absolute protected-loss capture floor η\eta, in the same units as E[πB]E[\pi B], and define Gη={D1w(π)−D0w(π):0≤π≤1, E[πA]≤β, E[πB]≥η}G_\eta=\{D_1^w(\pi)-D_0^w(\pi):0\le\pi\le1,\ E[\pi A]\le\beta,\ E[\pi B]\ge\eta\}. For a fractional floor ρ\rho relative to a budget-feasible baseline policy π0\pi_0, set η=ρE[π0B]\eta=\rho E[\pi_0B]. If this feasible set is nonempty, AA, BB and complaint coefficients are integrable, and group denominators are positive and finite, GηG_\eta is a compact interval. Exact complaint parity at that operating point is feasible if and only if 0∈Gη0\in G_\eta; otherwise the smallest achievable absolute disparity is inf⁡Δ∈Gη∣Δ∣>0\inf_{\Delta\in G_\eta}\lvert\Delta\rvert>0.

Proof. The randomized-policy feasible set is convex, and D1w(π)−D0w(π)D_1^w(\pi)-D_0^w(\pi) is linear in π\pi. A linear image of a convex set in R\mathbb{R} is an interval. The order interval 0≤π≤10\le\pi\le1 is weak-∗* compact in L∞L^\infty. Integration against the integrable objective and constraint coefficients is weak-∗* continuous, so the feasible set is closed and compact. Its complaint image is therefore a compact interval; the remaining claims follow by taking its distance from zero. □

Together, Propositions 4.1 and 4.2 identify the frontier we evaluate in the group experiments in § 5.3.

Experiments

We evaluate capture–objection trade-offs and validation-selected group-aware versus group-blind complaint pricing.

Experimental Settings

Score comparison. We compare fixed raw W-BCE with ordinary BCE on ten public recommendation, insurance, credit, and mortgage datasets. Value and loss represent engagement and negative feedback, exposure and claims, or lending value and credit loss. For cumulative loss capture c(v)c(v) at blocked-value fraction vv, yield Gini is 2∫01c(v) dv−12\int_0^1c(v)\,\mathrm{d}v-1 [23]. Figure 3 reports paired differences in yield Gini and capture at a 10% value budget, using ten splits. Capture thresholds are validation-frozen [V] or retrospectively interpolated on the test curve [T]; the latter is a ranking measure. Positive paired means occur on 9 of 10 datasets for yield Gini and 8 of 10 for capture; these counts do not establish significance for every dataset.

Figure 3. Raw W-BCE minus BCE on ten datasets. Each pair shows yield-Gini and loss-capture deltas; whiskers are 95% paired Student tt intervals over ten splits. Positive values favor W-BCE. Dataset suffixes describe capture: [V] validation-frozen threshold; [T] retrospective test-curve interpolation.

Fairness comparison. Account-level frontiers use Australian MTPL, Belgian MTPL, and KuaiRand. Group pricing uses seven independent source families (180 cohort-seed cells), with an expanded ten-source sensitivity (240 cells), adding Centers for Medicare & Medicaid Services (CMS) provider fraud, University of California, Irvine (UCI) credit, and German credit. A cell is one fitted cohort and seed; repeated cohorts and French portfolio panels are pooled within their source. We average source-median paired effects for source-level inference. Policies share a 10% value budget. Predictors are fitted on training rows and held fixed during pricing; thresholds and prices use validation rows, and selected policies are evaluated on held-out test rows. The primary capture floor is .95 of unpriced W-BCE capture. Appendix A gives datasets, outcome constructions, sample sizes, groups, splits, learners, and preprocessing; Section 5.3 gives the comparison protocol.

Account-Level Frontiers

A blocked account can object when a harm or net-harm criterion would pass it at the same value budget (Eq. (6)). Figure 4 compares fitted W-BCE, BCE, and unpriced plug-in ratios with union- and separate-objection price policies on the same accounts. A union-price policy charges one penalty when either norm would pass the account; a separate-price policy sums the harm and net-harm penalties and can charge both. Union prices apply to plug-in loss and fitted W-BCE implied loss; separate prices use plug-in loss. Appendix D gives the scores, validation-scaled price grids, and common budget correction.

Figure 4. Account-objection trade-offs relative to fitted W-BCE. Green points are nondominated priced policies. Gray shading is dominated by unpriced baselines; green shading shows the additional dominated region. Markers show W-BCE, BCE, and plug-in ratios with 95% paired-split intervals. Policies use validation-frozen thresholds and thinning at a 10% value cap. The test-mean frontier is descriptive; higher and farther right is better.

Every baseline marker is dominated by an evaluated priced policy on test means. Complete objection removal retains about 60% of W-BCE capture in Australian MTPL and 82% in Belgian MTPL (Appendix E); KuaiRand has zero-objection policies with greater capture than fitted W-BCE. These describe the evaluated bank. The group comparison below instead evaluates validation-selected policies.

Group-Level Disparity Frontiers

For group disparity, we use a deployment-style selection protocol rather than an in-hindsight envelope comparison. For each split, every family must clear the same validation capture floor; each family then selects the feasible policy with the smallest validation complaint gap and is scored once on held-out test rows. The main ablation is within-method: a group-blind account-price family uses one complaint price for all accounts, while the group-aware family uses an opposite signed group contrast. The common price is normalized by pooled validation mass, while the signed group contrast uses each group’s validation mass. Both families search a common axis that retains every original physical policy and extends geometrically using validation only. We require no selected policy at the largest price and two further extensions with unchanged selections. Appendix G gives the formulas, preservation mapping, coverage rule, and historical results. This enlargement is a post-review sensitivity on the existing test rows; the .95 capture floor was fixed before this correction.

Both price families use the fitted W-BCE implied loss side; auxiliary regressions construct objections and measurement weights. These plug-in policies have finite-model and clipping error, not population-optimality guarantees. Appendix G specifies their scores. Fixed block-, value-, and loss-parity postprocessors are reported as context. We retain the original .95 validation capture floor as the primary operating point and report the full .90–1.00 grid.

Before selection, let νval\nu_{\mathrm{val}} be a policy’s blocked value divided by total validation value. We apply independent thinning with probability α=min⁡(1,.10/νval)\alpha= \min(1, .10/\nu_{\mathrm{val}}), taking α=1\alpha= 1 when νval=0\nu_{\mathrm{val}} = 0, to every family and the unpriced baseline. Capture floors use the corrected baseline. We freeze this probability, the selected penalties, and thresholds on test, and report expected metrics under thinning. The validation blocked-value fraction is at most .10; the test fraction is measured separately.

FloorSeven meanppTen meanpp
0.90+0.03910.8426-0.03720.8008
0.95+0.21850.3071+0.08630.6007
0.98+0.19690.3214+0.07350.6333
1.00+0.10930.5482+0.04720.7127

Table 1. Corrected validation-selected value-unit comparison. Each mean averages source medians of paired held-out closure lift. The seven-source scope and all ten recorded sources are reported separately; pp is a two-sided source tt test. Prices are normalized and extended using validation only.

Let π0\pi_0 be unpriced W-BCE and LP=Etest[πPB]L_P = E_{\mathrm{test}}[\pi_P B] the loss capture of policy PP, including frozen thinning; L0=Etest[π0B]L_0 = E_{\mathrm{test}}[\pi_0 B]. For test complaint gaps dP=∣Δw(πP)∣d_P = \lvert\Delta^{w}(\pi_P) \rvert and d0=∣Δw(π0)∣d_0 = \lvert\Delta^{w}(\pi_0) \rvert, closure is CP=1−dP/d0C_P = 1 - d_P/d_0. Lift is Cgroup−CcomparatorC_{\mathrm{group}} - C_{\mathrm{comparator}}, the selected group-aware policy’s closure minus the comparator’s closure, without clipping. When d0=0d_0 = 0, closure is one if dP=0d_P = 0 and zero otherwise; relative improvement is undefined in these cells.

At the .95 floor, mean source-median lift is +.2185 across seven sources (two-sided p=.3071p = .3071) and +.0863 across ten (p=.6007p = .6007). Neither establishes a statistically significant average group-aware advantage. Raw cell means are −.2274 and −.5592, respectively; these measure different effects. Table 2 retains all source medians. Appendix G reports further aggregation and global-grid sensitivities.

SourceMedian lift
Australian MTPL+0.5393
Belgian MTPL+0.4448
Bondora+0.4892
CMS provider fraud-0.6403
EU health insurance+0.4899
French portfolio-0.0027
German credit+0.0000
Norwegian auto-0.8800
Swedish motorcycle+0.4489
UCI credit-0.0257

Table 2. Corrected value-unit group-aware versus group-blind complaint closure lift at the .95 capture floor. Each row is one source’s median paired held-out lift. Negative and zero effects are retained.

The median held-out capture shortfall max⁡(0,.95−LP/L0)\max(0, .95 - L_P/L_0) is zero in the seven-source scope; this does not guarantee the floor on every test split. Count-unit effects are mixed; loss-unit average lifts are negative at every reported floor under both source scopes and aggregations. Their comparisons and all source-mean sensitivities appear in Appendix G. Fixed parity postprocessors constrain quantities other than complaint disparity; their comparisons are diagnostic. The stronger historical result depended on the original truncated price bank and is recorded separately in that appendix.

We also compare with linear-program (LP) mixtures of the full group-blind bank under the same validation constraints. Mean ten-source median lift is −.0521 (p=.7339p = .7339), with five negative source medians, again not establishing average superiority. Appendix G gives the LP procedure and source results (Table 6).

Related Work, Limitations, and Conclusion

Related work. Binns connects algorithmic fairness to political philosophy [2]; Lyons et al. study contestability [24]. De Vos et al. constrain demographic parity within relevant decision regions [8], while Höppner et al. derive instance-dependent cost-sensitive fraud decisions [18]. These works motivate normative, decision, and cost-sensitive perspectives. Our target is disagreement between yield and accepted review policies at a common blocked-value budget, aggregated into group complaint disparity and priced explicitly. Fair-classification thresholds and reductions [1, 4, 6, 17, 21, 26] and fair-ranking trade-offs [33, 34] supply established tools. Budgeted allocation, ratio ranking, and proper losses provide the learning and optimization foundations [5, 7, 13, 16, 20, 22, 27, 29, 30, 31]; logged ranking and partial identification inform observation limits [19, 25].

Limitations. The admissible review criteria are normative inputs: disagreement identifies an objection under a specified criterion, without establishing which criterion should govern. The policy class is binary pass/block under a value budget; position-dependent ranking or intervention intensity requires a richer action space. The measured outcomes are proxies: insurance exposure measures coverage rather than profit, paid claims require causal assumptions to represent preventable loss, and KuaiRand hate feedback measures observed responses; its historical batch-size override cannot be verified because the original run arguments were not retained. CMS uses a provider potential-fraud flag, and UCI/German credit convert binary labels into amount-weighted loss proxies; these do not measure confirmed or preventable cash loss. Selective credit outcomes also limit identification[11, 19, 25]. The guarantees assume integrability, positive denominators, and dual attainment; they do not imply fitted-model or finite-grid optimality. Empirical frontiers

cover evaluated policies and the LP their bank mixtures, with possible headroom as in Norwegian auto. Raw closure divides by the baseline complaint gap and can amplify small-gap cells; the positive mean of source medians does not imply a positive mean across all cells. Group-aware and group-blind price families share a grid but need not nest. The seven-source scope is predominantly insurance; the ten-source sensitivity adds credit and provider fraud. The post-review correction reuses examined test rows; bank extension and selection use validation only. Independent replication remains necessary. Recorded binary sex or gender groups cannot establish causal discrimination or represent intersectional identities.

Conclusion. W-BCE elicits the value-per-loss order for budgeted blocking. Pricing objections from accepted alternative criteria gives capture–complaint trade-offs. Validation-selected group-aware prices improve closure in some cohorts; the normalized, enlarged comparison does not establish a statistically significant average advantage over group-blind pricing. Fixed parity comparisons remain mixed. The method makes the objection criterion and its capture cost explicit.

Ethical Considerations

Errors in predicted value or loss can wrongly block accounts, and reducing a group complaint gap can shift harm between groups. Group-aware deployment uses group membership at decision time; it requires justified criteria, privacy controls, subgroup audits, and appeal paths for consequential blocks. Our experiments use public retrospective data, with no production deployment or new personal-data collection.

References

Experiment Protocol Details

Public score-learning and candidate-log runners, with input and dependency instructions, are available at https://github.com/atavory/yielding_fairness. The reproduction artifact supplies the frozen table and figure inputs, original generators, commands, and checksums in artifacts/frozen_paper_evidence_20261011.zip; its bandha_paper_reproduction.md maps every paper artifact to its generator.

Datasets, cohorts, and model settings

Score setting. The score-learning check uses public recommendation, insurance, credit, and mortgage datasets: KuaiRand-Pure randomized short-video exposures [14]; motor third-party liability (MTPL) insurance data from CASdatasets [9]; Lending Club [15], Prosper [28], Small Business Administration (SBA) 7(a) [35], and Bondora [32] credit portfolios; and Fannie Mae [10] and Freddie Mac [12] mortgage samples. Loss BB is the hate indicator for videos, paid claims for insurance, and realized credit loss for loans. Value AA is video play time, insurance exposure, or realized lending return; the mortgage tasks use original principal on non-loss loans. Australian insurance masks exposure on claim rows; Belgian and French score tasks retain it. Lending Club, Prosper, and Bondora mask return on loss rows; SBA retains remaining principal. Thus the task constructions need not make AA and BB mutually exclusive. For a descending score order, let c(v)c(v) be cumulative loss capture at cumulative value fraction vv, with linear interpolation between accounts. Yield Gini is 2∫01c(v) dv−12\int_0^1 c(v)\,\mathrm{d}v-1, evaluated by trapezoidal integration [23]. We compare fixed raw W-BCE with BCE; log-damped W-BCE is a separate robustness comparison, never selected using test outcomes. Positive paired mean differences occur on 9 of 10 datasets for yield Gini and 8 of 10 for capture; these counts do not establish significance for every dataset.

DatasetSample sizeSplit
KuaiRand1,186,05964/16/20
Australian MTPL67,85664/16/20
freMTPL2100,00080/0/20
Belgian MTPL100,00080/0/20
Lending Club100,00080/0/20
Prosper55,05180/0/20
SBA 7(a)100,00080/0/20
Bondora52,79380/0/20
Fannie Mae300,00064/16/20
Freddie Mac300,00064/16/20

Table 3. Score samples and training/validation/test percentages. Each dataset uses ten paired splits; SBA denotes the Small Business Administration.

SourceCohortsTest rowsGroups
Australian MTPL213,571Female/Male
Belgian MTPL310,000–32,642Female/Male
French portfolio93,200–7,660Female/Male
Bondora110,006Female/Male
EU health135,207F/M
Norwegian auto110,0000/1
Swedish motorcycle110,000Female/Male

Table 4. Group-evaluation cohorts and held-out rows per split. Each fitted cohort has ten paired seeds; multiple fitted banks from one source remain one source for inference. F/M means Female/Male; these labels include their recorded case variants.

All score comparisons use ten paired splits and the same two-hidden-layer multilayer perceptron (MLP; 64/32 units with rectified linear unit (ReLU) activations), Adam with learning rate 10−310^{-3} and 20 epochs. The archived KuaiRand runner defaults to batch size 512, but its original run arguments were not retained; we cannot verify whether that default was overridden. Numeric standardization and categorical encoding are fit on training rows. Figure 3 uses [V] for validation-frozen capture thresholds and [T] for retrospective test-curve interpolation. Table 3 records the frozen sample sizes and training/validation/test proportions. In the four 64/16/20 evaluations, the 10% value threshold is chosen on validation and frozen on test. The six 80/0/20 evaluations report retrospective c(.10)c(.10) on the test curve, which is a ranking measure rather than a deployed threshold. The freMTPL2 comparison was rerun with training-only preprocessing; its artifact stores split identities, predictions, and transformation parameters for independent metric replay. Ordinary BCE uses the binary loss-occurrence label with unit weights.

Fairness setting. The account-objection frontier uses one natural public cohort from each of Australian MTPL, Belgian MTPL, and KuaiRand. The group-complaint evaluation uses public group-bearing evaluations from seven independent sources: Australian MTPL, Belgian MTPL, Bondora, EU health insurance, the French motor-insurance dataset family freMPL, Norwegian auto, and Swedish motorcycle. The freMPL1–freMPL9 panels are parts of one portfolio and are not counted as independent sources. Unless otherwise stated, policies share a 10% aggregate value budget. Fairness tasks use 64/16/20 training/validation/test splits and training-fitted preprocessing. Scores use linear weighted-logistic W-BCE and BCE fit by full-batch Adam (learning rate .05, ℓ2\ell_2 (squared Euclidean norm) regularization coefficient 10−410^{-4}). Separate raw-scale linear regressions supply nonnegative value and loss proxies (learning rate .03, ℓ2\ell_2 coefficient 10−410^{-4}). These predictors are fixed during postprocessor comparisons; only thresholds and complaint prices change. Appendix A gives split, learner, preprocessing, aggregation, and reporting details.

Group evaluation units. Table 4 records the held-out sizes and groups. A cell is one fitted cohort and split seed; a policy family is the common-price, signed group-price, or fixed-parity candidate set evaluated on that cell. Seeds are 0–9. The .95 comparison has 180 paired cells: 20 Australian, 30 Belgian, 90 French, and ten each for Bondora, EU health, Norwegian auto, and Swedish motorcycle. Cached Australian and Belgian names denote separate fitted banks, not independent sources; the French panels form one portfolio. Within each source we take the median paired test lift, then average the seven source medians and apply source-level inference. A family is eligible only if it has a validation-budget-feasible policy clearing the capture floor; both families must be eligible for a paired cell. For the seven-source scope, grouping fields are recorded binary sex/gender (Female/Male or F/M); Norwegian auto uses Male coded 0/1. The public insurance cohorts use exposure on no-claim rows and paid claims; EU health uses positive policy-years on no-claim rows and the sum of nonnegative claim-amount fields. Bondora uses the lending value/loss construction described above. Public outcome predictors exclude the group field. Appendix A documents cohort construction; the public artifact records per-cohort filters and feature lists.

The full ten-source sensitivity adds two grouping cohorts per source for Centers for Medicare & Medicaid Services (CMS) provider fraud, University of California, Irvine (UCI) credit, and German credit, giving 240 cells per floor. These use A=M(1−Y)A=M(1-Y) and B=MYB=MY: MM is reimbursed claim amount and YY the provider’s potential fraud flag in CMS; credit limit and next-month default in UCI; and credit amount and the bad-credit label in German credit. They are label-based outcome proxies, not measured preventable losses. CMS uses recorded gender and race code 1 versus other codes; UCI uses sex and age below 40 versus at least 40; German credit uses sex derived from the recorded personal-status code and age below 35 versus at least 35. The two grouping attributes are pooled within their source for inference.

Encoding and reporting. Categorical variables are one-hot encoded and numeric variables are standardized using training rows only. Policy thresholds or penalties are selected on validation rows, and reported policy results use untouched test rows. A paper-table aggregator in the data artifact generates manuscript table rows and confidence intervals from frozen seed-level CSVs. Tables report split means with 95% confidence intervals where space permits; compact slash-entry tables have per-entry intervals in the generated CSVs.

For existing protected-group metrics, we report loss capture and absolute gaps in counted blocks, blocked value, and protected loss. For a measurement unit w∈{1,A,B}w \in\{1, A, B\}, the group rate is

Hgw(π)=E[π(X,G)w∣G=g]E[w∣G=g],H_g^w(\pi) = \frac{E[\pi(X,G)w \mid G = g]}{E[w \mid G = g]},

and the reported gap is max⁡gHgw(π)−min⁡gHgw(π)\max_g H_g^w(\pi) - \min_g H_g^w(\pi).

Historical group-complaint ablation. The original-bank diagnostics below precede the normalized and enlarged comparison in Appendix G. Its original numbers are retained as provenance; they are not the corrected main results.

For the group-complaint selected-policy table, each policy family is selected only on validation rows. The baseline is the unpriced W-BCE value-budget rule. The group-blind account-price family subtracts the same complaint price from all objection-bearing accounts. The group-aware family uses the same lambda grid but applies opposite per-group multipliers to reduce the validation complaint gap. The fixed family consists of the global rule and the validation-fitted blocked-value, block-count, and protected-loss parity postprocessors. Within each split, family, unit, and capture floor, we keep policies whose validation loss capture is at least the requested fraction of the baseline validation capture, choose the smallest absolute validation complaint gap, and score that one policy on held-out test rows. Source-sign tests pool rows by independent source before assigning signs: freMPL1–freMPL9 count as one French portfolio, and Belgian MTPL is counted once across overlapping evidence families. The cluster test in the group-complaint selected-policy table uses the same source pooling but retains magnitudes: each source contributes its median held-out lift, and a one-sample tt test evaluates whether the mean source lift is positive. We report the two-sided tt p-value in the table, the directional p-value in text, and exact sign-flip and source-bootstrap checks because there are only seven source clusters.

Fixed-parity split. The pooled fixed-family comparison is not the headline claim because the family contains targets with different theoretical status. We therefore also compare the selected group-aware price with each fixed postprocessor separately, using the same validation-capture floor and the same held-out closure metric. At the .95 floor, raw value-unit closure, and materiality .01, the split is:

Fixed postprocessorMedian liftSource signsSign pp
Protected-loss parity+1.17386/0/00.0156
Block-count parity+0.16154/2/00.3438
Blocked-value parity+0.14915/1/00.1094

Table 1 (PDF p. 11).

These postprocessors constrain protected loss, block count, and blocked value, respectively. Their comparisons measure performance against those distinct targets. EU health remains in the seven-source group-aware-versus-group-blind table, but its fixed parity postprocessors do not clear the .95 validation-capture floor, so it is recorded in the eligibility table rather than in these paired rows.

Attainable-ceiling diagnostic. For each split, we compute the best held-out value-unit closure among group-aware rows satisfying the validation capture floor. This descriptive ceiling diagnoses available headroom and is not a performance estimate. At the .95 floor, Norwegian auto has median ceiling .6748 but median selected closure zero. For freMPL8, the corresponding medians are .3455 and .1450. This validation transfer can leave headroom even when complaint reduction is attainable within the evaluated bank. The artifact retains the cohort-level ceilings, selected closures, and eligibility rows.

Frozen cohorts and groups. The headline analysis retains seven source families. Each cohort contributes up to ten split cells, indexed by seeds 0–9, and infeasible cells are kept in the eligibility ledgers rather than silently paired. At the primary .95 floor, the 180 paired cells comprise 20 Australian, 30 Belgian, 90 French, and ten cells each for Bondora, EU health, Norwegian auto, and Swedish motorcycle. Australian and Belgian cached group-aware and group-blind cohorts are separate fitted-score experiments, not additional independent sources. French freMPL1–freMPL9 are nine datasets within one portfolio source. The public insurance cohorts use the recorded sex or gender field (sex for beMTPL97, Gender for freMPL1–freMPL9 and swmotorcycle, and Male for norauto); cached Australian, Belgian, and Bondora banks retain their recorded Female/Male labels. EU health uses gender, restricted to F/M. KuaiRand activity cohorts and constructed Australian stress cohorts are diagnostics outside this seven-source analysis. These cohort identifiers and their frozen row assignments are retained in the selection CSVs; overlapping cohorts are pooled at source level.

Data construction and splits. Fairness models use nested 80/20 splits followed by an 80/20 split of the training portion, giving 64/16/20 training/validation/test proportions; loss-occurrence stratification is used where the outcome permits it. These are separate experiments from score-learning tasks with 80/0/20 splits. For the public CASdatasets fairness cohorts, exposure is retained on no-claim rows and loss is the claim amount. EU-health loss sums non-negative amt_ fields and value is policy_years on no-claim rows; only positive policy-years and F/M group labels are retained. The group field is excluded from the public outcome-model predictors. Training rows determine numeric scaling and categorical encodings. The executable artifact records raw-file hashes, cohort filters, sample sizes, feature lists, split identities, and model parameters; these records distinguish fitted banks that share a source name.

Original policy-bank construction and selection. Policies are constructed from frozen fitted scores, without refitting models for each complaint price. The promoted public-auto and cached value-unit banks use a dense grid from 0 to .300 in steps of .005, followed by {.35, .4, .45, .5, .6, .75, 1, 1.5, 2, 3, 5, 10}. The lambda recorded on each frozen row indexes its actual available bank. Pricing uses the W-BCE-implied loss side b~=a^exp⁡(clip⁡(q,−50,50))\tilde{b} = \hat{a}\exp(\operatorname{clip}(q,-50,50)), where qq is the fitted W-BCE logit. Both price families share this loss side; their zero-price order reproduces fitted W-BCE up to numerical clipping. Separate raw-target value and loss regressions construct the harm and net-harm objection predicates and measurement weights. They do not replace the W-BCE loss side in this group-pricing comparison.

With z(x)=qc(x)u^w(x)z(x) = q_c(x)\widehat{u}_w(x), the common-price score is (b~−λz)/a^(\tilde{b} - \lambda z)/\hat{a} on its positive loss side. The group-aware score replaces λz\lambda z by sgλz/Rgs_g\lambda z/R_g, where sg=1s_g=1 for the validation high-complaint group and sg=−1s_g=-1 for the low-complaint group, and Rg=nval−1∑i:Gi=gu^w,iR_g=n_{\mathrm{val}}^{-1}\sum_{i:G_i=g}\hat{u}_{w,i}. This mass includes the group frequency required by the complaint-disparity postprocessor proposition. Zero price is included. Thus the families share the price axis; this is not a claim that they have identical expressive capacity.

Scores are sorted with row index as a deterministic tie-breaker. A validation threshold is frozen and applied to test scores. If the resulting validation blocked-value fraction vv exceeds β=.10\beta=.10, independent uniform thinning with probability α=min⁡(1,β/v)\alpha=\min(1,\beta/v) defines the evaluated randomized policy. Metrics are its expectations: blocked value, captured loss, and objection masses are multiplied by α\alpha. The same frozen α\alpha is used on test rows and applied separately to every family, including the baseline and fixed postprocessors. Test budgets are reported outcomes, not constraints used to select or adjust policies.

For ρ∈{.90,.95,.98,1}\rho\in\{.90,.95,.98,1\}, retain only validation-budget-feasible policies with capture at least ρ\rho times the adjusted baseline validation capture. Choose the smallest absolute validation complaint gap, breaking ties by higher validation capture, smaller price, and deterministic policy order. The fixed family searches its available parity postprocessors by the same eligibility and gap rules. A family with no eligible policy is recorded as ineligible; it does not enter a paired comparison. All comparisons use common eligible cohort/split cells and retain the baseline and family thinning factors.

Closure, aggregation, and uncertainty. Let d0d_0 be the unpriced baseline’s absolute test complaint gap and dPd_P the selected policy’s gap, using the same measurement unit and group denominators. Raw closure is CP=1−dP/d0C_P=1-d_P/d_0 for d0>0d_0>0, and comparison lift is Cgroup−CcomparatorC_{\mathrm{group}}-C_{\mathrm{comparator}}. Negative closure denotes a gap increase; raw values are not clipped. At d0=0d_0=0, the implemented convention gives closure one if dP=0d_P=0 and zero otherwise; such cells must be identifiable separately because relative improvement is undefined there. Clipped closure max⁡(0,min⁡(1,CP))\max(0,\min(1,C_P)) is a separately labeled diagnostic. The materiality threshold .01 classifies lift signs; it does not remove cells merely because their baseline gap is small.

Each source contributes its median paired lift across its retained cells. The headline cell median and the mean of seven source medians are distinct summaries. The source-cluster statistic is the one-sample tt statistic on those seven medians, with six degrees of freedom. Exact sign flips enumerate all 272^7 source-sign assignments. The source bootstrap resamples seven source medians with replacement 50,000 times and reports percentile endpoints; the implementation uses base seed 20261007 with deterministic analysis offsets. The artifact retains per-cell gaps, capture, budgets, eligibility, and source pooling, so these summaries can be recomputed without rerunning model fitting.

Preprocessing and Ranking

We prove the preprocessing proposition and give the exact condition under which weight damping preserves the raw W-BCE target.

Theorem B.1 (Outcome-side characterization). Let the outcome domain be the full cone Z={(B,A):B,A≥0, B+A>0}\mathcal{Z}=\{(B,A):B,A\ge0,\ B+A>0\}, and let U=W′Q′U=W'Q' and V=W′(1−Q′)V=W'(1-Q') be finite measurable transformations on this domain, where W′>0W'>0 and Q′∈[0,1]Q'\in[0,1]. Quantify over all finite-support distributions on Z\mathcal{Z} with positive means of AA and BB, and assume the transformed denominator has positive mean for each such distribution. The induced proper-loss ranking

E[U∣X=x]E[V∣X=x]\frac{E[U\mid X=x]}{E[V\mid X=x]}

is a fixed strictly decreasing function of Φ(x)\Phi(x) for all these distributions if and only if

[UV]=[p1p2r1r2]⏟M[BA],M≥0,det⁡(M)=p1r2−p2r1>0.\begin{bmatrix} U \\ V \end{bmatrix} = \underbrace{\begin{bmatrix} p_1 & p_2 \\ r_1 & r_2 \end{bmatrix}}_{M} \begin{bmatrix} B \\ A \end{bmatrix}, \qquad M\ge0,\qquad\det(M)=p_1r_2-p_2r_1>0.

Thus the universally valid reductions are exactly invertible non-negative linear combinations of the outcomes; W-BCE is the canonical (U,V)=(B,A)(U,V)=(B,A). The necessity claim uses the full outcome cone and arbitrary mixtures; preservation on a restricted dataset support need not imply this characterization. The same sufficiency holds for arbitrary distributions with finite outcome means and positive conditional means of AA and BB.

Proof. Let ρM(x)\rho_M(x) denote E[U∣X=x]/E[V∣X=x]E[U\mid X=x]/E[V\mid X=x]. Linearity of expectation gives

ρM(x)=p1+p2Φ(x)r1+r2Φ(x)\rho_M(x)=\frac{p_1+p_2\Phi(x)}{r_1+r_2\Phi(x)}

and dρM/dΦ=−(p1r2−p2r1)/(r1+r2Φ)2<0d\rho_M/d\Phi=-(p_1r_2-p_2r_1)/(r_1+r_2\Phi)^2<0, proving sufficiency. For necessity, let HH be the fixed decreasing function in the premise and fix a ratio tt. A conditional distribution satisfies E[A]/E[B]=tE[A]/E[B]=t exactly when E[A−tB]=0E[A-tB]=0. For the transformed outcomes, the same distribution must satisfy E[U−H(t)V]=0E[U-H(t)V]=0, because HH is one-to-one, the converse also holds.

Now consider two outcome points z1,z2z_1,z_2 for which Gt(z)=A(z)−tB(z)G_t(z)=A(z)-tB(z) has opposite signs. There is a unique mixture weight α\alpha with αGt(z1)+(1−α)Gt(z2)=0\alpha G_t(z_1)+(1-\alpha)G_t(z_2)=0. The same mixture must also make Ft(z)=U(z)−H(t)V(z)F_t(z)=U(z)-H(t)V(z) have zero expectation, so Ft(z1)/Gt(z1)=Ft(z2)/Gt(z2)F_t(z_1)/G_t(z_1)=F_t(z_2)/G_t(z_2). Since the two points were arbitrary, Ft=ctGtF_t=c_tG_t pointwise for some scalar ctc_t (points with Gt=0G_t=0 follow by mixing with any point off the hyperplane). Thus

U−H(t)V=ct(A−tB).U-H(t)V=c_t(A-tB).

Writing this identity for two distinct values of tt and solving the two linear equations shows that both UU and VV are linear combinations of BB and AA. Their nonnegativity gives M≥0M\ge0, and the strict decrease of HH gives det⁡(M)>0\det(M)>0. □\square

The first claim in the preprocessing proposition follows from the W-BCE theorem after replacing the original outcome sides by U=W′Q′U=W'Q' and V=W′(1−Q′)V=W'(1-Q'). The universal-preservation claim follows from Theorem B.1.

Proposition B.2 (Exact effect of weight damping). Suppose W>0W>0 and replace WW by g(W)>0g(W)>0 while retaining Q=B/WQ=B/W. Put h(W)=g(W)/Wh(W)=g(W)/W and assume both conditional outcome means are positive. The damped optimum has inverse ratio

R(x)=E[h(W)A∣X=x]E[h(W)B∣X=x].(14)R(x)=\frac{E[h(W)A\mid X=x]}{E[h(W)B\mid X=x]}. \tag*{(14)}

This equals Φ(x)\Phi(\mathbf{x}) exactly when

Cov⁡(h(W),AE[A∣x]−BE[B∣x] | X=x)=0.(15)\operatorname{Cov}\left(h(W),\frac{A}{E[A\mid\mathbf{x}]}-\frac{B}{E[B\mid\mathbf{x}]}\,\middle|\,X=\mathbf{x}\right)=0. \tag*{(15)}

Therefore only linear gg preserves the target for every conditional distribution.

Proof. The covariance term equals

E[h(W)A∣x]E[A∣x]−E[h(W)B∣x]E[B∣x].\frac{E[h(W)A\mid\mathbf{x}]}{E[A\mid\mathbf{x}]}-\frac{E[h(W)B\mid\mathbf{x}]}{E[B\mid\mathbf{x}]}.

Setting it to zero is equivalent to R(x)=Φ(x)R(\mathbf{x})=\Phi(\mathbf{x}). A nonconstant hh can be correlated with the centered bracket on a two-point distribution, so universal preservation forces hh to be constant. □\square

Prediction Domains and Frontier Regularity

Weighted proper-loss boundary cases. The population statement for W-BCE optimizes over the closed prediction interval [0,1][0,1], with each loss extended to its proper boundary values. Assume finite conditional means and 0<E[A+B∣X=x]<∞0 < E[A+B\mid X=\mathbf{x}] < \infty. Write q=E[B∣x]/E[A+B∣x]q=E[B\mid\mathbf{x}]/E[A+B\mid\mathbf{x}]. For 0<q<10<q<1, strict properness gives the unique minimizer qq. At q=0q=0 or q=1q=1, the optimum is the corresponding endpoint. For binary cross-entropy, use 0log⁡0=00\log0=0 in the weighted conditional risk. A finite-logit sigmoid takes values only in (0,1)(0,1): at an endpoint its risk has an infimum approached as the logit tends to −∞-\infty or +∞+\infty, rather than an attained finite-logit minimizer. A region with zero conditional total stake is indifferent. These population facts do not assert consistency of a restricted finite-sample model family.

Nonempty and closed attainable sets. For the frontier feasibility proposition, assume integrable nonnegative AA, BB, positive finite group denominators, and integrable coefficients in the complaint functional. Policies range over all measurable π(X,G)∈[0,1]\pi(X,G)\in[0,1], and the budget/capture feasible set is nonempty. A sufficient certificate for a fractional floor 0≤ρ≤10\le\rho\le1 is a baseline π0\pi_0 that satisfies the value budget, with η=ρE[π0B]\eta=\rho E[\pi_0B].

The order interval 0≤π≤10\le\pi\le1 is weak-* compact in L∞L^\infty. Integration against the stated L1L^1 coefficients is weak-* continuous, so budget and capture inequalities define a closed subset. Its linear complaint image is a nonempty compact interval. Hence if zero is outside that interval, its distance from zero is strictly positive and attained. If no policy meets the capture and budget requirements, the attainable set is empty; no policy or finite residual-disparity optimum is asserted. For a restricted finite candidate bank, the corresponding statement concerns its convex mixtures, not all measurable policies.

Duality and the supported policy form. For the complaint-disparity postprocessor proposition, retain the preceding integrability assumptions and fix the outcome proxies and objection indicator before optimization. In a finite empirical problem, randomized allocations form a finite-dimensional polytope. Feasibility and boundedness give linear programming strong duality and attained dual multipliers. In the population problem, additionally assume no duality gap and attainment by finite multipliers; a suitable constraint qualification, such as Slater feasibility for the inequality constraints after accounting for equality constraints, is sufficient.

A nonempty feasible set alone does not establish this population dual-attainment condition.

For a capture-maximizing supported point satisfying those conditions, let μ≥0\mu\ge0 be the budget multiplier and let the signed disparity multiplier be ν=α+−α−\nu=\alpha_+-\alpha_-. Conditioning the Lagrangian on (X,G)(X,G) gives coefficient

bg(x)−μag(x)−νκgc(x)uw(x,g)/Ugw,κ1=ν/P1,κ0=−ν/P0.b_g(\mathbf{x})-\mu a_g(\mathbf{x})-\nu\kappa_g c(\mathbf{x})u_w(\mathbf{x},g)/U_g^w,\qquad\kappa_1=\nu/P_1,\quad\kappa_0=-\nu/P_0.

An optimal policy blocks on its positive region, passes on its negative region, and randomizes on ties to satisfy active constraints. This establishes the stated policy form under dual attainment. A finite grid of prices followed by uniform thinning is a candidate-policy evaluation, not a proof that the selected grid policy attains this population optimum.

Account-Objection Validation Algorithm

The account-objection frontier figure uses ten paired seeds for each natural Australian MTPL, Belgian MTPL, and 300K KuaiRand cohort. Models are fit on training rows. Within a split, every policy uses the same validation and test accounts, observed value and loss, and frozen harm and net-harm criterion decisions. The latter are block indicators HH and NN; hence Fh=1−HF_h=1-H, Fn=1−NF_n=1-N, and the union objection is Fany=Fh∨FnF_{\mathrm{any}}=F_h\vee F_n. A block satisfying both objections contributes once to the plotted count-objection mass. The W-BCE reference is not an objecting criterion.

Let qq and qbq_b be fitted W-BCE and BCE logits, and let a^,b^\hat{a},\hat{b} be the nonnegative raw-target value and loss predictions. The three unpriced orders rank descending by qq, qbq_b, and b^/a^\hat{b}/\hat{a}, respectively. Negative fitted logits remain eligible. The plug-in union-price order ranks by (b^−λFany)/a^(\hat{b}-\lambda F_{\mathrm{any}})/\hat{a} on its positive loss side. A second union-price family replaces b^\hat{b} by a^exp⁡(q)\hat{a}\exp(q), so its zero-price order reproduces fitted W-BCE. The separate-price family uses (b^−λhFh−λnFn)/a^(\hat{b}-\lambda_hF_h-\lambda_nF_n)/\hat{a}; its penalty can count both criteria while the plotted union mass still counts each account once. The executable checks reproduce the unpriced masks and thinning factors at zero price and recover the stored BCE decisions.

Prices include the legacy absolute grids and validation-scaled values. For each loss-side family, positive predicted loss-side values among validation accounts with Fany=1F_{\mathrm{any}}=1 determine their median and the 90th, 95th, 99th percentiles and maximum. The grid includes those quantiles and the median times

{.01,.02,.05,.1,.2,.5,1,2,5,10,20,50}.\{.01,.02,.05,.1,.2,.5,1,2,5,10,20,50\}.

Separate harm and net-harm prices use pairs from the plug-in grid. Across seeds, summaries join the same relative price rule, retaining its seed-specific absolute prices in the policy bank.

For every family, a validation threshold targets 10% blocked value. Threshold ties can exceed that amount; uniform thinning with probability α=min⁡(1,.10/vval)\alpha=\min(1,.10/v_{\mathrm{val}}) corrects the resulting validation value fraction vvalv_{\mathrm{val}}. The threshold, absolute prices, and α\alpha are then frozen on test. Reported loss, value, and objection masses are expectations under thinning, and test value spend is recorded separately.

For policy PP in each test split, write LPL_P for loss capture and OPO_P for union objection counted divided by the number of accounts. The plot averages LP/L0L_P/L_0 and 1−OP/O01-O_P/O_0 over the ten paired splits, where zero denotes fitted W-BCE. All reference captures and objection shares are positive in these three cohorts. Intervals are paired Student tt intervals. Nondominance is computed after averaging; the frontier and its zero-objection endpoints are descriptive summaries of evaluated policies. They do not select a deployable policy using held-out outcomes. The group experiment uses validation-only family selection instead.

The data artifact scripts bandha_figure3_common_frontier.py and bandha_make_figure3_endpoint_table.py generate the policy bank, summaries, and endpoint table. Their provenance records source and candidate-input checksums. KuaiRand’s recovered BCE logits use the accepted raw-target runner: sample identities, splits, raw outcomes, and W-BCE logits match the earlier frozen bank; raw-target outcome proxies and their derived criterion decisions supersede its old proxy lineage.

Account-Level Endpoint Table

CohortPolicyL/L0L/L_01−O/O01 - O/O_0VV
Australian MTPLW-BCE1.000±0.0001.000 \pm0.0000.000±0.0000.000 \pm0.0000.100±0.0020.100 \pm0.002
Priced endpoint0.599±0.1250.599 \pm0.1251.000±0.0001.000 \pm0.0000.098±0.0040.098 \pm0.004
Belgian MTPLW-BCE1.000±0.0001.000 \pm0.0000.000±0.0000.000 \pm0.0000.101±0.0010.101 \pm0.001
Priced endpoint0.820±0.0740.820 \pm0.0741.000±0.0001.000 \pm0.0000.100±0.0020.100 \pm0.002
KuaiRandW-BCE1.000±0.0001.000 \pm0.0000.000±0.0000.000 \pm0.0000.101±0.0020.101 \pm0.002
Priced endpoint1.769±0.2581.769 \pm0.2581.000±0.0001.000 \pm0.0000.086±0.0040.086 \pm0.004

Table 5. Natural-cohort endpoints from the same corrected policy bank as the account-objection frontier figure. Entries are means ±\pm 95% paired-split intervals. L/L0L/L_0 is loss capture relative to fitted W-BCE, 1−O/O01 - O/O_0 is its count-objection share removed, and VV is held-out blocked value divided by total value. Each priced endpoint has zero objection in every test split and maximizes mean test capture among such evaluated policies. This endpoint selection is descriptive, not validation-selected performance. All families use the same validation cap and frozen thinning.

Group Disparity Decomposition

We give the denominator decomposition used in the main group-disparity section and in the empirical analysis.

Denote the main text’s group account-objection indicator by qC(x)q_C(x), so

OC(X,G;π)=π(X,G)qC(X).(16)O_C(X,G;\pi) = \pi(X,G)q_C(X). \tag*{(16)}

For groups with E[qC(X)Uw∣G=g]>0E[q_C(X)U_w \mid G=g] > 0, define

egw=E[qC(X)Uw∣G=g]E[Uw∣G=g],(17)e_g^w = \frac{E[q_C(X)U_w \mid G=g]}{E[U_w \mid G=g]}, \tag*{(17)}
Rgw(π)=E[π(X,G)qC(X)Uw∣G=g]E[qC(X)Uw∣G=g].(18)R_g^w(\pi) = \frac{E[\pi(X,G)q_C(X)U_w \mid G=g]}{E[q_C(X)U_w \mid G=g]}. \tag*{(18)}

Here egwe_g^w is fixed by the group distribution and objecting criteria, while Rgw(π)R_g^w(\pi) is chosen by the policy.

Proposition F.1 (Association–treatment decomposition). For each group with positive exposure,

Dgw(π)=egwRgw(π).D_g^w(\pi) = e_g^w R_g^w(\pi).

For two such groups, if treatment parity holds with R1w(π)=R0w(π)=rR_1^w(\pi) = R_0^w(\pi) = r, then

D1w(π)−D0w(π)=r(e1w−e0w).D_1^w(\pi) - D_0^w(\pi) = r(e_1^w - e_0^w).

If the treatment rates are positive, exact complaint parity D1w(π)=D0w(π)D_1^w(\pi) = D_0^w(\pi) is equivalent to

R1w(π)R0w(π)=e0we1w.\frac{R_1^w(\pi)}{R_0^w(\pi)} = \frac{e_0^w}{e_1^w}.

Proof. For each group with positive exposure,

egwRgw(π)=E[qC(X)Uw∣G=g]E[Uw∣G=g]⋅E[π(X,G)qC(X)Uw∣G=g]E[qC(X)Uw∣G=g]=E[π(X,G)qC(X)Uw∣G=g]E[Uw∣G=g]=Dgw(π),\begin{aligned} e_g^w R_g^w(\pi) = \frac{E[q_C(X)U_w \mid G=g]}{E[U_w \mid G=g]} \cdot \frac{E[\pi(X,G)q_C(X)U_w \mid G=g]}{E[q_C(X)U_w \mid G=g]} \\ &= \frac{E[\pi(X,G)q_C(X)U_w \mid G=g]}{E[U_w \mid G=g]} = D_g^w(\pi), \end{aligned}

using (16). If R1w(π)=R0w(π)=rR_1^w(\pi) = R_0^w(\pi) = r, then

D1w(π)−D0w(π)=e1wR1w(π)−e0wR0w(π)=r(e1w−e0w).\begin{aligned} D_1^w(\pi) - D_0^w(\pi) = e_1^w R_1^w(\pi) - e_0^w R_0^w(\pi) \\ &= r(e_1^w - e_0^w). \end{aligned}

The last claim follows by rearranging e1wR1w(π)=e0wR0w(π)e_1^w R_1^w(\pi) = e_0^w R_0^w(\pi). □

Normalized Price and Source Sensitivity

The original price bank used the same numerical axis for penalties expressed in different units. A common-price policy subtracted λz\lambda z, whereas a group-price policy subtracted sgλz/Rgs_g\lambda z/R_g. Consequently, matching numerical prices did not match their effective score shifts. Validation selections at the largest available price also indicated truncated coverage. We therefore report a post-review sensitivity on the same frozen candidate logs. This analysis is not a new, independently collected test evaluation.

Empirical score construction. For the empirical scores, let a^\hat a be the nonnegative value proxy and b~=a^exp⁡(clip⁡(sθ,−50,50))\tilde b = \hat a \exp(\operatorname{clip}(s_\theta,-50,50)) the fitted W-BCE logit’s implied loss side. The auxiliary loss regression constructs objections and loss-unit weights; it does not replace b~\tilde b. Let u^w\hat u_w be 1, a^\hat a, or the nonnegative auxiliary loss proxy for count, value, or loss units, respectively. Here λ≥0\lambda\ge0 is a scalar complaint price, and nvaln_{\mathrm{val}} is the number of validation accounts. With z=qCu^wz=q_C\hat u_w, define Rg=nval−1∑i:Gi=gu^w,iR_g=n_{\mathrm{val}}^{-1}\sum_{i:G_i=g}\hat u_{w,i}, and Rall=R0+R1R_{\mathrm{all}}=R_0+R_1. The normalized scores are

sblind=b~−λz/Ralla^,sgroup=b~−sgλz/Rga^.s_{\mathrm{blind}} = \frac{\tilde b-\lambda z/R_{\mathrm{all}}}{\hat a}, \qquad s_{\mathrm{group}} = \frac{\tilde b-s_g\lambda z/R_g}{\hat a}.

where sg=1s_g=1 for the validation high-complaint group and −1-1 for the low group. Positive adjusted loss is eligible; zero predicted value retains the original score convention. These are plug-in policies, with finite-model and clipping error, rather than guarantees of population-optimal capture.

Common price units and preservation. Write z=qCu^wz=q_C\hat u_w, with the estimated count, value, or loss weight used by the existing experiment, and Rall=nval−1∑iu^w,i=R0+R1R_{\mathrm{all}}=n_{\mathrm{val}}^{-1}\sum_i\hat u_{w,i}=R_0+R_1. The normalized common-price policy scores (b~−λz/Rall)/a^(\tilde b-\lambda z/R_{\mathrm{all}})/\hat a; the group-price policy scores (b~−sgλz/Rg)/a^(\tilde b-s_g\lambda z/R_g)/\hat a. Only positive adjusted loss is eligible for blocking, with the existing zero-predicted-value convention retained. Normalization puts the scalar multiplier in comparable aggregate units; it does not make the two families’ pointwise score shifts equal or make the families nested.

Figure 5. Denominator choice for group complaint disparity. The orange outline marks group 1, as in Figure 1. Panel A shows the same global value-per-loss threshold applied to both groups. Panel B aggregates the resulting objections in value units. DgAD_g^A counts blocked objection-bearing value and divides by all value in group gg. Dividing instead by all objection-bearing value would report only the conditional blocking rate among accounts that could object.

For every cohort, seed, and measurement unit, let G\mathcal{G} be the original scalar grid. Both families search the same union G∪RallG\mathcal{G} \cup R_{\mathrm{all}}\mathcal{G}. Original group policies occur at λ∈G\lambda\in\mathcal{G}, and original common-price physical policies occur at λ=Rallλold\lambda= R_{\mathrm{all}}\lambda_{\mathrm{old}}. Thus the corrected candidate sets retain both original families’ policies.

Validation-only coverage and freezing. Starting from the union’s maximum, we add geometric doublings to both families. We continue until no selected policy at any of the four capture floors is at the largest price and two further extensions leave all selected validation policies unchanged. This is a finite coverage diagnostic, not an optimality guarantee over continuous prices. All thresholds, thinning probabilities, selected policies, and LP mixture weights are written to a validation freeze before test-policy evaluation. Predictors, outcome proxies, objection predicates, splits, the .10 value cap, and deterministic ties are unchanged. The LP uses the full enlarged common-price bank, minimizing validation absolute signed gap, then maximizing capture and minimizing spend within numerical tolerances; its original constraints are checked after solution.

Source scope and aggregation. The seven insurance-dominated sources remain one reported scope. The extended scope includes all recorded applicable binary-group sources, adding CMS provider-fraud, UCI credit, and German credit cohorts. Multiple grouping attributes or fitted cohorts from one source are pooled before source-level inference. Neither negative nor zero effects are removed. Count and loss units are reported alongside value units. Source medians and source means are separate estimands; each averages one summary per source. Frozen input hashes, coverage traces, eligibility, selected component group masses, and all source-level results are available in the public reproduction artifact.

Additional aggregation checks. At the .95 floor, the seven-source cell median lift is +.2453 and its mean of source medians is +.2185 (two-sided p=.3071p = .3071). Across all ten recorded sources, these are +.0803 and +.0863 (p=.6007p = .6007). Neither comparison establishes a statistically significant average group-aware advantage. The seven-source mean of source means is +.0977 (p=.6997p = .6997); the ten-source mean is −.3979 (p=.3097p = .3097). Raw cell means are −.2274 and −.5592, respectively. These summaries measure different effects and retain the negative cells rather than clipping them. French and Norwegian source medians are negative in the seven-source scope; CMS and UCI credit also have negative medians in the extended scope. The seven-source exact sign-flip directional pp is .1094. An independent run with a common global axis extended through 512000 gives ten-source mean source-median lift +.0202 (p=.8990p = .8990), likewise not establishing an average advantage.

Independent global-grid sensitivity. A separate replay extends a common global base axis to 512000 before applying the same original-policy preservation mapping to every cell. All 240 value-unit cells are eligible at each of the four capture floors; selections are unchanged at maxima 128000, 256000, and 512000. At the .95 floor, the ten-source mean of source medians is +.0202 (two-sided source tt-test p=.8990p = .8990); its mean of source means is −.6209 (p=.1207p = .1207). This wider bank is a separate sensitivity, with the same frozen predictors and units. It supports the same absence of a significant average advantage; the LP table uses the per-cell bank defined above.

Historical comparison. Before normalization and extension, the .95 seven-source comparison had cell median lift +.5447, mean source-median lift +.4458, and two-sided source tt-test p=.002938p = .002938. Its mean of source means was +.3014 with p=.1482p = .1482; the cell mean was −.0085. Adding all ten recorded sources to that historical bank gave mean source-median lift +.2444 with p=.1009p = .1009. These historical numbers describe the original finite bank, rather than the corrected comparison reported in the main text.

SourceGroup gapLP gapLift
Australian MTPL0.00430.0132+0.3291
Belgian MTPL0.00190.0055+0.1903
Bondora0.02760.0508+0.4322
CMS provider fraud0.02640.0093−0.7152
EU health0.00260.0068+0.3010
French portfolio0.00790.0060−0.0373
German credit0.05090.0323−0.2659
Norwegian auto0.00970.0014−0.9302
Swedish motorcycle0.00580.0128+0.3428
UCI credit0.00870.0070−0.1684

Table 6. Group-aware prices versus linear-program (LP) mixtures of the group-blind account-price bank at the .95 capture floor. Gaps and lift Cgroup−CLPC_{\mathrm{group}}-C_{\mathrm{LP}} are separately computed source medians over held-out cells; CLPC_{\mathrm{LP}} is the mixture complaint closure and positive lift favors group pricing. Both families use the same validation constraints and the enlarged normalized price axis. All ten recorded sources are retained.

LP comparator and source-level results. We also compare with a linear program (LP) over the full frozen group-blind account-price bank; CLPC_{\mathrm{LP}} denotes its mixture’s closure. The LP chooses nonnegative mixture weights summing to one, minimizing the absolute signed validation complaint gap subject to the same capture floor and expected blocked-value cap. Among gap-optimal mixtures it maximizes capture, then minimizes spend; each stage is checked against the original constraints. Component thinning and mixture weights are frozen on test, where we report expected outcomes. The comparator optimizes the available policy bank, not all possible candidate-level decisions.

UnitFloorSeven mediansSeven meansTen mediansTen means
Value0.90+0.0391-0.1533-0.0372-0.5979
Value0.95+0.2185+0.0977+0.0863-0.3979
Value0.98+0.1969+0.1127+0.0735-0.3605
Value1.00+0.1093+0.0158+0.0472-0.4358
Count0.90-0.0514-0.2949-0.0699-0.2966
Count0.95+0.0995-0.2123+0.0321-0.2080
Count0.98+0.0580-0.2176+0.0406-0.1419
Count1.00+0.0892-0.1877+0.0624-0.1058
Loss0.90-0.1702-0.6889-0.2130-0.9735
Loss0.95-0.0375-0.4740-0.1121-0.7918
Loss0.98-0.0748-0.3661-0.0758-0.6242
Loss1.00-0.1367-0.4901-0.1021-0.3450

Table 7. Corrected closure-lift sensitivity across measurement units and capture floors. Each cell averages one source summary: medians or means. The seven-source and full ten-source scopes are distinct. Negative results are retained; all source test statistics and exact sign-flip results are in the frozen artifact.

References

  1. [1]Alekh Agrawal, Alina Beygelzimer, Miroslav Dudík, John Langford, and Hanna Wallach. 2018. A Reductions Approach to Fair Classification. In Proceedings of the 35th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 80). PMLR, Stockholm, Sweden, 60–69.
  2. [2]Reuben Binns. 2018. Fairness in Machine Learning: Lessons from Political Philosophy. In Conference on Fairness, Accountability and Transparency (Proceedings of Machine Learning Research, Vol. 81). https://proceedings.mlr.press/v81/binns18a.htmlarxiv.org/abs/1712.03586
  3. [3]Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. 2007. Learning to Rank: From Pairwise Approach to Listwise Approach. In International Conference on Machine Learning. Association for Computing Machinery, New York, NY, USA, 129–136.
  4. [4]Alexandra Chouldechova. 2017. Fair Prediction with Disparate Impact: A Study of Bias in Recidivism Prediction Instruments. Big Data 5, 2 (2017), 153–163. doi:10.1089/big.2016.0047DOI
  5. [5]Yves Clémençon, Gábor Lugosi, and Nicolas Vayatis. 2008. Ranking and Empirical Minimization of U-statistics. The Annals of Statistics 36, 2 (2008), 844–874. doi:10.1214/009053607000000910
  6. [6]Sam Corbett-Davies, Emma Pierson, Avi Feller, Sharad Goel, and Aziz Huq. 2017. Algorithmic Decision Making and the Cost of Fairness. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Association for Computing Machinery, New York, NY, USA, 797–806. doi:10.1145/3097983.3098095
  7. [7]George B. Dantzig. 1957. Discrete-Variable Extremum Problems. Operations Research 5, 2 (1957), 266–288.DOI
  8. [8]Simon De Vos, Jente Van Belle, Andres Algaba, Wouter Verbeke, and Sam Ver-boven. 2025. Decision-centric fairness: Evaluation and optimization for resource allocation problems. arXiv:2504.20642. https://arxiv.org/abs/2504.20642
  9. [9]Christophe Dutang and Arthur Charpentier. 2026. CASdatasets: Insurance Datasets. doi:10.57745/P0KHAG R package version 1.2-1.
  10. [10]Fannie Mae. 2026. Fannie Mae Single-Family Loan Performance Data. HARP loan performance public archive. https://capitalmarkets.fanniemae.com/credit-risk-transfer/single-family-credit-risk-transfer/fannie-mae-single-family-loan-performance-data Accessed 2026-08-24.
  11. [11]Jake Fawkes, Nic Fishman, Mel Andrews, and Zachary C. Lipton. 2024. The Fragility of Fairness: Causal Sensitivity Analysis for Fair Machine Learning. arXiv:2410.09600 https://arxiv.org/abs/2410.09600
  12. [12]Freddie Mac. 2026. Freddie Mac CRT Loan-Level Data. CRT Clarity STACR loan-level public archive. https://crt.freddiemac.com/tools-data/ctr-data Accessed 2026-08-24.
  13. [13]Rafael Frongillo and Ian A. Kash. 2015. Vector-Valued Property Elicitation. In Proceedings of the 28th Conference on Learning Theory (Proceedings of Machine Learning Research, Vol. 40). PMLR, Paris, France, 710–727.
  14. [14]Chongming Gao, Shijun Li, Yuan Zhang, Jiawei Chen, Biao Li, Wenqiang Lei, Peng Jiang, and Xianghan He. 2022. QuaiRand: An Unbiased Sequential Recommendation Dataset with Randomly Exposed Videos. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. Association for Computing Machinery, New York, NY, USA, 3953–3957. doi:10.1145/3511808.3557624DOI
  15. [15]Nathan George. 2019. All Lending Club Loan Data. Kaggle dataset. https://www.kaggle.com/datasets/wordsforthewise/lending-club Updated 2019-04-10; accessed 2026-08-24.
  16. [16]Tilmann Gneiting. 2011. Making and Evaluating Point Forecasts. J. Amer. Statist. Assoc. 106, 494 (2011), 746–762. doi:10.1198/jasa.2011.r10138DOI
  17. [17]Moritz Hardt, Eric Price, and Nati Srebro. 2016. Equality of Opportunity in Supervised Learning. In Advances in Neural Information Processing Systems, Vol. 29. Curran Associates, Inc., Red Hook, NY, USA, 3315–3323.arxiv.org/abs/1610.02413
  18. [18]Sebastian Höppner, Bart Baesens, Wouter Verbeke, and Tim Verdonck. 2020. Instance-Dependent Cost-Sensitive Learning for Detecting Transfer Fraud. arXiv:2005.02488 https://arxiv.org/abs/2005.02488
  19. [19]Thorsten Joachims, Adith Swaminathan, and Tobias Schnabel. 2017. Unbiased Learning-to-Rank with Biased Feedback. In Proceedings of the Tenth ACM International Conference on Web Search and Data Mining. Association for Computing Machinery, New York, NY, USA, 781–789. doi:10.1145/3018661.3018699DOI
  20. [20]Georgy Kalashnov, Evan Munro, Hao Sun, Shuyang Du, and Stefan Wager. 2021. Treatment Allocation under Uncertainty Costs. arXiv:2103.11066 [stat.ME] doi:10.48550/arXiv.2103.11066
  21. [21]Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan. 2017. Inherent Trade-Offs in the Fair Determination of Risk Scores. In 8th Innovations in Theoretical Computer Science Conference (Leibniz International Proceedings in Informatics, Vol. 67). Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik, Dagstuhl, Germany, 43:1–43:23. doi:10.4230/LIPIcs.ITCS.2017.43DOI
  22. [22]Nicolas O. Lorenz, Medhos Methods of Measuring the Concentration of Wealth. Publications of the American Statistical Association 9, 70 (1905), 209–219.
  23. [23]Max O. Lorenz. 1905. Methods of Measuring the Concentration of Wealth. Publications of the American Statistical Association 9, 70 (1905), 209–219.DOI
  24. [24]Henrietta Lyons, Eduardo Velloso, and Tim Miller. 2021. Conceptualising Contestability: Perspectives on Contesting Algorithmic Decisions. arXiv:2103.01774 https://arxiv.org/abs/2103.01774
  25. [25]Charles F. Manski. 1990. Nonparametric Bounds on Treatment Effects. American Economic Review 80, 2 (1990), 319–323.
  26. [26]Aditya Krishna Menon and Robert C. Williamson. 2018. The Cost of Fairness in Binary Classification. In Proceedings of the 1st Conference on Fairness, Accountability and Transparency (Proceedings of Machine Learning Research, Vol. 81). PMLR, New York, NY, USA, 107–118.
  27. [27]Jerzy Neyman and Egon S. Pearson. 1933. On the Problem of the Most Efficient Tests of Statistical Hypotheses. Philosophical Transactions of the Royal Society of London. Series A 231 (1933), 289–337.
  28. [29]Hongxiang Qiu, Marco Carone, and Alex Luedtke. 2022. Individualized Treatment Rules under Stochastic Treatment Cost Constraints. Journal of Causal Inference 10, 1 (2022), 480–493. doi:10.1515/jci-2022-0005DOI
  29. [30]Mark D. Reid and Robert C. Williamson. 2010. Composite Binary Losses. Journal of Machine Learning Research 11 (2010), 2387–2422.arxiv.org/abs/0912.3301
  30. [31]Leonard J. Savage. 1971. Elicitation of Personal Probabilities and Expectations. J. Amer. Statist. Assoc. 66, 336 (1971), 783–801. doi:10.1080/01621459.1971.10482346DOI
  31. [32]Manu Siddhartha. 2021. Bondora Peer to Peer Lending Loan Data. Kaggle dataset. https://www.kaggle.com/datasets/sid321axn/bondora-peer-to-peer-lending-loan-data Updated 2021-07-23; accessed 2026-08-24.
  32. [33]Ashudeep Singh and Thorsten Joachims. 2018. Fairness of Exposure in Rankings. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. Association for Computing Machinery, New York, NY, USA, 2219–2228. doi:10.1145/3219819.3220088DOI
  33. [34]Shubham Singh, Ian A. Kash, and Mesrob I. Ohannessian. 2026. Scoring Is Not Enough: Addressing Gaps in Utility-Fairness Trade-Offs for Ranking. arXiv:2606.26369 [cs.IR] doi:10.48550/arXiv.2606.26369
  34. [35]U.S. Small Business Administration. 2026. SBA 7(a) and 504 FOIA Data. SBA Open Data, quarterly FOIA export. https://data.sba.gov/dataset/7a-504-foia Accessed 2026-08-24.
  35. [36]Fen Xia, Tie-Yan Liu, Jue Wang, Wensheng Zhang, and Hang Li. 2008. Listwise Approach to Learning to Rank: Theory and Algorithm. In International Conference on Machine Learning. Association for Computing Machinery, New York, NY, USA, 1192–1199.

Paper details

vlnc:W_pmvqz89a
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents