betaFeedback

WorksComputer Science

Robust Aggregation of Quantile Statistics: from Median to Interval

Abstract

Quantile-based summaries, such as prediction intervals and medians, are commonly used in economic forecasting, climate science, and public health. For highly skewed samples, experimental evidence suggests that estimates of the mean can be biased toward the median, whereas estimates of the median are relatively accurate. This motivates aggregating explicitly elicited quantiles rather than reported means. We study robust aggregation of quantile reports for bounded outcomes under squared loss, minimizing worst-case regret relative to an omniscient Bayesian aggregator who knows the information structure and observes all private signals. Without restrictions on the information structure, no rule improves on the constant forecast at the range midpoint in worst-case regret. We therefore consider weak substitutes, a diminishing-returns condition on the value of information, and characterize optimal aggregation rules under this condition. For central prediction intervals with common nominal coverage αα, the optimal rule, which we call regularized simple average for prediction intervals, shrinks the average reported interval midpoint toward the range midpoint, with a weight determined by αα and the number nn of experts. This optimal rule strictly improves on the constant baseline in worst-case regret when α<1/nα<1/\sqrt{n}; at or above this threshold, the constant rule is optimal. The optimal worst-case regret is minimized at a unique interior coverage level, revealing a nonmonotonic relationship between coverage and aggregation performance. For median reports, regularized simple average for medians similarly shrinks the average reported median toward the range midpoint, assigning weight 1/n1/n to the reported average and strictly improving on the constant baseline in worst-case regret for every finite n≥2n\ge2.

Introduction

Suppose you wish to forecast how much U.S. real GDP will grow next year. You consult several experts, each of whom is uncertain about the outcome and communicates this uncertainty through lower and upper forecasts. In the fourth-quarter 2018 Duke CFO survey, for example, respondents were asked for “worst-case” and “best-case” GDP growth estimates associated with a ten-percent chance in each tail [11] (Figure 1). The two tail estimates define a central prediction interval with nominal coverage α=0.8\alpha= 0.8 (the 10th and 90th percentiles). If only these interval reports are available, how should we combine them into a single point forecast? And would asking for narrower or wider intervals lead to a more reliable aggregate? These questions are particularly relevant to online platforms that elicit and aggregate quantile forecasts.

GDP-growth question from the fourth-quarter 2018 Duke CFO survey

Figure 1. The GDP-growth question from the fourth-quarter 2018 Duke CFO survey [11]. The highlighted ten-percent tail probabilities define the 10th and 90th percentiles, forming a central eighty-percent prediction interval.

Quantile-based summaries, including prediction intervals and medians, are natural ways to communicate subjective uncertainty. They are used in finance [5], climate science [4], public health [6], and energy forecasting [19]. Behavioral evidence provides an additional motivation for studying quantiles directly. As reviewed by Garthwaite et al. [14], experimental evidence on skewed samples suggests that participants’ estimates of sample means were biased toward the corresponding sample medians, whereas their estimates of sample medians were relatively accurate.

However, quantile reports might not contain enough information to yield a point forecast that outperforms even the best constant predictor. Three challenges arise. First, quantile reports can obscure information relevant to prediction under squared loss, for which the posterior mean is optimal. For a binary outcome, for example, posterior probabilities of 0.6 and 0.9 both imply a median report of 1, despite corresponding to different posterior means. Second, the information structure—the joint distribution of the outcome and the experts’ private signals—is unknown. Without this knowledge, we can neither apply the Bayesian combiner nor determine whether experts’ information is redundant or complementary, leaving it unclear how their reports should be combined. Third, when eliciting prediction intervals, the choice of coverage level determines which quantiles are reported and thus what information is available for prediction; how to choose this level is itself an open question.

We study the above challenges under a robust aggregation framework. We consider a bounded outcome, normalized to [0,1][0, 1], and nn experts who each observe a private signal. Each expert truthfully reports a central posterior prediction interval at a common nominal coverage level α∈(0,1)\alpha\in(0, 1). The aggregator observes these reports and knows the reporting format, but not the underlying information structure. Its task is to produce a point forecast evaluated by squared loss. We measure its performance by regret relative to an omniscient Bayesian aggregator who knows the information structure and observes all private signals, and therefore predicts the full-signal posterior mean. We seek an aggregation rule that minimizes the worst-case regret over a class of possible information structures, following the robust forecast aggregation framework [1, 25].

Existing work on robust forecast aggregation has primarily considered posterior probability reports [1, 18, 22] or posterior mean reports [25]. However, these results do not directly determine how to aggregate prediction intervals or medians, nor do they answer which interval coverage level should be elicited.

As in these prior settings, obtaining a nontrivial worst-case guarantee requires restrictions on the information structure. When arbitrary dependence among private signals is allowed, no aggregation rule based on quantile reports can achieve lower worst-case regret than the constant forecast 1/21/2. This is because the general case allows the XOR construction (Example 3.1) which makes each expert’s posterior distribution given their own signal—and hence every quantile report—uninformative even though the signals jointly reveal the outcome.

We focus on information structures satisfying weak substitutes, a diminishing-returns condition on the value of information introduced by Chen and Waggoner [8] and used in forecast aggregation by Neyman and Roughgarden [25]. Informally, this condition implies that the improvement in expected squared loss from observing an additional signal cannot increase as more signals become available. This assumption rules out the pure XOR example while permitting experts’ information to overlap. Under weak substitutes, we derive explicit expressions for the optimal aggregators of interval and median reports, and identify the best interval coverage levels at which the optimal worst-case regret achieves its minimum.

Summary of Results

Our main findings are as follows:

  • Optimal aggregation of prediction intervals. We characterize an optimal rule, which we call regularized simple average for prediction intervals (Theorem 4.1). It combines the average of the reported interval midpoints with the midpoint of the outcome range, placing a weight on the reports that depends only on the number of experts nn and the nominal coverage level α\alpha. Figure 2 compares its worst-case regret with simple averaging of interval midpoints (Corollary 4.3) and with the optimal worst-case regret attainable from posterior-mean reports (Proposition C.1).

Worst-case regret under weak substitutes, showing means: SA, medians: RSA, intervals: SA, intervals: RSA, constant 1/2, and medians: SA as a function of number of experts $n$

Figure 2. Worst-case regret under weak substitutes. SA is robust optimal for posterior-mean reports, and RSA for intervals and medians. Interval bands span the infimum and supremum over 0<α<10 < \alpha< 1; SA averages interval midpoints. Median SA coincides with the upper orange boundary. The constant 1/2 has worst-case regret 0.25.

  • A phase transition and an optimal interval reporting level. We show that regularized simple average strictly improves upon the optimal constant forecast in worst-case regret if α<1/n\alpha< 1/\sqrt{n}, whereas the constant forecast is robust optimal when α≥1/n\alpha\ge1/\sqrt{n}. Moreover, the optimal worst-case regret is minimized at a unique interior coverage level α∗(n)\alpha^*(n), characterized by n(α∗(n))2(3+2α∗(n))=1n(\alpha^*(n))^2(3 + 2\alpha^*(n)) = 1, strictly below the phase threshold. Both lower and higher reporting levels yield worse guarantees. In particular, for bounded outcomes, the eighty-percent coverage level used in the motivating survey lies in the constant-optimal regime for any n≥2n \ge2. Figure 3 illustrates the threshold and the optimal reporting level as the number of experts varies (Theorem 4.2).

Phase threshold and optimal reporting level as functions of the number of experts

Figure 3. The phase threshold αthre(n)=1/n\alpha_{\mathrm{thre}}(n) = 1/\sqrt{n} and the regret-minimizing interval reporting level α∗(n)\alpha^*(n) both decrease with nn. Pale blue allows strict improvement over the constant rule in worst-case regret; pale orange marks levels where the constant is robust optimal, including the threshold. The regret-minimizing level lies strictly between zero and the threshold.

  • Optimal aggregation of medians. For median reports, regularized simple average for medians is also optimal. It combines the average reported median with the midpoint of the outcome range, assigning weight 1/n1/n to the reported average and the remaining weight to the constant midpoint. This rule strictly improves upon the optimal constant forecast in worst-case regret for every finite n≥2n \ge2. Its optimal weight and worst-case regret agree with the corresponding limits of the interval results as α↓0\alpha\downarrow0, although central

prediction intervals need not shrink to a single point when the posterior has multiple medians (Theorem 5.1). Figure 2 compares its worst-case regret with that of simple averaging of medians (Corollary 5.2).

Together, these results show that robust aggregation involves not only how to combine experts’ quantile reports, but also which quantiles to ask them to report.

Related Work

Aggregation of quantile statistics. Quantile-based summaries, including prediction intervals and medians, are natural ways to communicate subjective uncertainty. Proper quantile and interval scoring rules provide truthful mechanisms for eliciting such statistics [6, 17]. Recent studies adopt such mechanisms and investigate how to aggregate the elicited quantile statistics and interval forecasts [17, 29]. We focus on aggregation in this paper, and for simplicity, assume truthful reports without specifying an elicitation mechanism.

Given the experts’ quantile reports, existing studies either aggregate them into a point estimate [24, 29], or aggregate them into distributional summaries represented by quantiles, intervals, or full distributions [13, 15, 16, 17, 20, 23, 28]. Regarding the form of aggregation rules, existing studies combine reported values or corresponding quantiles through weighted average [15, 17, 24, 29] or use linear combinations of quantile forecasts with coefficients estimated from historical data [16, 28]. Other approaches pool predictive distributions through pointwise averages or medians of cumulative distribution functions [13, 20, 23] or logarithmic pooling of densities [7]. We characterize a robust optimal point prediction rule, regularized simple average, that shrinks the average of interval midpoints or reported medians toward the fixed midpoint of the outcome range, with a weight given explicitly in terms of the number of experts and, for interval reports, the nominal coverage level.

For theoretical guarantees, existing theoretical work characterizes when quantile aggregation improves on random selection under proper scoring rules [17] and assesses gains from combining forecasts through encompassing tests [16]. Other work derives maximum-likelihood aggregation rules under specific signal models [29]. In this paper, we characterize the robust optimal aggregation rule with minimum worst-case regret relative to the omniscient aggregator under weak substitutes [8, 25], optimizing over all possible aggregation rules.

Robust Aggregation Frameworks. The robust aggregation framework, introduced by Arieli et al. [1], seeks rules with worst-case performance guarantees over a specified class of information structures. In their work, they consider a binary outcome scenario where two Bayesian experts truthfully report their posterior beliefs.

Subsequent work varies the assumptions on information structures and the information available to the aggregator. Several studies assume that experts’ signals are conditionally independent [2, 9, 18]. Other approaches study unknown dependence with known individual information structures [3, 10, 22]. Neyman and Roughgarden Neyman and Roughgarden [25] study posterior-mean reports for general real-valued outcomes under weak substitutes and projective substitutes.

Another line of work varies the information elicited from experts. Second-order reports can improve aggregation [21, 26], while Frongillo et al. [12] develop a general framework of elicitable queries and relate query complexity to aggregation accuracy. To our knowledge, our work is the first to characterize robust optimal aggregation of prediction intervals and medians for general bounded outcomes under squared loss and weak substitutes.

Model

Fix an integer n≥2n \ge2 and a nominal coverage level α∈(0,1)\alpha\in(0,1). Write β=(1−α)/2\beta= (1-\alpha)/2 and [n]={1,…,n}[n] = \{1,\ldots,n\}. The outcome is a random variable YY taking values in Y=[0,1]\mathcal{Y} = [0,1]. Each expert ii observes a private signal SiS_i taking values in Si\mathcal{S}_i. We use lowercase letters yy and sis_i for their realizations, and write S=(S1,…,Sn)S = (S_1,\ldots,S_n) and s=(s1,…,sn)s = (s_1,\ldots,s_n). An information structure is a joint distribution θ∈Δ(Y×S1×⋯×Sn)\theta\in\Delta(\mathcal{Y} \times\mathcal{S}_1 \times\cdots\times\mathcal{S}_n) of the outcome and signals. Expert ii forms a Bayesian posterior and truthfully reports the central prediction interval

Zi=[Li,Ui]=[Qθ,i(β∣Si),Qθ,i(1−β∣Si)],Z_i = [L_i,U_i] = [Q_{\theta,i}(\beta\mid S_i),Q_{\theta,i}(1-\beta\mid S_i)],
Qθ,i(p∣si)=inf⁡{y∈[0,1]:Pθ(Y≤y∣Si=si)≥p}.(1)Q_{\theta,i}(p \mid s_i) = \inf\{y \in[0,1] : \mathbb{P}_{\theta}(Y \le y \mid S_i = s_i) \ge p\}. \tag*{(1)}

We use this lower generalized quantile convention throughout. The two quantile levels assign the same nominal tail probability β\beta on each side. The resulting interval has posterior coverage at least α\alpha, with possibly larger coverage when the posterior has atoms. A realized interval report is denoted by zi=[ℓi,ui]z_i = [\ell_i,u_i].

For median reports, we set Zi=Qθ,i(1/2∣Si)Z_i = Q_{\theta,i}(1/2 \mid S_i), using the same symbol ZiZ_i for the random report and ziz_i for its realization. This replaces the two quantile levels by their common limit 1/21/2 as α↓0\alpha\downarrow0. Figure 4 illustrates both reporting formats.

Posterior density given private signal, showing interval and median reports

Figure 4. Quantile reports summarize an expert’s posterior belief. For this continuous example, [ℓi,ui]=[Qθ,i(β∣Si=si),Qθ,i(1−β∣Si=si)][\ell_i,u_i] = [Q_{\theta,i}(\beta\mid S_i=s_i), Q_{\theta,i}(1-\beta\mid S_i=s_i)] covers the central α\alpha probability, and the dashed green line is the median Qθ,i(1/2∣Si=si)Q_{\theta,i}(1/2 \mid S_i=s_i). With atoms, the interval’s coverage can exceed α\alpha.

The aggregator knows nn and the reporting format, including α\alpha for interval reports, but not the information structure θ\theta or the private signals. After observing the report vector Z=(Z1,…,Zn)Z = (Z_1,\ldots,Z_n), it gives a single point prediction of the outcome. Let Z\mathcal{Z} denote the space of a single expert’s report under the chosen reporting format. An aggregation rule is a measurable map f:Zn→[0,1]f : \mathcal{Z}^n \to[0,1], with output f(z)f(z) at a realized report vector z=(z1,…,zn)z = (z_1,\ldots,z_n).

Benchmark. We evaluate predictions using squared loss. An omniscient aggregator knows both θ\theta and all private signals, so its optimal prediction is the full-signal posterior mean

Mθ=Eθ[Y∣S],mθ(s)=Eθ[Y∣S=s].M_\theta= \mathbb{E}_\theta[Y \mid S], \qquad m_\theta(s) = \mathbb{E}_\theta[Y \mid S=s].

Given aggregation rule ff, its regret relative to this benchmark is

R(f,θ)=Eθ[(f(Z)−Y)2]−Eθ[(Mθ−Y)2].R(f,\theta) = \mathbb{E}_\theta[(f(Z)-Y)^2] - \mathbb{E}_\theta[(M_\theta-Y)^2].

We seek a rule minimizing worst-case regret, which we call a robust optimal rule. For interval reports with nominal coverage level α\alpha, define the optimal worst-case regret over a class Θ\Theta of information structures by

VΘ(n,α)=inf⁡f:Zn→[0,1]sup⁡θ∈ΘR(f,θ).(2)V_\Theta(n,\alpha)=\inf_{f:\mathcal{Z}^n\to[0,1]}\sup_{\theta\in\Theta}R(f,\theta). \tag*{(2)}

If Θ\Theta is unrestricted, no aggregation rule has lower worst-case regret than the constant rule f≡1/2f\equiv1/2, as the following example shows.

Example 3.1 (XOR information structure). Let (S1,S2)(S_1,S_2) be uniformly distributed on {0,1}2\{0,1\}^2 and set Y=S1⊕S2Y=S_1\oplus S_2, where ⊕\oplus denotes addition modulo two. For n>2n>2, let the remaining signals be constant.

Every expert’s posterior is Bernoulli(1/2)(1/2), so all interval reports are [0,1][0,1] at every 0<α<10<\alpha<1; all lower median reports are 0. Thus the report vector is constant, while the full signals reveal YY exactly. If a rule outputs hh at this report vector, then

R(f,θ)=12h2+12(1−h)2=14+(h−12)2≥14.R(f,\theta)=\frac{1}{2}h^2+\frac{1}{2}(1-h)^2=\frac{1}{4}+\left(h-\frac{1}{2}\right)^2\geq\frac{1}{4}.

The constant rule f≡1/2f\equiv1/2 satisfies R(f,θ)=Eθ[(Mθ−1/2)2]≤1/4R(f,\theta)=\mathbb{E}_\theta[(M_\theta-1/2)^2]\leq1/4 for every information structure. Together with the XOR lower bound, this shows that the optimal worst-case regret over unrestricted information structures is 1/41/4 for either reporting format.

In the XOR example, the first two signals are information complements: neither is informative alone, but together they determine the outcome. To exclude such scenarios, we introduce the weak substitutes condition [8, 25].

Definition 3.2 (Weak substitutes). For A⊆[n]A\subseteq[n], write SA=(Si)i∈AS_A=(S_i)_{i\in A} and define the information value of this coalition by

vθ(A)=Eθ[(Y−EθY)2]−Eθ[(Y−Eθ[Y∣SA])2].v_\theta(A)=\mathbb{E}_\theta[(Y-\mathbb{E}_\theta Y)^2]-\mathbb{E}_\theta[(Y-\mathbb{E}_\theta[Y\mid S_A])^2].

This is the improvement in expected squared loss from observing SAS_A, relative to the uninformed prediction EθY\mathbb{E}_\theta Y; in particular, vθ(∅)=0v_\theta(\varnothing)=0. An information structure satisfies weak substitutes if

vθ(A∪{i})−vθ(A)≥vθ(B∪{i})−vθ(B)v_\theta(A\cup\{i\})-v_\theta(A)\geq v_\theta(B\cup\{i\})-v_\theta(B)

for every A⊆B⊆[n]A\subseteq B\subseteq[n] and i∉Bi\notin B.

Weak substitutes means that the ex ante marginal value of an additional signal is nonincreasing as the observed coalition grows. Figure 5 illustrates this property in a symmetric information structure, plotting the information value of a coalition’s signals against its size. The following equivalent variance representation justifies this measure of information value and will be used throughout the paper.

Proposition 3.3 (Variance representation of weak substitutes). For A⊆[n]A\subseteq[n], let Mθ,A=Eθ[Y∣SA]M_{\theta,A}=\mathbb{E}_\theta[Y\mid S_A], so Mθ,∅=EθYM_{\theta,\varnothing}=\mathbb{E}_\theta Y and Mθ,[n]=MθM_{\theta,[n]}=M_\theta. Then vθ(A)=Var⁡θ(Mθ,A)v_\theta(A)=\operatorname{Var}_\theta(M_{\theta,A}). Consequently, the weak substitutes condition is equivalent to

Var⁡θ(Mθ,A∪{i})−Var⁡θ(Mθ,A)≥Var⁡θ(Mθ,B∪{i})−Var⁡θ(Mθ,B)\operatorname{Var}_\theta(M_{\theta,A\cup\{i\}})-\operatorname{Var}_\theta(M_{\theta,A})\geq\operatorname{Var}_\theta(M_{\theta,B\cup\{i\}})-\operatorname{Var}_\theta(M_{\theta,B})
Weak substitutes in a symmetric example

caption: Weak substitutes in a symmetric example: each signal independently reveals a fair binary outcome with probability 1/21/2 and otherwise reveals nothing, giving vθ(A)=(1−2−∣A∣)/4v_{\theta}(A)=(1-2^{-|A|})/4.

for every A⊆B⊆[n]A\subseteq B\subseteq[n] and i∉Bi\notin B. In particular, it implies

∑i=1nVar⁡θ(Mθ,{i})≥Var⁡θ(Mθ).(3)\sum_{i=1}^{n}\operatorname{Var}_{\theta}(M_{\theta,\{i\}})\ge\operatorname{Var}_{\theta}(M_{\theta}). \tag*{(3)}

The proof is given in Appendix A. Let ΘWS\Theta_{\mathrm{WS}} be the class of information structures satisfying weak substitutes, and abbreviate VΘWS(n,α)V_{\Theta_{\mathrm{WS}}}(n,\alpha) as V(n,α)V(n,\alpha). When θ\theta is fixed, we write Mi=Mθ,{i}M_i=M_{\theta,\{i\}} and M=MθM=M_{\theta} for the individual and full-signal posterior means.

Aggregation of Prediction Intervals

For interval reports zi=[ℓi,ui]z_i=[\ell_i,u_i], write c(zi)=(ℓi+ui)/2c(z_i)=(\ell_i+u_i)/2 for each reported midpoint and cˉ(z)=n−1∑ic(zi)\bar{c}(z)=n^{-1}\sum_i c(z_i) for their simple average.

Theorem 4.1 (Regularized Simple Average for Prediction Intervals). For every n≥2n\ge2 and 0<α<10<\alpha<1, a robust optimal aggregation rule for Θ=ΘWS\Theta=\Theta_{\mathrm{WS}} in (2) is

fn,α∗(z)=ηn,α∗cˉ(z)+(1−ηn,α∗)12f_{n,\alpha}^{*}(z)=\eta_{n,\alpha}^{*}\bar{c}(z)+(1-\eta_{n,\alpha}^{*})\frac{1}{2}
=12+ηn,α∗(cˉ(z)−12).(4)=\frac{1}{2}+\eta_{n,\alpha}^{*}\left(\bar{c}(z)-\frac{1}{2}\right). \tag*{(4)}

Let an=(1+8/n−1)/4a_n=(\sqrt{1+8/n}-1)/4, so that nan(1+2an)=1na_n(1+2a_n)=1. The weight is

ηn,α∗={1+nαn(1+α),0<α≤an,2α,an<α<min⁡{1/2,1/n},1,1/2≤α<1/n,0,1/n≤α<1.(5)\eta_{n,\alpha}^{*}= \begin{cases} \frac{1+n\alpha}{n(1+\alpha)}, & 0<\alpha\le a_n,\\ 2\alpha, & a_n<\alpha<\min\{1/2,1/\sqrt{n}\},\\ 1, & 1/2\le\alpha<1/\sqrt{n},\\ 0, & 1/\sqrt{n}\le\alpha<1. \end{cases} \tag*{(5)}

We call this rule regularized simple average for prediction intervals. Its worst-case regret equals V(n,α)V(n,\alpha), where

V(n,α)={(n−1)(n+1+2nα)4n2(1+α)2,0<α≤an,(n−1)αn(1+α)+(12−α)2,an<α<min⁡{1/2,1/n},n−14n(1−α2),1/2≤α<1/n,14,1/n≤α<1.(6)V(n,\alpha)= \begin{cases} \frac{(n-1)(n+1+2n\alpha)}{4n^{2}(1+\alpha)^{2}}, & 0<\alpha\le a_n,\\ \frac{(n-1)\alpha}{n(1+\alpha)}+\left(\frac{1}{2}-\alpha\right)^{2}, & a_n<\alpha<\min\{1/2,1/\sqrt{n}\},\\ \frac{n-1}{4n(1-\alpha^{2})}, & 1/2\le\alpha<1/\sqrt{n},\\ \frac{1}{4}, & 1/\sqrt{n}\le\alpha<1. \end{cases} \tag*{(6)}

The third branch in both formulas applies only when n=2n=2 or 33. In particular, V(n,α)<1/4V(n,\alpha)<1/4 if and only if α<1/n\alpha<1/\sqrt{n}.

The rule moves the average reported midpoint toward the center 1/21/2, placing weight ηn,α∗\eta^*_{n,\alpha} on the reports and the remaining weight on 1/21/2. This adjustment accounts for the fact that an interval midpoint need not equal the expert’s posterior mean. At the threshold αthre(n)=1/n\alpha_{\mathrm{thre}}(n)=1/\sqrt{n}, every weight in [0,min⁡{2/n,1}][0,\min\{2/\sqrt{n},1\}] is optimal, and in Equation (5) we select the constant rule. The adjacent formulas for worst-case regret V(n,α)V(n,\alpha) agree at every boundary.

Optimal reporting level. Before proving Theorem 4.1 and explaining the weights ηn,α∗\eta^*_{n,\alpha}, we use the formulas to identify the nominal coverage level α∗(n)\alpha^*(n) that minimizes the worst-case regret.

Theorem 4.2 (Optimal reporting level). For each integer n≥2n\ge2, the optimal worst-case regret V(n,α)V(n,\alpha) has a unique minimizer over 0<α<10<\alpha<1, denoted by α∗(n)\alpha^*(n). This minimizer satisfies

n(α∗(n))2(3+2α∗(n))=1,n\left(\alpha^*(n)\right)^2\left(3+2\alpha^*(n)\right)=1,
an<α∗(n)<min⁡{1/2,1/n},(7)a_n<\alpha^*(n)<\min\{1/2,1/\sqrt{n}\}, \tag*{(7)}

and strictly decreases as nn increases, with α∗(n)∼(3n)−1/2\alpha^*(n)\sim(3n)^{-1/2} as n→∞n\to\infty. Moreover,

inf⁡0<α<1V(n,α)=14−2(α∗(n))3=14−2+o(1)33n3/2,\inf_{0<\alpha<1}V(n,\alpha)=\frac{1}{4}-2\left(\alpha^*(n)\right)^3=\frac{1}{4}-\frac{2+o(1)}{3\sqrt{3}n^{3/2}},
sup⁡0<α<1V(n,α)=14.\sup_{0<\alpha<1}V(n,\alpha)=\frac{1}{4}.

The infimum is attained only at α=α∗(n)\alpha=\alpha^*(n), and the supremum is attained exactly when 1/n≤α<11/\sqrt{n}\le\alpha<1.

Proof. The second branch of Equation (6) has derivative

nα2(3+2α)−1n(1+α)2.\frac{n\alpha^2(3+2\alpha)-1}{n(1+\alpha)^2}.

Its numerator is strictly increasing and changes sign from negative to positive between ana_n and min⁡{1/2,1/n}\min\{1/2,1/\sqrt{n}\}. The first branch strictly decreases and the third (when present) strictly increases, so continuity gives the unique minimizer. The supremum follows directly from Theorem 4.1. Substituting Equation (7) gives the minimum value. Finally, α2(3+2α)\alpha^2(3+2\alpha) is strictly increasing, so its solution at 1/n1/n strictly decreases with nn. Equation (7) gives α∗(n)→0\alpha^*(n)\to0 and n α∗(n)→1/3\sqrt{n}\,\alpha^*(n)\to1/\sqrt{3}, yielding the stated asymptotics.

Low reporting levels can pool weak and strong evidence into the same extreme report, whereas high levels can pool beliefs on opposite sides of 1/21/2 into the same wide interval. The optimal reporting level balances these two losses of information. The minimum worst-case regret inf⁡αV(n,α)\inf_\alpha V(n,\alpha) increases toward 1/41/4 as nn grows. Thus, more expert reports yield a larger worst-case regret relative to the corresponding omniscient aggregator benchmark.

Simple-average comparisons. Taking η=1\eta=1 gives simple averaging of interval midpoints, f1(z)=cˉ(z)f_1(z)=\bar{c}(z). Its exact worst-case regret determines the orange band in Figure 2.

Corollary 4.3 (Simple average for prediction intervals). For every n≥2n\ge2 and 0<α<10<\alpha<1, write Vmid(n,α)=sup⁡θ∈ΘWSR(f1,θ)V_{\mathrm{mid}}(n,\alpha)=\sup_{\theta\in\Theta_{\mathrm{WS}}}R(f_1,\theta) for the worst-case regret achieved by the simple average. Then

Vmid(n,α)={min⁡{1−α2,n−12n(1+α)},0<α≤12,min⁡{14,n−14n(1−α2)},12<α<1.(8)V_{\mathrm{mid}}(n,\alpha)= \begin{cases} \min\left\{\dfrac{1-\alpha}{2},\dfrac{n-1}{2n(1+\alpha)}\right\}, & 0<\alpha\le\dfrac{1}{2},\\[6pt] \min\left\{\dfrac{1}{4},\dfrac{n-1}{4n(1-\alpha^2)}\right\}, & \dfrac{1}{2}<\alpha<1. \end{cases} \tag*{(8)}

Consequently,

inf⁡0<α<1Vmid(n,α)=min⁡{14,n−13n},sup⁡0<α<1Vmid(n,α)=n−12n.(9)\begin{aligned} \inf_{0<\alpha<1} V_{\mathrm{mid}}(n,\alpha) &= \min\left\{\frac{1}{4},\frac{n-1}{3n}\right\},\\ \sup_{0<\alpha<1} V_{\mathrm{mid}}(n,\alpha) &= \frac{n-1}{2n}. \tag*{(9)} \end{aligned}

Appendix C.2 proves the corollary using a method similar to that for Theorem 4.1. Under posterior mean reports, simple average is robust optimal with worst-case regret (n−1)/(4n)(n-1)/(4n). The argument of Neyman and Roughgarden [25] implies this optimality result in our bounded-outcome setting. Appendix C.1 gives a direct proof.

We prove Theorem 4.1 in three steps. Section 4.1 considers binary outcomes, deriving the upper bound and explaining the weights in Equation (5). Section 4.2 extends this upper bound to all bounded outcomes in [0,1][0,1] through Lemma 4.6, which also applies to median reports. Section 4.3 then gives matching lower bounds using binary information structures, completing the proof.

Upper Bound for Binary Outcomes

For now suppose Y∈{0,1}Y \in\{0,1\}. For a weight η∈[0,1]\eta\in[0,1], we can write the resulting regularized simple average as

fη(z)=1n∑i=1nqη(zi),qη(zi)=ηc(zi)+(1−η)12.f_\eta(z)=\frac{1}{n}\sum_{i=1}^{n}q_\eta(z_i), \qquad q_\eta(z_i)=\eta c(z_i)+(1-\eta)\frac{1}{2}.

where qη(zi)q_\eta(z_i) denotes the contribution of expert ii’s report. We also write qη(ℓ,u)q_\eta(\ell,u) for qη([ℓ,u])q_\eta([\ell,u]).

For binary outcomes, each expert must hold a Bernoulli posterior. When expert ii’s posterior mean is mi=Eθ[Y∣Si=si]m_i=\mathbb{E}_\theta[Y\mid S_i=s_i], using β=(1−α)/2\beta=(1-\alpha)/2 and the lower-quantile convention in Equation (1) we have:

posterior mean mim_ireport ziz_iqη(zi)q_\eta(z_i)
0≤mi≤β0\le m_i\le\beta[0,0][0,0](1−η)/2(1-\eta)/2
β<mi≤1−β\beta<m_i\le1-\beta[0,1][0,1]1/21/2
1−β<mi≤11-\beta<m_i\le1[1,1][1,1](1+η)/2(1+\eta)/2

{#eq:4.7}

Fix an information structure θ∈ΘWS\theta\in\Theta_{\mathrm{WS}}. Write μ=EθY\mu=\mathbb{E}_\theta Y, vi=Var⁡θ(Mi)v_i=\operatorname{Var}_\theta(M_i), and vN=Var⁡θ(M)v_N=\operatorname{Var}_\theta(M) for the prior mean and the variances of the individual and full-signal posterior means. Jensen’s inequality and Eθ[M∣Si]=Mi\mathbb{E}_\theta[M\mid S_i]=M_i give

R(fη,θ)=Eθ[(M−fη(Z))2]≤1n∑iEθ[(M−qη(Zi))2]=vN−1n∑ivi+1n∑iEθ[(Mi−qη(Zi))2].(10)\begin{aligned} R(f_\eta,\theta)&=\mathbb{E}_\theta[(M-f_\eta(Z))^2]\\ &\le\frac{1}{n}\sum_i\mathbb{E}_\theta[(M-q_\eta(Z_i))^2]\\ &=v_N-\frac{1}{n}\sum_i v_i+\frac{1}{n}\sum_i\mathbb{E}_\theta[(M_i-q_\eta(Z_i))^2]. \tag*{(10)} \end{aligned}

The weak substitutes condition implies ∑ivi≥vN\sum_i v_i\ge v_N by Proposition 3.3, while vN≤Var⁡(Y)=μ(1−μ)v_N\le\operatorname{Var}(Y)=\mu(1-\mu). To control the remaining reporting error, we bound it in terms of the posterior variance mi(1−mi)m_i(1-m_i):

(mi−qη(zi))2≤(γ−1)mi(1−mi)+(1−η)24.(11)(m_i-q_\eta(z_i))^2\le(\gamma-1)m_i(1-m_i)+\frac{(1-\eta)^2}{4}. \tag*{(11)}

The last term comes from degenerate Bernoulli posteriors mi=0m_i=0 or 11.

Lemma 4.4. For 0<α<10<\alpha<1 and 0≤η≤10\le\eta\le1, inequality (4.9) holds for every Bernoulli posterior with mean mim_i and its corresponding contribution qη(zi)q_\eta(z_i) if and only if

γ≥Γα(η):=max⁡{2η1+α,η(2−η)1−α2}.\gamma\ge\Gamma_\alpha(\eta):=\max\left\{\frac{2\eta}{1+\alpha},\frac{\eta(2-\eta)}{1-\alpha^2}\right\}.

The two terms come from the extreme-report and central-report regions, respectively. Appendix § 8.§ 1 proves the lemma. For any valid γ∈[0,n]\gamma\in[0,n], substituting inequality (4.9) into inequality (4.8) gives

R(fη,θ)≤vN−γn∑ivi+(γ−1)μ(1−μ)+(1−η)24≤n−1nγμ(1−μ)+(1−η)24≤n−14nγ+(1−η)24.\begin{aligned} R(f_{\eta},\theta) \le v_N-\frac{\gamma}{n}\sum_i v_i+(\gamma-1)\mu(1-\mu)+\frac{(1-\eta)^2}{4} \\ &\le\frac{n-1}{n}\gamma\mu(1-\mu)+\frac{(1-\eta)^2}{4} \\ &\le\frac{n-1}{4n}\gamma+\frac{(1-\eta)^2}{4}. \end{aligned}

Here E[Mi(1−Mi)]=μ(1−μ)−vi\mathbb{E}[M_i(1-M_i)]=\mu(1-\mu)-v_i; the second inequality uses both ∑ivi≥vN\sum_i v_i\ge v_N and vN≤μ(1−μ)v_N\le\mu(1-\mu), with 1−γ/n≥01-\gamma/n\ge0.

For α≤1/n\alpha\le1/\sqrt{n}, the coefficient Γα(η)\Gamma_{\alpha}(\eta) lies in [0,n][0,n] for every η∈[0,1]\eta\in[0,1]. Choosing γ=Γα(η)\gamma=\Gamma_{\alpha}(\eta) yields

R(fη,θ)≤Bn,α(η)R(f_{\eta},\theta)\le B_{n,\alpha}(\eta)
:=(1−η)24+n−14nmax⁡{2η1+α,η(2−η)1−α2}.(12):=\frac{(1-\eta)^2}{4}+\frac{n-1}{4n}\max\left\{\frac{2\eta}{1+\alpha},\frac{\eta(2-\eta)}{1-\alpha^2}\right\}. \tag*{(12)}

The two requirements coincide at η=2α\eta=2\alpha, apart from η=0\eta=0, so

Bn,α(η)={14−1−nα24n(1−α2)η(2−η),0≤η≤min⁡{2α,1},(n−1)η2n(1+α)+(1−η)24,2α≤η≤1.(13)B_{n,\alpha}(\eta)= \begin{cases} \frac{1}{4}-\frac{1-n\alpha^2}{4n(1-\alpha^2)}\eta(2-\eta), & 0\le\eta\le\min\{2\alpha,1\},\\[6pt] \frac{(n-1)\eta}{2n(1+\alpha)}+\frac{(1-\eta)^2}{4}, & 2\alpha\le\eta\le1. \end{cases} \tag*{(13)}

For α<1/n\alpha<1/\sqrt{n}, the first piece decreases with η\eta, and the second is a quadratic minimized at (1+nα)/(n(1+α))(1+n\alpha)/(n(1+\alpha)). Minimizing these two pieces gives the positive weights in Equation (5). At small levels the quadratic minimum determines the weight. It meets 2α2\alpha when nα(1+2α)=1n\alpha(1+2\alpha)=1, giving the transition ana_n. At intermediate levels the best weight lies where the two constraints meet; if 2α2\alpha reaches one before the phase threshold, the best weight is one. These are the three positive-weight cases in Equation (5).

When α=1/n\alpha=1/\sqrt{n}, both branches in Bn,α(η)B_{n,\alpha}(\eta) are no less than 1/41/4, indicating that this upper bound no longer yields a strict improvement over the constant rule. The matching lower bound in § 4.3 shows that no aggregation rule can achieve worst-case regret below 1/41/4 at or above this threshold. This explains the phase threshold αthre(n)=1/n\alpha_{\mathrm{thre}}(n)=1/\sqrt{n}.

Proposition 4.5 (Binary upper bound). For every n≥2n\ge2 and 0<α<10<\alpha<1, regularized simple average for prediction intervals satisfies

R(fn,α∗,θ)≤V(n,α)R(f^*_{n,\alpha},\theta)\le V(n,\alpha)

for every θ∈ΘWS\theta\in\Theta_{\mathrm{WS}} with Y∈{0,1}Y\in\{0,1\}, where the right-hand side is the expression in Equation (6).

Proof. For α<1/n\alpha<1/\sqrt{n}, substitute the three positive weights in Equation (5) into inequality (4.10). They give the first three branches of Equation (6), respectively. For α≥1/n\alpha\ge1/\sqrt{n}, the proposed rule is fn,α∗≡1/2f^*_{n,\alpha}\equiv1/2, whose regret is E[(M−1/2)2]≤1/4\mathbb{E}[(M-1/2)^2]\le1/4.

Extension to Bounded Outcomes

The core idea is that we can decompose the general posterior over [0,1][0,1] into a family of binary random variables. Specifically, given expert ii's posterior Wi∼P(Y∈⋅∣Si=si)W_i\sim\mathbb{P}(Y\in\cdot\mid S_i=s_i), we can construct

Bi,t=1{Wi>t}B_{i,t}=\mathbf{1}\{W_i>t\}

for every threshold t∈(0,1)t \in(0,1). Moreover, we have a similar decomposition for the quantiles:

ℓi,t=1{ℓi>t},ui,t=1{ui>t}.\ell_{i,t} = \mathbf{1}\{\ell_i > t\}, \quad u_{i,t} = \mathbf{1}\{u_i > t\}.

For each threshold tt, Bi,tB_{i,t} is a binary random variable, and the thresholded endpoints ℓi,t\ell_{i,t} and ui,tu_{i,t} are the lower quantiles of Bi,tB_{i,t} at levels β\beta and 1−β1-\beta, respectively. Then we have

Wi=∫01Bi,t dt,W_i = \int_0^1 B_{i,t}\,dt,
qη(ℓi,ui)=∫01qη(ℓi,t,ui,t) dt.(14)q_\eta(\ell_i,u_i) = \int_0^1 q_\eta(\ell_{i,t},u_{i,t})\,dt. \tag*{(14)}

The following lemma transfers a binary posterior inequality to arbitrary posterior distributions supported on [0,1][0,1] through such representations.

Lemma 4.6 (From binary posterior bounds to regret). Fix n≥2n \ge2, 0≤γ≤n0 \le\gamma\le n, and κ≥0\kappa\ge0. Let WW take values in [0,1][0,1], set Bt=1{W>t}B_t = \mathbf{1}\{W > t\}, and suppose a measurable family of predictions ht∈[0,1]h_t \in[0,1] satisfies h=∫01ht dth = \int_0^1 h_t\,dt and, for almost every t∈(0,1)t \in(0,1),

(EBt−ht)2≤(γ−1)Var⁡(Bt)+κ.(15)(\mathbb{E}B_t-h_t)^2 \le(\gamma-1)\operatorname{Var}(B_t)+\kappa. \tag*{(15)}

Then

(EW−h)2≤(γ−1)Var⁡(W)+(n−1)γnE[W(1−W)]+κ.(16)(\mathbb{E}W-h)^2 \le(\gamma-1)\operatorname{Var}(W) + \frac{(n-1)\gamma}{n}\mathbb{E}[W(1-W)]+\kappa. \tag*{(16)}

For any θ∈ΘWS\theta\in\Theta_{\mathrm{WS}}, suppose these hypotheses hold for each expert’s conditional posterior WiW_i and almost every private-signal realization, with report-based prediction h=qi(zi)h=q_i(z_i) and the same γ\gamma and κ\kappa. Then the aggregation rule f(z)=n−1∑iqi(zi)f(z)=n^{-1}\sum_i q_i(z_i) satisfies

R(f,θ)≤(n−1)γ4n+κ.(17)R(f,\theta) \le\frac{(n-1)\gamma}{4n}+\kappa. \tag*{(17)}

The proof is given in Appendix § 8.§ 2.

Proposition 4.7 (Upper bound for interval aggregation). For every n≥2n \ge2, 0<α≤1/n0 < \alpha\le1/\sqrt{n}, θ∈ΘWS\theta\in\Theta_{\mathrm{WS}} with bounded outcome Y∈[0,1]Y \in[0,1], and η∈[0,1]\eta\in[0,1],

R(fη,θ)≤(n−1)Γα(η)4n+(1−η)24=Bn,α(η).(18)R(f_\eta,\theta) \le\frac{(n-1)\Gamma_\alpha(\eta)}{4n}+\frac{(1-\eta)^2}{4}=B_{n,\alpha}(\eta). \tag*{(18)}

Consequently, regularized simple average for prediction intervals satisfies R(fn,α∗,θ)≤V(n,α)R(f_{n,\alpha}^{*},\theta) \le V(n,\alpha) for every 0<α<10<\alpha<1, where the right-hand side is the expression in Equation (6).

For interval reports, Lemma 4.4 supplies inequality (4.13) with ht=qη(ℓi,t,ui,t)h_t=q_\eta(\ell_{i,t},u_{i,t}), γ=Γα(η)\gamma=\Gamma_\alpha(\eta), and κ=(1−η)2/4\kappa=(1-\eta)^2/4. When 0<α≤1/n0<\alpha\le1/\sqrt{n}, we have γ∈[0,n]\gamma\in[0,n], so Lemma 4.6 yields upper bound Bn,α(η)B_{n,\alpha}(\eta). Above this threshold, the selected rule is constant and its regret is at most 1/41/4. Appendix § 8.§ 3 gives the detailed proof.

Matching Lower Bound

Theorem 4.8 (Secret sharing with a common posterior). For every n≥2n \ge2, 0<α<10<\alpha<1, and ϵ>0\epsilon>0, there is a finite information structure θ∈ΘWS\theta\in\Theta_{\mathrm{WS}} with a binary outcome such that every aggregation rule ff satisfies

R(f,θ)≥V(n,α)−ϵ,R(f,\theta)\ge V(n,\alpha)-\epsilon,

where the right-hand side uses the expression in Equation (6). If α>1/n\alpha>1/\sqrt{n}, one such structure satisfies R(f,θ)≥1/4R(f,\theta)\ge1/4 for every ff.

The construction builds on Shamir’s secret-sharing scheme, where sufficiently many private signals recover the outcome while fewer shares reveal nothing [27]. Neyman and Roughgarden [25] randomize the recovery threshold uniformly to satisfy the weak substitutes condition and obtain the lower bound (n−1)/(4n)(n-1)/(4n). In that construction, interval reports introduce no additional information loss relative to posterior means, so the bound falls strictly below V(n,α)V(n,\alpha).

To create that additional ambiguity and obtain a matching lower bound, we add an additional common signal UU. Each expert observes UU, the recovery threshold KK, and a private share. This construction ensures that each expert’s posterior is Bernoulli(U)(U), while all signals together reveal YY. We choose the distribution of UU to hide as much information as possible within each report category. Conditional on UU and KK, the private shares can exhibit threshold complementarities, while the ex ante coalition information values satisfy weak substitutes. The aggregator cannot distinguish different posterior means that induce the same reports, indicating a lower bound on the regret relative to the omniscient benchmark which predicts YY exactly in this construction. By choosing the distribution of UU appropriately, we can match every branch of Equation (6). Appendix B.4 gives the full construction.

Combining Theorem 4.8 with Proposition 4.7 proves Theorem 4.1. Appendix B.5 gives the complete argument.

Aggregation of Medians

For median reports z=(z1,…,zn)z=(z_1,\ldots,z_n), write zˉ=n−1∑i=1nzi\bar{z}=n^{-1}\sum_{i=1}^{n}z_i for their simple average. Denote the robust optimal worst-case regret by

Vmed(n)=inf⁡f:[0,1]n→[0,1]sup⁡θ∈ΘWSR(f,θ).V_{\mathrm{med}}(n)=\inf_{f:[0,1]^n\to[0,1]}\sup_{\theta\in\Theta_{\mathrm{WS}}}R(f,\theta).

Theorem 5.1 (Regularized Simple Average for Medians). For every n≥2n\ge2, a robust optimal aggregation rule for median reports under weak substitutes is

fn,med∗(z)=ηn,med∗zˉ+(1−ηn,med∗)12f^*_{n,\mathrm{med}}(z)=\eta^*_{n,\mathrm{med}}\bar{z}+(1-\eta^*_{n,\mathrm{med}})\frac{1}{2}
=12+1n(zˉ−12),(19)=\frac{1}{2}+\frac{1}{n}\left(\bar{z}-\frac{1}{2}\right), \tag*{(19)}

where ηn,med∗=1n\eta^*_{n,\mathrm{med}}=\frac{1}{n}. We call this rule regularized simple average for medians. Its worst-case regret is

sup⁡θ∈ΘWSR(fn,med∗,θ)=Vmed(n)=14−14n2.(20)\sup_{\theta\in\Theta_{\mathrm{WS}}}R(f^*_{n,\mathrm{med}},\theta)=V_{\mathrm{med}}(n)=\frac{1}{4}-\frac{1}{4n^2}. \tag*{(20)}

The same rule and worst-case regret apply under any method for selecting a median based on each expert’s signal, including when the worst case also ranges over such selections.

As for prediction intervals, regularized simple average combines the average report with the center 1/21/2. The rule improves on the optimal constant prediction in worst-case regret by 1/(4n2)1/(4n^2) for every finite n≥2n\ge2.

Relation to prediction intervals. The optimal weight and worst-case regret agree with the corresponding limits in Theorem 4.1:

ηn,med∗=lim⁡α↓0ηn,α∗=1n,Vmed(n)=lim⁡α↓0V(n,α).\eta^*_{n,\mathrm{med}}=\lim_{\alpha\downarrow0}\eta^*_{n,\alpha}=\frac{1}{n},\qquad V_{\mathrm{med}}(n)=\lim_{\alpha\downarrow0}V(n,\alpha).

Nevertheless, taking the limit α↓0\alpha\downarrow0 in the interval theorem does not by itself prove the median result. For a fair Bernoulli posterior, every positive-level central interval is [0,1][0,1], whereas the lower generalized median is 00. If every expert has this posterior, regularized simple average for prediction intervals always outputs 1/21/2, while regularized simple average for medians outputs (n−1)/(2n)(n-1)/(2n). We therefore establish the median result directly.

Simple-average comparisons. Taking η=1\eta= 1 gives simple averaging of medians, fmed,1(z)=zˉf_{\mathrm{med},1}(z) = \bar{z}. Its exact worst-case regret determines the median SA curve in Figure 2, which coincides with the upper interval-SA boundary in Equation (9).

Corollary 5.2 (Simple averaging for medians). For every n≥2n \ge2, simple averaging fmed,1(z)=zˉf_{\mathrm{med},1}(z) = \bar{z} has exact worst-case regret

sup⁡θ∈ΘWSR(fmed,1,θ)=n−12n,\sup_{\theta\in\Theta_{\mathrm{WS}}} R(f_{\mathrm{med},1}, \theta) = \frac{n-1}{2n},

under any median-selection method.

Appendix § 9.§ 3 proves the corollary.

The proof of Theorem 5.1 follows the same structure as the interval analysis in § 4. § 5.1 first establishes a binary posterior inequality analogous to inequality (11), then applies Lemma 4.6 to extend it to bounded outcomes and derive the upper bound. § 5.2 obtains a matching lower bound by choosing a different distribution for the common signal UU in the secret-sharing construction of § 4.3. § 5.3 combines the two bounds to prove Theorem 5.1.

Upper Bound

Fix n≥2n \ge2 and 0≤η≤10 \le\eta\le1. Define the contribution of expert ii’s median report and the resulting aggregation rule by

qmed,η(zi)=ηzi+1−η2,q_{\mathrm{med},\eta}(z_i) = \eta z_i + \frac{1-\eta}{2},
fmed,η(z)=1n∑i=1nqmed,η(zi).f_{\mathrm{med},\eta}(z) = \frac{1}{n}\sum_{i=1}^{n} q_{\mathrm{med},\eta}(z_i).

Binary posterior inequality. First focus on binary outcomes. For a Bernoulli posterior with mean mim_i, the unique median is zero if mi<1/2m_i < 1/2 and one if mi>1/2m_i > 1/2; when mi=1/2m_i = 1/2, every zi∈[0,1]z_i \in[0,1] is a median. For every such median report, we have

(mi−qmed,η(zi))2≤(2η−1)mi(1−mi)+(1−η)24.(21)(m_i-q_{\mathrm{med},\eta}(z_i))^2 \le(2\eta-1)m_i(1-m_i)+\frac{(1-\eta)^2}{4}. \tag*{(21)}

This is the median counterpart of inequality (11) with γ=2η\gamma= 2\eta.

Extension to bounded outcomes. Lemma 4.6 transfers inequality (21) to bounded outcomes with γ=2η\gamma= 2\eta and κ=(1−η)2/4\kappa= (1-\eta)^2/4.

Proposition 5.3 (Upper bound for median aggregation). For every θ∈ΘWS\theta\in\Theta_{\mathrm{WS}} with bounded outcome Y∈[0,1]Y \in[0,1] and every η∈[0,1]\eta\in[0,1],

R(fmed,η,θ)≤η(n−1)2n+(1−η)24(22)R(f_{\mathrm{med},\eta},\theta) \le\frac{\eta(n-1)}{2n}+\frac{(1-\eta)^2}{4} \tag*{(22)}
=14−14n2+14(η−1n)2.= \frac{1}{4}-\frac{1}{4n^2}+\frac{1}{4}\left(\eta-\frac{1}{n}\right)^2.

This bound holds for any median selection method.

Proof. Fix θ∈ΘWS\theta\in\Theta_{\mathrm{WS}}. For each expert’s posterior Wi∼P(Y∈⋅∣Si=si)W_i \sim\mathbb{P}(Y \in\cdot\mid S_i=s_i) and any reported median ziz_i, define

Bi,t=1{Wi>t},zi,t=1{zi>t}.B_{i,t} = 1\{W_i > t\}, \qquad z_{i,t} = 1\{z_i > t\}.

If zi>tz_i > t, then P(Bi,t=1)≥P(Wi≥zi)≥1/2\mathbb{P}(B_{i,t}=1) \ge\mathbb{P}(W_i \ge z_i) \ge1/2; if zi≤tz_i \le t, then P(Bi,t=0)≥P(Wi≤zi)≥1/2\mathbb{P}(B_{i,t}=0) \ge\mathbb{P}(W_i \le z_i) \ge1/2. Hence zi,tz_{i,t} is a median of Bi,tB_{i,t}, including when that binary posterior has multiple medians. Threshold integration and the affine form of qmed,ηq_{\mathrm{med},\eta} give

qmed,η(zi)=∫01qmed,η(zi,t) dt.q_{\mathrm{med},\eta}(z_i) = \int_0^1 q_{\mathrm{med},\eta}(z_{i,t})\,\mathrm{d}t.

Applying Lemma 4.6 with inequality (5.3), ht=qmed,η(zi,t)h_t = q_{\mathrm{med},\eta}(z_{i,t}), γ=2η∈[0,n]\gamma= 2\eta\in[0,n], and κ=(1−η)2/4\kappa= (1-\eta)^2/4 gives inequality (5.4). The argument applies to any signal-dependent median selection, so the bound holds for every such method.

Matching Lower Bound

Proposition 5.4 (Lower bound for median aggregation). For every n≥2n \ge2 and ε>0\varepsilon> 0, there is a finite binary information structure θ∈ΘWS\theta\in\Theta_{\mathrm{WS}} with unique posterior medians such that every aggregation rule f:[0,1]n→[0,1]f : [0,1]^n \to[0,1] satisfies

R(f,θ)≥14−14n2−ε.R(f,\theta) \ge\frac{1}{4} - \frac{1}{4n^2} - \varepsilon.

Proof. We use the common-posterior construction of § 4.3, choosing the following law of UU to hide different posterior probabilities behind identical median reports.

Fix 0<δ<1/20 < \delta< 1/2, and put ρ=(n−1)/n\rho= (n-1)/n and b=1/2−δb = 1/2-\delta. Define

P(U=0)=P(U=1)=1−ρ2,P(U=b)=P(U=1−b)=ρ2.\mathbb{P}(U=0)=\mathbb{P}(U=1)=\frac{1-\rho}{2}, \qquad\mathbb{P}(U=b)=\mathbb{P}(U=1-b)=\frac{\rho}{2}.

This common posterior mean UU has expectation 1/21/2 and variance

Var⁡(U)=14−ρb(1−b)=14n+n−1nδ2≥14n.\operatorname{Var}(U)=\frac{1}{4}-\rho b(1-b)=\frac{1}{4n}+\frac{n-1}{n}\delta^2 \ge\frac{1}{4n}.

Applying the common-posterior construction of Appendix § 8.§ 4.1 to this law yields a finite information structure θδ\theta_\delta with Mi=UM_i=U for every expert and M=YM=Y. Since Var⁡(U)≥1/(4n)\operatorname{Var}(U) \ge1/(4n), Lemma B.1 ensures that θδ∈ΘWS\theta_\delta\in\Theta_{\mathrm{WS}}.

Every singleton posterior has a unique median: zero when U∈{0,b}U \in\{0,b\} and one when U∈{1−b,1}U \in\{1-b,1\}. Consequently, the only report vectors are 0n0^n and 1n1^n, each with probability 1/21/2. Denote h0=f(0n)h_0=f(0^n) and h1=f(1n)h_1=f(1^n) for an arbitrary aggregation rule ff. Since M=YM=Y, the regret satisfies

R(f,θδ)=ρb(1−ρb)+12(h0−ρb)2+12(h1−(1−ρb))2≥ρb(1−ρb)⟶14−14n2as δ↓0.\begin{aligned} R(f,\theta_\delta)=\rho b(1-\rho b)+\frac{1}{2}(h_0-\rho b)^2+\frac{1}{2}(h_1-(1-\rho b))^2 \\ &\ge\rho b(1-\rho b)\longrightarrow\frac{1}{4}-\frac{1}{4n^2}\qquad\text{as }\delta\downarrow0. \end{aligned}

The lower bound is independent of ff. For every ε>0\varepsilon> 0, sufficiently small δ\delta therefore gives the required finite structure in ΘWS\Theta_{\mathrm{WS}} simultaneously for all rules. Its unique medians make the construction independent of the median selection method.

Proof of Theorem 5.1

Proof. Choosing η=1/n\eta=1/n in Proposition 5.3 gives the rule in (19) and the upper bound

sup⁡θ∈ΘWSR(fn,med∗,θ)≤14−14n2.\sup_{\theta\in\Theta_{\mathrm{WS}}} R(f^*_{n,\mathrm{med}},\theta) \le\frac{1}{4}-\frac{1}{4n^2}.

The rule lies in [0,1][0,1] by convexity. Proposition 5.4 implies that every rule has worst-case regret at least 1/4−1/(4n2)−ε1/4-1/(4n^2)-\varepsilon for every ε>0\varepsilon>0. Letting ε↓0\varepsilon\downarrow0 and taking the infimum over rules yields the matching lower bound for Vmed(n)V_{\mathrm{med}}(n). Together these bounds prove Equation (20) and optimality of the regularized simple average in Equation (19). The upper bound allows any signal-dependent median selection, while the lower-bound structures have unique medians. Thus the result holds for each median selection method and when the worst case also ranges over such methods. □\square

Discussion and Future Work

We characterize robust optimal rules for interval and median reports and show how interval coverage affects optimal worst-case regret.

There are several future directions. First, we assume truthful reports and central prediction intervals at a common nominal coverage level. Future work could study asymmetric intervals or heterogeneous reporting levels across experts. Second, we mainly focus on the weak substitutes condition. Considering more general information structures could provide a richer representation of the aggregator’s knowledge. Finally, the proposed rules have not been evaluated empirically. Robust optimality concerns worst-case regret over the class of information structures, while other rules may achieve lower regret in particular structures. Experiments with real-world data could evaluate how the proposed rules compare with alternatives in practice.

Proof of Proposition 3.3

Proof. Fix θ\theta and write μ=EθY\mu= \mathbb{E}_{\theta}Y. For every A⊆[n]A \subseteq[n], the tower property gives EθMθ,A=μ\mathbb{E}_{\theta}M_{\theta,A} = \mu. Since Mθ,A−μM_{\theta,A} - \mu is determined by SAS_A and Eθ[Y−Mθ,A∣SA]=0\mathbb{E}_{\theta}[Y - M_{\theta,A} \mid S_A] = 0, we have

Eθ[(Y−Mθ,A)(Mθ,A−μ)]=Eθ[(Mθ,A−μ)Eθ[Y−Mθ,A∣SA]]=0.\mathbb{E}_{\theta}[(Y - M_{\theta,A})(M_{\theta,A} - \mu)] = \mathbb{E}_{\theta}[(M_{\theta,A} - \mu)\mathbb{E}_{\theta}[Y - M_{\theta,A} \mid S_A]] = 0.

Expanding Y−μ=(Y−Mθ,A)+(Mθ,A−μ)Y - \mu= (Y - M_{\theta,A}) + (M_{\theta,A} - \mu) therefore yields

Eθ[(Y−μ)2]=Eθ[(Y−Mθ,A)2]+Eθ[(Mθ,A−μ)2].\mathbb{E}_{\theta}[(Y - \mu)^2] = \mathbb{E}_{\theta}[(Y - M_{\theta,A})^2] + \mathbb{E}_{\theta}[(M_{\theta,A} - \mu)^2].

By the definition of vθ(A)v_\theta(A) and EθMθ,A=μ\mathbb{E}_\theta M_{\theta,A}=\mu, it follows that

vθ(A)=Eθ[(Mθ,A−μ)2]=Var⁡θ(Mθ,A).v_\theta(A)=\mathbb{E}_\theta[(M_{\theta,A}-\mu)^2]=\operatorname{Var}_\theta(M_{\theta,A}).

To obtain inequality (3), let Ai={1,…,i}A_i=\{1,\ldots,i\} for i=1,…,ni=1,\ldots,n and A0=∅A_0=\varnothing. Apply weak substitutes with ∅⊆Ai−1\varnothing\subseteq A_{i-1} and i∉Ai−1i\notin A_{i-1}:

vθ(Ai)−vθ(Ai−1)≤vθ({i})−vθ(∅)=vθ({i}).v_\theta(A_i)-v_\theta(A_{i-1})\leq v_\theta(\{i\})-v_\theta(\varnothing)=v_\theta(\{i\}).

Summing these inequalities and using vθ(∅)=0v_\theta(\varnothing)=0, we obtain

vθ([n])=∑i=1n(vθ(Ai)−vθ(Ai−1))≤∑i=1nvθ({i}).v_\theta([n])=\sum_{i=1}^{n}\bigl(v_\theta(A_i)-v_\theta(A_{i-1})\bigr)\leq\sum_{i=1}^{n}v_\theta(\{i\}).

The variance representation now gives inequality (3).

Proofs for Optimal Interval Aggregation

Proof of Lemma 4.4

Proof. Fix expert ii’s Bernoulli posterior with mean mim_i and corresponding interval report zi=[ℓi,ui]z_i=[\ell_i,u_i]. In the low-report region 0≤mi≤β0\leq m_i\leq\beta, we have qη(zi)=(1−η)/2q_\eta(z_i)=(1-\eta)/2. Subtracting the right-hand side of inequality (11) from the left-hand side gives

mi[η−γ(1−mi)],0≤mi≤β.m_i[\eta-\gamma(1-m_i)],\qquad0\leq m_i\leq\beta.

This is nonpositive throughout the region exactly when γ≥η/(1−β)=2η/(1+α)\gamma\geq\eta/(1-\beta)=2\eta/(1+\alpha). Necessity follows by taking mi=βm_i=\beta. The high-report region 1−β<mi≤11-\beta<m_i\leq1 gives the same condition after replacing mim_i by 1−mi1-m_i.

For the central report β<mi≤1−β\beta<m_i\leq1-\beta, we have qη(zi)=1/2q_\eta(z_i)=1/2, and the corresponding difference is

η(2−η)4−γmi(1−mi),β<mi≤1−β.\frac{\eta(2-\eta)}{4}-\gamma m_i(1-m_i),\qquad\beta<m_i\leq1-\beta.

The smallest posterior variance in this region is β(1−β)=(1−α2)/4\beta(1-\beta)=(1-\alpha^2)/4, attained at mi=1−βm_i=1-\beta. Hence the equivalent condition for the central-report region is γ≥η(2−η)/(1−α2)\gamma\geq\eta(2-\eta)/(1-\alpha^2). Combining these two conditions gives Γα(η)\Gamma_\alpha(\eta).

Proof of Lemma 4.6

Proof. On a fixed probability space, define the residual

F(W,h)=(EW−h)2−(γ−1)Var⁡(W)−(n−1)γnE[W(1−W)]=(EW−h)2+(1−γn)Var⁡(W)+(n−1)γn((EW)2−EW).\begin{aligned} \mathcal{F}(W,h)=(\mathbb{E}W-h)^2-(\gamma-1)\operatorname{Var}(W)-\frac{(n-1)\gamma}{n}\mathbb{E}[W(1-W)] \\ &=(\mathbb{E}W-h)^2+\left(1-\frac{\gamma}{n}\right)\operatorname{Var}(W)+\frac{(n-1)\gamma}{n}\bigl((\mathbb{E}W)^2-\mathbb{E}W\bigr). \end{aligned}

The second expression is jointly convex in (W,h)(W,h) because 0≤γ≤n0\leq\gamma\leq n. Since W=∫01Bt dtW=\int_0^1B_t\,dt and Bt(1−Bt)=0B_t(1-B_t)=0, inequality (15) and Jensen’s inequality yield

F(W,h)≤∫01F(Bt,ht) dt≤κ,\mathcal{F}(W,h)\leq\int_0^1\mathcal{F}(B_t,h_t)\,dt\leq\kappa,

which proves inequality (16). For the regret bound, fix θ∈ΘWS\theta\in\Theta_{\mathrm{WS}} and write μ=EY\mu= \mathbb{E}Y, σ2=Var⁡(Y)\sigma^2 = \operatorname{Var}(Y), vi=Var⁡(Mi)v_i = \operatorname{Var}(M_i), vN=Var⁡(M)v_N = \operatorname{Var}(M), and dY=E[Y(1−Y)]d_Y = \mathbb{E}[Y(1-Y)]. The decomposition in inequality (10) also holds for bounded outcomes and contributions qi(Zi)q_i(Z_i):

R(f,θ)≤vN−1n∑ivi+1n∑iE[(Mi−qi(Zi))2].R(f,\theta) \le v_N - \frac{1}{n}\sum_i v_i + \frac{1}{n}\sum_i \mathbb{E}[(M_i-q_i(Z_i))^2].

Apply inequality (16) to each expert’s conditional posterior, take expectations over that expert’s signal realizations, and then average over the nn experts. By the law of total variance, the averaged posterior variance is σ2−n−1∑ivi\sigma^2-n^{-1}\sum_i v_i, while the averaged E[Wi(1−Wi)]\mathbb{E}[W_i(1-W_i)] term equals dYd_Y. Thus

R(f,θ)≤vN−γn∑ivi+(γ−1)σ2+(n−1)γndY+κ≤(1−γn)vN+(γ−1)σ2+(n−1)γndY+κ≤(n−1)γn(σ2+dY)+κ≤(n−1)γ4n+κ.\begin{aligned} R(f,\theta) \le v_N-\frac{\gamma}{n}\sum_i v_i+(\gamma-1)\sigma^2+\frac{(n-1)\gamma}{n}d_Y+\kappa\\ &\le\left(1-\frac{\gamma}{n}\right)v_N+(\gamma-1)\sigma^2+\frac{(n-1)\gamma}{n}d_Y+\kappa\\ &\le\frac{(n-1)\gamma}{n}(\sigma^2+d_Y)+\kappa\\ &\le\frac{(n-1)\gamma}{4n}+\kappa. \end{aligned}

The second inequality uses ∑ivi≥vN\sum_i v_i \ge v_N from weak substitutes and γ≥0\gamma\ge0. The third uses vN≤σ2v_N \le\sigma^2 and γ≤n\gamma\le n, and the last uses σ2+dY=μ(1−μ)≤1/4\sigma^2+d_Y=\mu(1-\mu)\le1/4.

Proof of Proposition 4.7

Proof. Fix n≥2n \ge2, 0<α≤1/n0 < \alpha\le1/\sqrt{n}, θ∈ΘWS\theta\in\Theta_{\mathrm{WS}}, and η∈[0,1]\eta\in[0,1]. Set γ=Γα(η)∈[0,n]\gamma=\Gamma_\alpha(\eta)\in[0,n] and κ=(1−η)2/4\kappa=(1-\eta)^2/4. For each expert’s posterior, the thresholded endpoints in Section 4.2 are valid binary interval reports, so Lemma 4.4 verifies inequality (15) with ht=qη(ℓi,t,ui,t)h_t=q_\eta(\ell_{i,t},u_{i,t}). Equation (14) gives the required representation of h=qη(zi)h=q_\eta(z_i). Lemma 4.6 therefore yields

R(fη,θ)≤(n−1)Γα(η)4n+(1−η)24,R(f_\eta,\theta)\le\frac{(n-1)\Gamma_\alpha(\eta)}{4n}+\frac{(1-\eta)^2}{4},

which is inequality (18). For α<1/n\alpha<1/\sqrt{n}, substituting the weights in Equation (5) gives the first three branches of Equation (6). For α≥1/n\alpha\ge1/\sqrt{n}, the selected rule is constant and has regret E[(M−1/2)2]≤1/4\mathbb{E}[(M-1/2)^2]\le1/4 because M∈[0,1]M\in[0,1]. Thus the proposed rule has the stated upper bound for every 0<α<10<\alpha<1.

Proof of Theorem 4.8

We first define a common-posterior information structure and characterize when it satisfies weak substitutes. We then give the specific choice of the distribution of UU for the low and intermediate reporting levels and for the midpoint and constant regions, respectively. Combining these constructions proves Theorem 4.8.

A common-posterior information structure

We adapt the construction of Neyman and Roughgarden [25, Proposition 2.5] by adding a common signal UU. Let U∈[0,1]U\in[0,1] have finite support and EU=1/2\mathbb{E}U=1/2, and draw Y∣U∼Bernoulli⁡(U)Y\mid U\sim\operatorname{Bernoulli}(U). Independently of (U,Y)(U,Y), draw KK uniformly from {2,…,n}\{2,\ldots,n\}. Fix a prime q>nq>n and, conditional on KK, draw independent uniform coefficients A1,…,AK−1∈FqA_1,\ldots,A_{K-1}\in\mathbb{F}_q, independently of (U,Y)(U,Y). Following Shamir’s scheme [27], define

P(t)=Y+∑j=1K−1Ajtj,Si=(U,K,P(i)),i∈[n].P(t)=Y+\sum_{j=1}^{K-1}A_jt^j,\qquad S_i=(U,K,P(i)),\quad i\in[n].

Let θ\theta be the resulting joint distribution of (Y,S1,…,Sn)(Y,S_1,\ldots,S_n). Every expert observes UU and KK together with a private share P(i)P(i); the aggregator observes only the reports.

Lemma B.1 (Weak substitutes condition). In the common-posterior construction above, Y∼Bernoulli⁡(1/2)Y \sim\operatorname{Bernoulli}(1/2), each expert’s posterior is Bernoulli⁡(U)\operatorname{Bernoulli}(U), and the full set of signals reveals YY. Hence Mi=UM_i=U for every ii and M=YM=Y.

For every nonempty A⊆[n]A \subseteq[n], its information value is

vθ(A)=Var⁡(U)+∣A∣−1n−1(14−Var⁡(U)),(23)v_\theta(A)=\operatorname{Var}(U)+\frac{|A|-1}{n-1}\left(\frac{1}{4}-\operatorname{Var}(U)\right), \tag*{(23)}

with vθ(∅)=0v_\theta(\varnothing)=0. Therefore, the information structure θ\theta defined above satisfies weak substitutes if and only if

Var⁡(U)≥14n.\operatorname{Var}(U)\geq\frac{1}{4n}.

Proof. Any KK distinct shares (P(i1),…,P(iK))(P(i_1),\ldots,P(i_K)) recover Y=P(0)Y=P(0) by polynomial interpolation. For fewer than KK shares, the map from the uniform coefficients to the evaluations has full row rank, so they provide no further information about YY conditional on (U,K)(U,K). Hence, for every nonempty coalition AA of size kk,

MA={Y,K≤k,U,K>k.M_A= \begin{cases} Y, & K\leq k,\\ U, & K>k. \end{cases}

Since K≥2K\geq2, each expert’s posterior is Bernoulli⁡(U)\operatorname{Bernoulli}(U), and since K≤nK\leq n, M=YM=Y. Both UU and YY have mean 1/21/2, and Var⁡(Y)=1/4\operatorname{Var}(Y)=1/4. Independence of KK from (U,Y)(U,Y) and P(K≤k)=(k−1)/(n−1)\mathbb{P}(K\leq k)=(k-1)/(n-1) therefore give Equation (B.1) by Proposition 3.3.

The information value depends only on coalition size. Its first marginal increment is Var⁡(U)\operatorname{Var}(U), and every subsequent increment is (1/4−Var⁡(U))/(n−1)(1/4-\operatorname{Var}(U))/(n-1). Thus the weak substitutes condition holds if and only if

Var⁡(U)≥1/4−Var⁡(U)n−1⟺Var⁡(U)≥14n.\operatorname{Var}(U)\geq\frac{1/4-\operatorname{Var}(U)}{n-1} \quad\Longleftrightarrow\quad \operatorname{Var}(U)\geq\frac{1}{4n}.

The low and intermediate reporting levels

Lemma B.2. If 0<α<min⁡{1/2,1/n}0<\alpha<\min\{1/2,1/\sqrt{n}\}, then for every ε>0\varepsilon>0 there is a finite information structure θ∈ΘWS\theta\in\Theta_{\mathrm{WS}} with a binary outcome such that R(f,θ)≥V(n,α)−εR(f,\theta)\geq V(n,\alpha)-\varepsilon for every aggregation rule ff.

Proof. Set

xn(α)=n−12n(1+α),x=min⁡{xn(α),1/2−α},x_n(\alpha)=\frac{n-1}{2n(1+\alpha)},\qquad x=\min\{x_n(\alpha),1/2-\alpha\},
p=1/(4n)−α2/42(1−β)(β−x).p=\frac{1/(4n)-\alpha^2/4}{2(1-\beta)(\beta-x)}.

Since β−xn(α)=(1−nα2)/(2n(1+α))>0\beta-x_n(\alpha)=(1-n\alpha^2)/(2n(1+\alpha))>0, these choices give 0<x<β0<x<\beta, 0<p≤1/20<p\leq1/2, and

α24+2p(1−β)(β−x)=14n.(24)\frac{\alpha^2}{4}+2p(1-\beta)(\beta-x)=\frac{1}{4n}. \tag*{(24)}

For sufficiently small 0<δ<min⁡{β,α/2}0<\delta<\min\{\beta,\alpha/2\}, put b=β−δb=\beta-\delta and d=β+δd=\beta+\delta, and choose tδt_\delta to satisfy

2p tδb(1−b)+(1−2p)d(1−d)=n−14n.(25)2p\,t_\delta b(1-b)+(1-2p)d(1-d)=\frac{n-1}{4n}. \tag*{(25)}

Equation (B.2) implies tδ→x/β∈(0,1)t_\delta\to x/\beta\in(0,1), so 0<tδ<10 < t_\delta< 1 for all sufficiently small δ\delta. Let CC have probabilities pp, 1−2p1-2p, pp on low, mid, high, respectively, and define

U∣C∼{0 or b,with probabilities 1−tδ,tδ,C=low,d or 1−d,with probabilities 1/2,1/2,C=mid,1 or 1−b,with probabilities 1−tδ,tδ,C=high.U \mid C \sim \begin{cases} 0 \text{ or } b, & \text{with probabilities } 1-t_\delta,t_\delta, & C=\text{low},\\ d \text{ or } 1-d, & \text{with probabilities } 1/2,1/2, & C=\text{mid},\\ 1 \text{ or } 1-b, & \text{with probabilities } 1-t_\delta,t_\delta, & C=\text{high}. \end{cases}

Zero-probability categories are omitted. Symmetry gives EU=1/2EU=1/2, and Equation (B.3) gives

Var⁡(U)=14−E[U(1−U)]=14n\operatorname{Var}(U)=\frac{1}{4}-E[U(1-U)]=\frac{1}{4n}

for every such δ\delta.

Use this law of UU in the construction of Appendix § 8.§ 4.1, and denote the resulting structure by θδ\theta_\delta. Lemma B.1 gives θδ∈ΘWS\theta_\delta\in\Theta_{\mathrm{WS}}, Mi=UM_i=U, M=YM=Y, and vθδ(A)=∣A∣/(4n)v_{\theta_\delta}(A)=|A|/(4n). Since b<β<d<1/2b<\beta<d<1/2, every expert reports [0,0][0,0], [0,1][0,1], or [1,1][1,1] according to C=C= low, mid, high. The conditional outcome means are xδ=tδbx_\delta=t_\delta b, 1/21/2, and 1−xδ1-x_\delta, respectively. Since M=YM=Y, every aggregation rule satisfies

R(f,θδ)≥2pxδ(1−xδ)+1−2p4.R(f,\theta_\delta)\ge2px_\delta(1-x_\delta)+\frac{1-2p}{4}.

As δ↓0\delta\downarrow0, xδ→xx_\delta\to x. The limiting lower bound is

L(p,x)=2px(1−x)+1−2p4={xn(α)(1−xn(α)),α≤an,14−α(1−nα2)n(1+α),α>an.\begin{aligned} L(p,x)=2px(1-x)+\frac{1-2p}{4} \\ &=\begin{cases} x_n(\alpha)(1-x_n(\alpha)), & \alpha\le a_n,\\ \frac{1}{4}-\frac{\alpha(1-n\alpha^2)}{n(1+\alpha)}, & \alpha>a_n. \end{cases} \end{aligned}

These are the first two branches of Equation (6). For any ε>0\varepsilon>0, one sufficiently small δ\delta therefore gives R(f,θδ)≥V(n,α)−εR(f,\theta_\delta)\ge V(n,\alpha)-\varepsilon simultaneously for every ff.

The midpoint and constant regions

Lemma B.3. For every n≥2n\ge2, 0<α<10<\alpha<1, and ε>0\varepsilon>0, there is a finite information structure θ∈ΘWS\theta\in\Theta_{\mathrm{WS}} with a binary outcome such that every aggregation rule ff satisfies

R(f,θ)≥min⁡{14,n−14n(1−α2)}−ε.R(f,\theta)\ge\min\left\{\frac{1}{4},\frac{n-1}{4n(1-\alpha^2)}\right\}-\varepsilon.

If α>1/n\alpha>1/\sqrt{n}, one such structure gives R(f,θ)≥1/4R(f,\theta)\ge1/4 for every ff. Consequently,

V(n,α)≥min⁡{14,n−14n(1−α2)}.V(n,\alpha)\ge\min\left\{\frac{1}{4},\frac{n-1}{4n(1-\alpha^2)}\right\}.

Proof. Choose β<ε<1/2\beta<\varepsilon<1/2 and define

P(U=ε)=P(U=1−ε)=w2,\mathbb{P}(U=\varepsilon)=\mathbb{P}(U=1-\varepsilon)=\frac{w}{2},
P(U=0)=P(U=1)=1−w2.(26)\mathbb{P}(U=0)=\mathbb{P}(U=1)=\frac{1-w}{2}. \tag*{(26)}

where

w=min⁡{1,n−14nε(1−ε)}.w=\min\left\{1,\frac{n-1}{4n\varepsilon(1-\varepsilon)}\right\}.

Then EU=1/2EU=1/2 and Var⁡(U)=1/4−wε(1−ε)≥1/(4n)\operatorname{Var}(U)=1/4-w\varepsilon(1-\varepsilon)\ge1/(4n). The construction of Appendix § 8.§ 4.1 and Lemma B.1 give a finite structure θε∈ΘWS\theta_\varepsilon\in\Theta_{\mathrm{WS}} with Mi=UM_i=U and M=YM=Y. All experts report [0, 1] if and only if U∈{ϵ,1−ϵ}U \in\{\epsilon,1-\epsilon\}. This event happens with probability ww, and conditional on this event, YY is Bernoulli(1/2)(1/2). Thus, writing h=f([0,1],…,[0,1])h=f([0,1],\ldots,[0,1]), every rule satisfies

R(f,θϵ)≥w{12h2+12(h−1)2}≥w4.R(f,\theta_\epsilon)\ge w\left\{\frac{1}{2}h^2+\frac{1}{2}(h-1)^2\right\}\ge\frac{w}{4}.

As ϵ↓β\epsilon\downarrow\beta, this lower bound tends to

min⁡{14,n−14n(1−α2)}.\min\left\{\frac{1}{4},\frac{n-1}{4n(1-\alpha^2)}\right\}.

For any ϵ>0\epsilon>0, choosing ϵ\epsilon sufficiently close to β\beta, independently of ff, gives the stated lower bound for every rule. If α>1/n\alpha>1/\sqrt{n}, some ϵ>β\epsilon>\beta permits w=1w=1, so one structure gives regret at least $1/4. □

Proof of Theorem 4.8. If 0<α<min⁡{1/2,1/n}0<\alpha<\min\{1/2,1/\sqrt{n}\}, Lemma B.2 supplies the required structure for the first two branches of Equation (6). If 1/2≤α<1/n1/2\le\alpha<1/\sqrt{n}, Lemma B.3 gives the third branch. For 1/n≤α<11/\sqrt{n}\le\alpha<1, the same lemma gives the last branch, including the approximating sequence at equality and the single structure above the threshold. These cases cover every 0<α<10<\alpha<1, and each chosen structure works simultaneously for all aggregation rules. □

Proof of Theorem 4.1

Proof. Proposition 4.7 bounds the worst-case regret of fn,α∗f^*_{n,\alpha} by the expression in Equation (6). Theorem 4.8 bounds every rule’s worst-case regret below by the same expression minus any ϵ>0\epsilon>0, using binary structures within ΘWS\Theta_{\mathrm{WS}}. Letting ϵ↓0\epsilon\downarrow0 and taking the infimum over rules proves Equation (6) and optimality of the rule in Equation (4).

The first three branches of Equation (6) are strictly below 1/41/4, and adjacent expressions agree at their boundaries. At α=1/n\alpha=1/\sqrt{n}, Proposition 4.7 and Equation (13) give the upper bound 1/41/4 for every η∈[0,min⁡{2/n,1}]\eta\in[0,\min\{2/\sqrt{n},1\}]; the matching lower bound makes each of these weights optimal. □

Simple Average Comparisons

This appendix establishes the simple-average benchmarks in Figure 2. All worst cases range over the bounded-outcome class ΘWS\Theta_{\mathrm{WS}}, and regret is measured against the omniscient aggregator predicting the full-signal posterior mean M=E[Y∣S]M=\mathbb{E}[Y\mid S].

Simple Average of Posterior Means

Proposition C.1 (Simple average for posterior means). When each expert reports Zi=Mi=E[Y∣Si]Z_i=M_i=\mathbb{E}[Y\mid S_i], the simple average fmeanSA(z)=n−1∑izif^{\mathrm{SA}}_{\mathrm{mean}}(z)=n^{-1}\sum_i z_i is robust optimal, with

inf⁡fsup⁡θ∈ΘWSR(f,θ)=sup⁡θ∈ΘWSR(fmeanSA,θ)=n−14n.\inf_f\sup_{\theta\in\Theta_{\mathrm{WS}}}R(f,\theta)=\sup_{\theta\in\Theta_{\mathrm{WS}}}R(f^{\mathrm{SA}}_{\mathrm{mean}},\theta)=\frac{n-1}{4n}.

Here ff ranges over all measurable rules for aggregating the reported posterior means.

Proof. For each expert’s posterior WiW_i, set Bi,t=1{Wi>t}B_{i,t}=\mathbf{1}\{W_i>t\} and ht=EBi,th_t=\mathbb{E}B_{i,t}. Then ∫01ht dt=EWi=mi\int_0^1h_t\,\mathrm{d}t=\mathbb{E}W_i=m_i, and the binary certificate in Lemma 4.6 holds with γ=1\gamma=1 and κ=0\kappa=0. Its regret bound gives

R(fmeanSA,θ)≤n−14n.R(f^{\mathrm{SA}}_{\mathrm{mean}},\theta)\le\frac{n-1}{4n}.

For the matching lower bound, apply the common-posterior information structure in Appendix § 8.§ 4.1 with common posterior mean UU distributed as follows:

P(U=0)=P(U=1)=12n,P(U=1/2)=n−1n.\mathbb{P}(U=0)=\mathbb{P}(U=1)=\frac{1}{2n},\qquad\mathbb{P}(U=1/2)=\frac{n-1}{n}.

Then EU=1/2\mathbb{E}U=1/2 and Var⁡(U)=1/(4n)\operatorname{Var}(U)=1/(4n), so Lemma B.1 gives a finite information structure θ∈ΘWS\theta\in\Theta_{\mathrm{WS}} with Mi=UM_i=U and M=YM=Y. On the event U=1/2U=1/2, the report vector is (1/2,…,1/2)(1/2,\ldots,1/2) and the conditional outcome variance is 1/41/4. Thus every aggregation rule incurs regret at least (n−1)/(4n)(n-1)/(4n).

Simple Average of Interval Midpoints

Proof of Corollary 4.3. Write f1(z)=cˉ(z)f_1(z)=\bar{c}(z) for simple averaging of interval midpoints, and set β=(1−α)/2\beta=(1-\alpha)/2.

Upper bound. Taking η=1\eta=1 in Lemma 4.4 gives

Γα(1)={2/(1+α),0<α≤1/2,1/(1−α2),1/2<α<1.\Gamma_{\alpha}(1)= \begin{cases} 2/(1+\alpha), & 0<\alpha\le1/2,\\ 1/(1-\alpha^2), & 1/2<\alpha<1. \end{cases}

Whenever Γα(1)≤n\Gamma_{\alpha}(1)\le n, the binary certificate and threshold representation in § 4.2 satisfy Lemma 4.6 with γ=Γα(1)\gamma=\Gamma_{\alpha}(1) and κ=0\kappa=0. Hence, for every θ∈ΘWS\theta\in\Theta_{\mathrm{WS}} with bounded outcome,

R(f1,θ)≤n−14nΓα(1).R(f_1,\theta)\le\frac{n-1}{4n}\Gamma_{\alpha}(1).

For α≤1/2\alpha\le1/2, the required condition holds because Γα(1)≤2≤n\Gamma_{\alpha}(1)\le2\le n. For α>1/2\alpha>1/2, it also holds when α≤1/n\alpha\le1/\sqrt{n}.

A second certificate covers the remaining reporting levels. For a Bernoulli posterior with mean mim_i and the midpoint c(zi)c(z_i) of the reported interval, the three report categories in Equation (4.7) give

EY∼Bernoulli⁡(mi)[(Y−c(zi))2]={mi,0≤mi≤β,1/4,β<mi≤1−β,1−mi,1−β<mi≤1.\mathbb{E}_{Y\sim\operatorname{Bernoulli}(m_i)}[(Y-c(z_i))^2]= \begin{cases} m_i, & 0\le m_i\le\beta,\\ 1/4, & \beta<m_i\le1-\beta,\\ 1-m_i, & 1-\beta<m_i\le1. \end{cases}

Thus the binary squared loss is at most κ=max⁡{β,1/4}\kappa=\max\{\beta,1/4\}. Equivalently, its squared prediction bias is at most −mi(1−mi)+κ-m_i(1-m_i)+\kappa, which is the certificate in Lemma 4.6 with γ=0\gamma=0. The midpoint threshold representation in Equation (14) with η=1\eta=1 lets us apply that lemma directly, yielding

R(f1,θ)≤max⁡{β,1/4}R(f_1,\theta)\le\max\{\beta,1/4\}

for every θ∈ΘWS\theta\in\Theta_{\mathrm{WS}} with bounded outcome.

For 0<α≤1/20<\alpha\le1/2, combining the two certificates gives

R(f1,θ)≤min⁡{1−α2,n−12n(1+α)}.R(f_1,\theta)\le\min\left\{\frac{1-\alpha}{2},\frac{n-1}{2n(1+\alpha)}\right\}.

For 1/2<α<11/2<\alpha<1, the second certificate gives 1/41/4. The first certificate improves this bound exactly when α<1/n\alpha<1/\sqrt{n}, where it is valid, so

R(f1,θ)≤min⁡{14,n−14n(1−α2)}.R(f_1,\theta)\le\min\left\{\frac{1}{4},\frac{n-1}{4n(1-\alpha^2)}\right\}.

Matching lower bounds. Reuse the distribution of UU in Equation (B.4) from the proof of Lemma B.3 with

0<ϵ<1/2,w=min⁡{1,n−14nϵ(1−ϵ)}.0<\epsilon<1/2,\qquad w=\min\left\{1,\frac{n-1}{4n\epsilon(1-\epsilon)}\right\}.

Then EU=1/2EU = 1/2 and Var⁡(U)=1/4−wϵ(1−ϵ)≥1/(4n)\operatorname{Var}(U) = 1/4 - w_{\epsilon}(1-\epsilon) \ge1/(4n). The construction of Appendix § 8.§ 4.1 and Lemma B.1 therefore give a finite structure θϵ∈ΘWS\theta_{\epsilon} \in\Theta_{\mathrm{WS}} with Mi=UM_i = U and M=YM = Y.

If 0<α≤1/20 < \alpha\le1/2, choose ϵ<β\epsilon< \beta. All experts report [0,0][0,0] when U∈{0,ϵ}U \in\{0,\epsilon\}, and report [1,1][1,1] when U∈{1−ϵ,1}U \in\{1-\epsilon,1\}. Thus, as ϵ↑β\epsilon\uparrow\beta,

R(f1,θϵ)=wϵ⟶min⁡{1−α2,n−12n(1+α)}.R(f_1,\theta_{\epsilon}) = w_{\epsilon} \longrightarrow\min\left\{\frac{1-\alpha}{2},\frac{n-1}{2n(1+\alpha)}\right\}.

If 1/2<α<11/2 < \alpha< 1, choose ϵ>β\epsilon> \beta. All experts report [0,1][0,1] when U∈{ϵ,1−ϵ}U \in\{\epsilon,1-\epsilon\}, and report [U,U][U,U] when U∈{0,1}U \in\{0,1\}. Thus, as ϵ↓β\epsilon\downarrow\beta,

R(f1,θϵ)=w4⟶min⁡{14,n−14n(1−α2)}.R(f_1,\theta_{\epsilon}) = \frac{w}{4} \longrightarrow\min\left\{\frac{1}{4},\frac{n-1}{4n(1-\alpha^2)}\right\}.

Together with the upper bounds, the two limits prove Equation (8).

Range over reporting levels. Both expressions inside the minimum in the first branch of Equation (8) strictly decrease on (0,1/2](0,1/2]. The second branch is nondecreasing on [1/2,1)[1/2,1), increasing until it reaches 1/41/4. The branches agree at α=1/2\alpha= 1/2, with value min⁡{1/4,(n−1)/(3n)}\min\{1/4,(n-1)/(3n)\}. This proves the infimum in Equation (9).

For the supremum, the first branch has limit (n−1)/(2n)(n-1)/(2n) as α↓0\alpha\downarrow0 and lies strictly below this limit, while the second is at most 1/4≤(n−1)/(2n)1/4 \le(n-1)/(2n). This proves the supremum in Equation (9). □

Simple Average of Medians

Proof of Corollary 5.2. Taking η=1\eta= 1 in inequality (5.4)(5.4) gives the upper bound for simple averaging:

R(fmed,1,θ)≤n−12nR(f_{\mathrm{med},1},\theta) \le\frac{n-1}{2n}

for every θ∈ΘWS\theta\in\Theta_{\mathrm{WS}}.

For the matching lower bound, use the finite structures θδ\theta_{\delta} from the proof of Proposition 5.4. Under these information structures, the only median-report vectors are 0n0^n and 1n1^n, each with probability 1/21/2. Simple average sets h0=0h_0 = 0 and h1=1h_1 = 1 in each case. Hence

R(fmed,1,θδ)=ρb⟶n−12nR(f_{\mathrm{med},1},\theta_{\delta}) = \rho_b \longrightarrow\frac{n-1}{2n}

as δ↓0\delta\downarrow0.

The unique medians make this matching lower bound valid for every selection method. □

References

  1. [1]Itai Arieli, Yakov Babichenko, and Rann Smorodinsky. Robust forecast aggregation. Proceedings of the National Academy of Sciences, 115(52):E12135–E12143, 2018. doi: 10.1073/pnas.1813934115. URL https://www.pnas.org/doi/abs/10.1073/pnas.1813934115.DOI
  2. [2]Itai Arieli, Yakov Babichenko, and Rann Smorodinsky. Identifiable information structures. Games and Economic Behavior, 120:16–27, 2020. ISSN 0899-8256. doi: https://doi.org/10.1016/j.geb.2019.12.006. URL https://www.sciencedirect.com/science/article/pii/S0899825619301824.
  3. [3]Itai Arieli, Yakov Babichenko, Inbal Talgam-Cohen, and Konstantin Zabarnyi. A random dictator is all you need. American Economic Journal: Microeconomics, 17(1):66–96, February 2025. doi: 10.1257/mic.20230255. URL https://www.aeaweb.org/articles?id=10.1257/mic.20230255.DOI
  4. [4]J. L. Bamber, M. Oppenheimer, R. E. Kopp, W. P. Aspinall, and Roger M. Cooke. Ice sheet and climate processes driving the uncertainty in projections of future sea level rise: Findings from a structured expert judgement approach. Earth’s Future, 10(10):e2022EF002772, 2022. doi: https://doi.org/10.1029/2022EF002772. URL https://agupubs.onlinelibrary.wiley.com/doi/abs/10.1029/2022EF002772.
  5. [5]Itzhak Ben-David, John R. Graham, and Campbell R. Harvey. Managerial miscalibration*. The Quarterly Journal of Economics, 128(4):1547–1584, 11 2013. ISSN 0033-5533. doi: 10.1093/qje/qjt023. URL https://doi.org/10.1093/qje/qjt023.
  6. [6]Johannes Bracher, Evan L. Ray, Tilmann Gneiting, and Nicholas G. Reich. Evaluating epidemic forecasts in an interval format. PLOS Computational Biology, 17(2):1–15, 02 2021. doi: 10.1371/journal.pcbi.1008618. URL https://doi.org/10.1371/journal.pcbi.1008618.
  7. [7]Fabio Busetti. Quantile aggregation of density forecasts. Oxford Bulletin of Economics and Statistics, 79(4):495–512, 2017. doi: https://doi.org/10.1111/obes.12163. URL https://onlinelibrary.wiley.com/doi/abs/10.1111/obes.12163.
  8. [8]Yiling Chen and Bo Waggoner. Informational substitutes. In 2016 IEEE 57th Annual Symposium on Foundations of Computer Science (FOCS), pages 239–247, 2016. doi: 10.1109/FOCS.2016.33.DOI
  9. [9]Zhi Chen, Cheng Peng, and Wei Tang. Prior-agnostic robust forecast aggregation, 2026. URL https://arxiv.org/abs/2604.24517.
  10. [10]Henrique de Oliveira, Yuhta Ishii, and Xiao Lin. Robust aggregation of correlated information. The Review of Economic Studies, page rdag117, 10 2026. ISSN 0034-6527. doi: 10.1093/restud/rdag117. URL https://doi.org/10.1093/restud/rdag117.
  11. [11]Duke University’s Fuqua School of Business. Duke CFO Global Business Outlook Survey: U.S. questionnaire, fourth quarter 2018, 2018. URL https://cfosurvey.fuqua.duke.edu/2018q4/survey/index.htm. Question 6.
  12. [12]Rafael Frongillo, Mary Monroe, Eric Neyman, and Bo Waggoner. Robust forecast aggregation via additional queries, 2025. URL https://arxiv.org/abs/2512.05271.
  13. [13]Anil Gaba, Ilia Tsetlin, and Robert L. Winkler. Combining interval forecasts. Decision Analysis, 14(1):1–20, 2017. doi: 10.1287/deca.2016.0340. URL https://doi.org/10.1287/deca.2016.0340.
  14. [14]Paul H Garthwaite, Joseph B Kadane, and Anthony O’Hagan. Statistical methods for eliciting probability distributions. Journal of the American Statistical Association, 100 (470):680–701, 2005. doi: 10.1198/016214505000000105. URL https://doi.org/10.1198/016214505000000105.
  15. [15]Christian Genest. Vincentization Revisited. The Annals of Statistics, 20(2):1137 – 1142, 1992. doi: 10.1214/aos/1176348676. URL https://doi.org/10.1214/aos/1176348676.
  16. [16]Raffaella Giacomini and Ivana Komunjer. Evaluation and combination of conditional quantile forecasts. Journal of Business & Economic Statistics, 23(4):416–431, 2005. doi: 10.1198/073500105000000018. URL https://doi.org/10.1198/073500105000000018.
  17. [17]Yael Grushka-Cockayne, Kenneth C. Lichtendahl, Victor Richmond R. Jose, and Robert L. Winkler. Quantile evaluation, sensitivity to bracketing, and sharing business payoffs. Operations Research, 65(3):712–728, 2017. doi: 10.1287/opre.2017.1588. URL https://doi.org/10.1287/opre.2017.1588.
  18. [18]Yongkang Guo, Jason D. Hartline, Zhihuan Huang, Yuqing Kong, Anant Shah, and Fang-Yi Yu. Algorithmic robust forecast aggregation. In Proceedings of the 26th ACM Conference on Economics and Computation, EC ’25, page 1110–1129, New York, NY, USA, 2025. Association for Computing Machinery. ISBN 9798400719431. doi: 10.1145/3736252.3742674. URL https://doi.org/10.1145/3736252.3742674.
  19. [19]Tao Hong, Pierre Pinson, Shu Fan, Hamidreza Zareipour, Alberto Troccoli, and Rob J. Hyndman. Probabilistic energy forecasting: Global energy forecasting competition 2014 and beyond. International Journal of Forecasting, 32(3):896–913, 2016. ISSN 0169-2070. doi: https://doi.org/10.1016/j.ijforecast.2016.02.001. URL https://www.sciencedirect.com/science/article/pii/S0169207016000133.
  20. [20]Stephen C. Hora, Benjamin R. Fransen, Natasha Hawkins, and Irving Susel. Median aggregation of distribution functions. Decision Analysis, 10(4):279–291, 2013. doi: 10.1287/deca.2013.0282. URL https://doi.org/10.1287/deca.2013.0282.
  21. [21]Yuqing Kong. Peer expectation in robust forecast aggregation: The possibility/impossibility, 2024. URL https://arxiv.org/abs/2402.06062.
  22. [22]Gilat Levy and Ronny Razin. Combining forecasts in the presence of ambiguity over correlation structures. Journal of Economic Theory, 199:105075, 2022. ISSN 0022-0531. doi: https://doi.org/10.1016/j.jet.2020.105075. URL https://www.sciencedirect.com/science/article/pii/S0022053118303144. Symposium Issue on Ambiguity, Robustness, and Model Uncertainty.
  23. [23]Kenneth C. Lichtendahl, Yael Grushka-Cockayne, and Robert L. Winkler. Is it better to average probabilities or quantiles? Management Science, 59(7):1594–1611, 2013. doi: 10.1287/mnsc.1120.1667. URL https://doi.org/10.1287/mnsc.1120.1667.
  24. [24]Aidan Lyon, Bonnie C. Wintle, and Mark Burgman. Collective wisdom: Methods of confidence interval aggregation. Journal of Business Research, 68(8):1759–1767, 2015. ISSN 0148-2963. doi: https://doi.org/10.1016/j.jbusres.2014.08.012. URL https://www.sciencedirect.com/science/article/pii/S0148296315001526. Special Issue on Simple Versus Complex Forecasting.
  25. [25]Eric Neyman and Tim Roughgarden. Are you smarter than a random expert? the robust aggregation of substitutable signals. In Proceedings of the 23rd ACM Conference on Economics and Computation, EC ’22, page 990–1012, New York, NY, USA, 2022. Association for Computing Machinery. ISBN 9781450391504. doi: 10.1145/3490486.3538243. URL https://doi.org/10.1145/3490486.3538243.
  26. [26]Yuqi Pan, Zhaohua Chen, and Yuqing Kong. Robust decision aggregation with second-order information. In Proceedings of the ACM Web Conference 2024, WWW ’24, page 77–88, New York, NY, USA, 2024. Association for Computing Machinery. ISBN 9798400701719. doi: 10.1145/3589334.3645384. URL https://doi.org/10.1145/3589334.3645384.
  27. [27]Adi Shamir. How to share a secret. Commun. ACM, 22(11):612–613, November 1979. ISSN 0001-0782. doi: 10.1145/359168.359176. URL https://doi.org/10.1145/359168.359176.
  28. [28]James W. Taylor and Derek W. Bunn. Combining forecast quantiles using quantile regression: Investigating the derived weights, estimator bias and imposing constraints. Journal of Applied Statistics, 25(2):193–206, 1998. doi: 10.1080/02664769823188. URL https://doi.org/10.1080/02664769823188.
  29. [29]Johan Ugander, Ryan Drapeau, and Carlos Guestrin. The wisdom of multiple guesses. In Proceedings of the Sixteenth ACM Conference on Economics and Computation, EC ’15, page 643–660, New York, NY, USA, 2015. Association for Computing Machinery. ISBN 9781450334105. doi: 10.1145/2764468.2764529. URL https://doi.org/10.1145/2764468.2764529.

Paper details

vlnc:W_kjjk3mfy
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 2

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