| paper | Forecasting Competitions with Correlated Events |
| authors | Rafael Frongillo, Manuel Lladser, Anish Thilagar, Bo Waggoner |
| venue | AAAI 2025 |
| filed under | frontier · tools-data |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1) |
| judge confidence | high |
| authors would recognise it | yes |
The paper has no named worst-case, algorithmic, or parameterized computational result: Theorem 2 gives statistical event complexity rather than running-time complexity. The weighted type-cohort selector is nevertheless a coherent continuous population extension, so the failure is bit (a), not an impossibility of modelling the population continuously. The opponent's source-level objection therefore decides the grade.
fails bit a — no named computational result to mirror
Any exact-probability complexity may come from the chosen succinct representation of \(D\) and \(D_t\) rather than from mass aggregation; the proponent identifies this concern but does not resolve it.
fatal: False
The mirror covers Theorem 2's accuracy and event-complexity guarantee and reuses Theorem 1 and Theorem 3 as supporting ideas, but covers no computational-complexity result.
The strongest positive case is a population-level forecasting competition, with the forecasters—not the events—continuized.
There is an important qualification first. Under the programme’s strict anchor rule, this paper has no qualifying computational anchor. Theorem 1, proved here, gives approximate truthfulness; Theorem 2, proved here, gives an accuracy and event-complexity guarantee; and Theorem 3, also proved here, gives a concentration inequality. None asserts NP-hardness, membership in P, FPT, W[1]-hardness, or a comparable computational-complexity classification. Lemma 4 is cited from Frongillo et al. (2021) and is not a complexity result either. Thus the formal answer is that the paper supplies no computational anchor. The best positive case must therefore use Theorem 2 as a non-qualifying, computationally adjacent anchor.
The natural mirror is a high-multiplicity forecasting competition. There are many forecasters, but only finitely many complete forecasting types. A type \(t\) consists of the forecaster’s full belief distribution \(D_t\) over the \(m\) binary events, its marginal vector \(p_t\), and any report or utility constraints relevant to the competition. A mass \(\mu_t\) is the fraction of the forecasting population of that type. The true event distribution is \(D\), with marginal vector \(\theta\). Events remain discrete: only the forecaster population is continuous.
This is plausible in a large election-forecasting or prediction-market platform where thousands of participants use one of a finite number of public models, calibration pipelines, data feeds, or institutional forecasting templates. The regime is \(n\gg\tau\): many forecasters, few distinct belief archetypes. Rational \(\mu_t=n_t/n\) expands exactly to a finite competition with \(n_t\) copies of type \(t\).
For a type-level report \(r_t\in[0,1]^m\), let
\[ q_t(y)=\sum_{j=1}^m\bigl(1-(r_{tj}-y_j)^2\bigr). \]
The continuous Multiplicative Weights selector is
\[ P_r(t\mid y)= \frac{\mu_t e^{\eta q_t(y)}} {\sum_{u\in T}\mu_u e^{\eta q_u(y)}}. \]
This is not an arbitrary relaxation: if \(n_t\) identical forecasters submit \(r_t\), their total selection probability in the original mechanism is exactly the expression above after dividing numerator and denominator by \(n\). The objective is also inherited unchanged. Define
\[ a_t=1-\frac1m\|p_t-\theta\|_2^2. \]
The selector should choose a type with \(a_t\geq \max_{u:\mu_u>0}a_u-\epsilon\) with probability at least \(1-\delta\).
A precise candidate problem is:
Weighted Block-Correlated MW Selection\(_\infty\). Given rational masses \(\mu\), a finite type set \(T\), rational descriptions of \(D\) and the \(D_t\), parameters \(b,c,\eta,\epsilon,\delta\), and \(m\) binary events, decide whether
\[ \Pr_{Y\sim D,\;t\sim P_r(\cdot\mid Y)} \left[ a_t\geq \max_u a_u-\epsilon \right]\geq 1-\delta \]
for every type-undominated report profile \(r\). Here a type report is undominated if no alternative report gives that type strictly higher expected selection probability against every profile of the other types. For an ordinary finite-input version, \(D\) and the \(D_t\) could be given by rational block-factor descriptions, including the paper’s disjoint correlated-block examples.
This is the mirror of Theorem 2, which states—proved in this paper—that with \(\eta=\epsilon/(80b)\), \((b,\epsilon/20)\)-block correlation, and
\[ m\geq \frac{400b^2\ln(8n/\delta)}{\epsilon^2}, \]
Multiplicative Weights selects an \(\epsilon\)-optimal forecaster with probability at least \(1-\delta\).
I would expect the continuous problem to be tractable under an explicit block or sampling representation. The weighted selector itself is computable in \(O(\tau m)\) time per outcome, and a weighted version of the paper’s proof should replace the dependence on \(n\) by dependence on \(\tau\) and the smallest positive type mass. A plausible bound is of the form
\[ m= O\!\left( \frac{b^2}{\epsilon^2} \left(\log\frac{\tau}{\delta} +\log\frac1{\mu_{\min}}\right) \right). \]
When every type has macroscopic mass, this removes the irrelevant dependence on the number of duplicated forecasters. Theorem 1 supplies the strategic motivation for retaining undominated reports, while Theorem 3 supplies the concentration mechanism. The expected classification is therefore Class A for this promised, explicitly represented version—not because the paper proves a polynomial-time theorem, but because its statistical argument appears to survive aggregation.
The main further questions are whether the weighted approximate-truthfulness statement really holds for type-level deviations; whether the benchmark should ignore negligible-mass types; and whether exact ex ante winning probabilities can be computed for arbitrary succinct correlated distributions. The last question could itself become continuum-specifically hard: evaluating the expectation of a nonlinear softmax over a succinct correlated distribution resembles a partition-function computation.
The weakest point is substantial. The paper’s central contribution is statistical and incentive-theoretic, not computational. Continuizing the forecaster population may merely compress repeated labels, rather than expose the LP, separation, or hardness phenomena sought by ChoCo. Moreover, aggregating identical forecasters into a strategic type-cohort changes the deviation model: an individual nonatomic deviation has no effect on the denominator, whereas the paper studies named forecasters. Thus this is a credible continuous mirror of Theorem 2, but only a weak programme fit—and formally it remains an unanchored positive case because the paper contains no named computational-complexity result.
The decisive objection is source-level: this paper has no qualifying computational anchor. Theorem 1 is an approximate-truthfulness theorem, Theorem 2 is a statistical event-complexity guarantee, and Theorem 3 is a concentration inequality. None gives a worst-case complexity classification, an algorithmic running-time result, or a parameterized result. Defining a new decision problem around Theorem 2 does not make that problem a computational result of the paper; it manufactures the missing anchor. Under ChoCo’s strict rule, the paper fails before modelling begins.
The proposed forecaster-population mirror is nevertheless the strongest possible candidate, and it is worth examining. If \(n_t\) identical forecasters are duplicated, then their total Multiplicative Weights probability is exactly
\[ \frac{\mu_t e^{\eta q_t(y)}}{\sum_u \mu_u e^{\eta q_u(y)}}. \]
That is a legitimate high-multiplicity identity. It establishes that weighted cohorts are mathematically coherent. But it also shows the limitation: the continuous model is merely the exact aggregation of identical contestants. It introduces no new population-level computational object, and the paper contains no computational problem whose high-multiplicity version could be studied.
More seriously, Theorem 1 is about individual incentives, and those do not survive the atomless limit. In the paper, one named forecaster changes one report while facing the other named forecasters. In a population model, an individual has zero mass and hence zero effect on the softmax denominator and zero probability of being selected. Its incentive problem becomes vacuous. If instead an entire type-cohort controls its mass and changes its common report, that is a coalition or representative-agent game, not the paper’s individual deviation model. Keeping each clone as a positive-mass strategic atom avoids this issue only by abandoning the continuum. No choice of belief type repairs this agency mismatch.
Theorem 2 has a related problem. Its benchmark is the best among \(n\) named forecasters. In the proposed model, a best type may have arbitrarily small or zero mass. If the objective still requires selecting that best model, a mass-weighted selector makes it effectively invisible. If the objective instead asks for a good type with substantial mass, or ignores types below \(\mu_{\min}\), then the benchmark has been changed. The proposed dependence on \(\log(1/\mu_{\min})\) is therefore not a harmless technical replacement for \(\log n\); it encodes a new objective about the population’s mass distribution.
Nor is Theorem 2 computational in the sense ChoCo needs. Its “event complexity” counts observations required for statistical identification. The proposed exact-probability decision problem would owe its difficulty to how the joint distribution \(D\) and the beliefs \(D_t\) are represented—explicit tables, factor graphs, block samplers, or some other succinct encoding—not to continuization. With an explicit outcome table it is a direct finite calculation; with a succinct correlated distribution it may resemble partition-function evaluation. Either way, the computational phenomenon comes from distributional representation, not from replacing forecaster counts by masses.
Theorem 3 is even less suitable. Its variables are event-indexed, and its blocks encode which events influence which other events. Continuizing forecasters leaves this theorem entirely unchanged. Continuizing events would destroy the block topology, temporal positions, and joint-distribution structure that the theorem uses. Reintroducing those features through a structured type space would produce a new model of correlated processes, not a continuous population of agents.
One can propose stronger variants: public forecasting templates used by millions of participants, a continuum of belief vectors, or optimization over the mass of forecasters that can be induced to change reports. The first is a sensible high-multiplicity forecasting model, but its direct MW formula is just weighted clone aggregation. The second loses finite-type fidelity, and its best belief type may have measure zero, creating new selection and incentive semantics. The third could be worthwhile research, but it is a new intervention or mechanism-design problem absent from the paper.
The honest weakness of this negative case is that a weighted, block-correlated forecasting competition may be a perfectly reasonable extension in its own right. What cannot honestly be claimed is that this paper supplies a computational anchor for it, or that Theorems 1–3 become ChoCo population-complexity results merely by replacing \(n\) forecasters with type masses. The positive proposal is therefore a plausible new forecasting model, but not a worthwhile continuous mirror of this paper under the programme’s computational and source-faithfulness requirements.
The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.