acceptodds
Under review as a conference paper at ICLR 2027

The Strategic Cost of Rationality

Abstract

When predictions influence the outcomes they forecast, learning can become impossible: under adversarial responses, regret against an oracle that knows the response is . We show that learnability is restored when the world is rational, meaning it best-responds to each published prediction by minimizing a convex cost under which acting more becomes relatively cheaper as the prediction rises. Rationality then makes the honest prediction self-certifying: the outcome at any prediction reveals its residual, whose square is the excess loss, and the rational response makes that residual monotone, so it points to a unique honest prediction identified from the data at that prediction alone. This gives regret at every response strength. Self-certification needs an honest prediction to exist, but when the world consists of agents of discrete types it may not: no prediction is calibrated, every learner's loss stays a constant above the noise, and zero bias and accuracy can no longer both be achieved. The same rationality that makes the world learnable also makes it predictable enough to steer. A learner with a stake in the outcome gains by steering it: at the honest prediction its marginal incentive splits into conflict of interest and manipulation in the ratio , where is the elasticity of the world's response, and manipulation adds harm to users, beyond what a learner ignoring its influence would cause, exactly when . But steering moves the learner away from the one prediction that certifies itself. The strategic target depends on the slope of the world's response, whereas the outcome at that target reveals only its level, so when the fundamental is unknown, learning the strategic target costs rather than . Rationality thus makes honest prediction self-certifying, but leaves a strategic learner needing information its own outcomes do not reveal.

Primary area:
learning theory
Keywords:
performative prediction, regret bounds, Stackelberg games, rational agents, fixed-point equilibria, strategic manipulation, sample complexity, learning theory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_U9cvOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This submission studies settings where forecasts influence the outcomes they predict, along the lines of performative prediction. In Section 2, the authors introduce a repeated interaction where a "learner" publishes a forecast, and the "world" best-responds by picking a "response intensity" according to a quadratic cost function that results in a response intensity that is monotone in the learner's forecast. In Section 3, the authors show that, under sufficient regularity conditions, the learner can learn a calibrated forecast with logarithmic regret by using a simple update rule. In Section 4, they instead consider a setting where responses jump when a forecast crosses a participation threshold. They show that in this setting, unlike the previous section, achieving zero bias requires sacrificing performance accuracy. In Section 5, the authors study the setting where the learner cares about both prediction accuracy and the outcome itself, and characterize the "strategic forecast" which minimizes the forecaster's combined loss function over these two objectives. In Section 6, the authors show that learning the strategic forecast requires \Theta(T^{1/2}) regret, compared to the logarithmic regret, compared to the O(log T) regret which is required for learning an "honest" forecast which just tries to predict the expected outcome after strategization. Finally in Section 7, the authors ask whether it is possible to discourage manipulation when some parts of the model are unknown. They show that a binning-based approach results in O(T^{2/3}) regret in this setting.

Critical strengths

I appreciated that the authors took the time to list several open directions in an appendix.

Critical weaknesses

I had an extremely difficult time reading this submission and parsing its results. The paper is packed full of technical results and jargon, with very little room left for exposition. It is hard to evaluate the quality of the contributions as a result. Some examples of confusing or unnatural writing from just the first few sections are included in the questions section. There is also no justification for the “quadratic world” assumption, and key details such as model timing are often relegated to an appendix.

Writing issues aside, another weakness of this submission is that its conceptual novelty is incremental compared to the vast literature on performative prediction. For example, the original Performative Prediction paper already gives an example of a setting where discontinuous feedback leads to no stable forecast, and the setting of Section 5 where the learner wants to incentivize good outcomes is well-studied. The paper would greatly benefit from an expanded related work section which better delineates this paper's conceptual and technical contributions from the existing literature.

Questions for authors

None. Examples of confusing writing include:

L46: “Under this condition the response is monotone”, but no condition is mentioned

L46: “The outcome then reveals the residual of the current forecast, and monotonicity makes that residual point to a unique honest forecast, which can be learned at logarithmic regret at every level of responsiveness”

L48: “If the response jumps, as it does when agents of discrete types switch on, no forecast is calibrated”

L52: “A stake in the outcome does, and at the honest forecast its marginal incentive splits into conflict of interest and manipulation in the ratio 1 : ε, where ε is the elasticity of the world’s intensity” -> It's unclear what “elasticity of the world’s intensity” means, as it was not introduced before this point

L69: “A strategic learner is therefore miscalibrated, and a binned audit caps its gain at O(T 2/3 ); in the gap, though, the audit sees a departure from calibration but not the objective behind it” -> the “binned audit” terminology was never introduced

L98: “The timing is Stackelberg, with the learner leading”

L127: “If the world is a population of agents who pay a fixed cost to participate, type k switches on only when the forecast exceeds a threshold”

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

This paper studies sequential forecasting when published forecasts influence outcomes through the response of a rational follower. It distinguishes honest forecasting, where the goal is to accurately predict the response, from strategic forecasting, where the goal is to balance the forecaster’s own profit and the accuracy of the prediction. The main message is that the strategic target can be harder to learn (Theta(sqrt T) vs. Theta(log T) regret) even though both problems involve the same rational world and the same observations.

Critical strengths

The paper addresses an interesting problem at the intersection of online learning and strategic behavior.

Critical weaknesses

The (Theta(sqrt T))-versus-(Theta(log T)) regret separation is closely related to results in dynamic pricing by Keskin and Zeevi (2014) and Broder and Rusmevichientong (2012). The model and objective here differ, but the paper acknowledges that its lower bound technique and level versus slope distinction follow this prior literature.

The paper is also difficult to read as it written in jargon and covers a lot of ground. A clearer narrative centered on the main conceptual contribution would strengthen the paper.

Questions for authors

What is the main conceptual contribution compared to prior work?

Reviewer_L4yPOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents an online learning setting where a learner publishes forecasts p^t\hat p_t to minimize expected regret against what they define a calibrated and honest forecast. In particular, the authors first consider an honest learner whose objective is to minimize prediction error relative to a calibrated forecast (a forecast that agrees with the expected outcome it induces). The authors show that when the world’s response is monotone and a calibrated forecast exists, a stochastic approximation algorithm achieves optimal logarithmic regret. They also consider discontinuous responses arising from agents with discrete types, where a calibrated forecast may not exist. Then they study a learner with an additional stake in the outcome, which may incentivize it to deviate from calibrated predictions. They further show a separation in learning complexity: learning the honest forecast incurs logarithmic regret, whereas learning the strategically optimal forecast incurs \Theta(\sqrt{T}) regret when the underlying fundamental is unknown. Finally, they investigate calibration-based auditing mechanisms for limiting strategic manipulation.

Critical strengths
  • It has lot of theoretical results accompanied with proofs
Critical weaknesses
  • It is hard to understand the main contribution, in comparison to relevant literature. Is it that no one before used the same outcome structure that equation 2.1 introduces?
  • I had a hard time navigating the results; the paper is dense, with a lot of notation and modeling assumptions which are not clearly motivated.
Questions for authors
  • An honest forecast is defined as one satisfying p∗=E[y∣p^=p∗]p^*=\mathbb E[y\mid \hat p=p^*]. Why in practice does the learner aim for calibration? For example a prediction of high congestion might cause people to avoid a certain highway, resulting in low congestion. But why would the learner then aim to predict the congestion that its own announcement induces? Shouldn't they aim to provide information about congestion before people react?
  • The paper emphasizes that rationality restores learnability. However, the relevant mathematical condition is not rationality alone. It is the monotonicity of the response, obtained through decreasing differences in the world’s cost function. Is rationality really the central property driving the results, or is it monotonicity? Can we still hope to derive a similar conclusion under non monodic structures?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.