acceptodds
Withdrawn from ICLR 2027

Learning with Less: Machine Learning from Partial Market Information

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Financial market data provide incomplete views of complex underlying interactions, raising questions about what machine learning models can infer and which information is useful for downstream tasks. We investigate how the availability, granularity, and representation of market observations shape learning in financial market microstructure. Our focus is on the relationship between observed information, learned representations, and their relevance to prediction and decision-making. We consider how useful structure may be extracted from partial observations while accounting for uncertainty about information that is not directly available. A central question is when additional detail provides meaningful value and when a more compact representation may be sufficient for the task of interest. This perspective connects learning under partial observability with the broader problem of aligning representations with downstream objectives. The research aims to clarify the opportunities and limitations of learning from incomplete market information and to inform the development of models whose representations are appropriate to both the available observations and their intended use.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Financial market microstructure, representation learning, partial observability, sequential learning, information sufficiency, decision-making

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_EMD7OpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses an important and often-neglected question: whether richer limit-order-book (LOB) information genuinely improves downstream trading decisions, rather than merely improving predictive metrics. Through carefully matched experiments on Hyperliquid data, the authors examine hidden-book reconstruction, the forecast-to-utility translation, and the effects of depth, history, and delay. The results are predominantly negative and appropriately qualified.

Critical strengths
  • The paper correctly distinguishes between predictive quality and decision utility—a distinction frequently blurred in the LOB forecasting literature.

  • The separation of reconstruction accuracy, forecasting performance, and policy utility is analytically valuable.

  • The paper is easy to follow.

Critical weaknesses
  • The novelty and technical contributions of this paper are unclear.

  • The experimental section can be further enhanced. The evidence rests on a single venue and product family, two assets, a small number of paired dates for the decision and information-sufficiency studies, and a reconstruction cohort that is not a pristine prospective holdout.

  • The decision study uses quote endpoints and assumed costs rather than realized fills, queue priority, market impact, or funding.

  • Some neural fits reach their fixed iteration budget, which the paper notes may explain full-depth underperformance.

Questions for authors

see the weaknesses.

Reviewer_59SCOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper tests whether richer or better-reconstructed order-book information improves trading decisions, using Hyperliquid BTC/ETH perpetual-futures data. It holds the model family, budget, and dates fixed and varies only the information, across several studies: (1) reconstructing hidden book summaries to predict a fill/adverse/no-move target, (2) passing forecasts from several models through fixed trading policies under costs. All return negative results, which the authors interpret as showing that information value is small and fragile at the decision.

Critical strengths
  • An interesting and practically relevant question: whether finer market data actually buys better decisions.
  • Include a limitations section.
  • Code and experiment packages released.
Critical weaknesses
  • The premise is flawed. Holding the model fixed while varying information is not a controlled comparison. The appropriate hypothesis class depends on the data available, and more information requires at least retuning, if not different model classes. The claim that retuning per view is a confound held constant “by construction” has it backwards: fixing the learner is itself the confound.
  • The reconstruction study can’t answer its question. P is reconstructed from C, so any decision rule could extract that information from C directly. And the observed-rich control R is worse than C (+0.018 in August), even though C can always be recovered from R. This shows only that the setup can’t use finer-grained information, so whether reconstruction helps is meaningless here.
  • The decision study is degenerate. Even the oracle loses money (−5.67 / −5.46 bps), and it knows only the direction of the move, not its size. Under 5 bps per-side costs no forecast can be profitable, so absolute utility can’t distinguish forecasts. Negative utility reflects the cost setup, not forecast quality. A forecast of both size and direction could clear the costs.
  • Vague metrics and missing experimental details. Examples:
    • The fill/adverse/no-move target behind the half-Brier score is never defined: no order, horizon, or meaning of “adverse.”
    • The “relative-loss gap” never specifies its loss or target.
    • The horizon, flat threshold, and train/test splits are not stated.
    • Appendix J describes targets as “past-only,” but targets are by definition about the future.
  • Poor writing. Grammar and sentence structure are often garbled, and the text is hard to follow and reads as AI-generated. For instance, “decision-focused learning trains to the decision precisely because training to the forecast is insufficient” relies on ill-defined notions (“training to the decision” vs. “training to the forecast”).
Questions for authors

Please consider addressing the weaknesses listed above. Willing to increase the recommended score upon additional clarifying information.

Reviewer_e68FOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper asks whether more detailed market data improves predictions and decisions. It uses Hyperliquid data for three studies: reconstructing hidden order-book information, comparing forecast scores with policy utility, and changing book depth, history, and delay. The authors report that better intermediate predictions do not always lead to better decisions. The paper also describes thirty experiment packages.

Critical strengths

The paper evaluates forecast scores and policy utility separately. In Section 5, the same frozen forecasts are tested through argmax, confidence, and scheduling policies under explicit costs. This is useful because it shows how the value of a prediction depends on the decision rule. I consider this a meaningful part of the paper's contribution.

Critical weaknesses

The central statement about policy losses is too broad. The abstract and conclusion say every learned policy loses money. However, Table 5 shows positive mean scheduling utility for logistic regression, gradient boosting, and MLP on both assets, while confidence-policy utility is zero. The all-negative result holds for argmax. Positive means do not prove significant or realized trading profits, but they do not support the blanket loss statement either. Since decision value is a main question of the paper, the conclusion needs to identify the policy and criterion that support its negative result.

Questions for authors

Please reconcile the statement that every learned policy loses money with Table 5. Is the intended conclusion about negative argmax utility or failure of the predefined joint gate? Please state the supported conclusion consistently in the abstract and conclusion.

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

Related papers

Discussion (0)

Sign in to comment.