acceptodds
Under review as a conference paper at ICLR 2027

Learning Monotone Information Time for Fixed-Budget Intraday Representations

Abstract

We study representations of intraday market and order-book data for predicting stocks' relative next-day close-to-close returns. Uniform tokenization partitions 240 one-minute records at a fixed temporal resolution, although trading activity varies across intervals, stocks, and trading dates. We introduce a tokenizer that learns sample-dependent temporal coverage under a fixed token budget. A local causal encoder assigns a positive information rate to each minute; normalized cumulative rates define a monotone time coordinate. Soft aggregation around a fixed number of equally spaced centers in this coordinate yields tokens with adaptive spans in natural time while preserving chronological order. The rates are learned jointly with the prediction objective, without supervision of discrete boundaries. A causal convolutional Mixer and a constrained pooling head integrate these representations to predict next-day stock-ranking scores. Historical validation results support the tested adaptive encoding configuration and its learned rate-based allocation. A system built on the proposed method achieved champion status in an external quantitative modeling competition. The organizers executed the submitted training code and evaluated the resulting model on hidden test data using the submitted inference code; test labels and scores were not used for training or model selection.

Primary area:
learning on time series and dynamical systems
Keywords:
time series, adaptive tokenization, representation learning, monotone time warping, financial forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_9z2iOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a sample-dependent tokenizer for fixed-budget intraday time-series representations. Given 240 one-minute market and order-book records, a local causal encoder predicts a positive rate for each minute. The normalized cumulative rates define a monotone time coordinate, and the minute representations are softly aggregated around 48 uniformly spaced centers in this coordinate. This produces a fixed number of ordered tokens with variable coverage in natural time. A causal convolutional Mixer then predicts stocks’ relative next-day returns. The model is trained on data from 2019–2023 and evaluated on a 2024 validation period. The reported comparisons favor sample-dependent allocation over fixed, uniform, and position-dependent alternatives. The complete system also reportedly achieved first place in an external competition with hidden evaluation data.

Critical strengths
  1. Clear and practically motivated method. The proposed tokenizer provides a simple differentiable mechanism for reallocating a fixed token budget across each stock-day. Positive rates guarantee chronological order, while soft aggregation avoids discrete boundary prediction and always produces a fixed-size representation.
  2. Encouraging matched comparisons. The matched-context experiment retains the minute encoder, projection, backbone, heads, and training procedure while changing the allocation mechanism. Sample-dependent allocation outperforms all three controls across the reported seeds, and the same ordering remains at the final training epoch. The paper is also clear about its evidence boundaries and limitations.
Critical weaknesses
  1. The paper does not provide independent evidence that the tokenizer generalizes to new data. All scientific comparisons use the same 2024 period for checkpoint selection and reported evaluation. Multiple training seeds do not address temporal generalization. The hidden competition result evaluates the complete system under a different training procedure and does not include a matched comparison between sample-dependent and fixed allocation. It therefore cannot establish that the tokenizer itself generalizes out of period.
  2. The empirical scope and comparison with prior work are insufficient for a general representation-learning contribution. The method is evaluated on one proprietary dataset, one market, one token budget, and one prediction task. There are no comparisons with recent adaptive patching or learned time-warping methods, despite the authors explicitly stating that such comparisons remain necessary. Runtime, token-budget sensitivity, assignment-temperature sensitivity, and aggregate behavior of the learned rates are also not reported. These omissions make both the generality and novelty of the contribution difficult to assess.
Questions for authors
  1. Can the authors evaluate sample-dependent allocation and the matched controls on an untouched later period using a checkpoint-selection rule fixed before examining that period? Alternatively, can the same matched ablation be evaluated on the competition’s hidden period?
  2. How does the method compare with existing adaptive patching and learned time-warping methods? Can the authors provide results on at least one public time-series benchmark so that the contribution can be reproduced and evaluated outside the proprietary financial dataset?
  3. How sensitive are the results to the number of tokens and the learned assignment temperature? Although soft assignment avoids hard boundaries, can the authors test whether the aggregation preserves abrupt or multi-minute market events rather than averaging them across tokens?
  4. Can the authors provide aggregate diagnostics of the learned allocation, including token-width distributions, stability across seeds and time periods, and relationships with observable quantities such as volume, volatility, and order-book changes?
  5. Can the authors provide exact definitions of IC, ICIR, Sharpe proxy, stress IR proxy, and the competition composite score, together with uncertainty estimates that account for temporal dependence and checkpoint selection?
Reviewer_C8pTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an adaptive tokenizer for intraday financial time series that learns sample-dependent information rates to transform 240 minute-level observations into 48 fixed-budget tokens. These tokens are processed by a causal convolutional Mixer to predict next-day stock rankings. Experiments show improvements over fixed-window tokenization (IC 0.0676 vs. 0.0549), with three-seed ablations supporting the benefits of adaptive allocation. The model also outperforms the tested Transformer baselines and achieves champion status in an external competition. However, most experimental comparisons rely on the same validation period used for model selection, limiting evidence of out-of-sample generalization.

Critical strengths
  1. The proposed monotone information-time formulation provides an elegant, differentiable mechanism for allocating temporal resolution based on sample-specific characteristics while preserving chronological ordering and a fixed token budget.
  2. The matched-context experiments isolate the effect of sample-dependent allocation from the backbone and local encoder, demonstrating consistent improvements over uniform, position-dependent, and fixed-interval alternatives across three seeds.
Critical weaknesses
  1. Most experimental comparisons rely on the same 2024 validation period used for checkpoint selection, introducing potential selection bias. Although the authors report success in an external competition, the hidden-test results do not directly validate the proposed tokenization mechanism. This weakens confidence in the robustness of the reported improvements.
  2. The ablations primarily compare against fixed or simplified allocation strategies, without evaluating established adaptive tokenization or temporal resampling approaches. This makes it difficult to assess the novelty and practical advantages of the proposed method relative to existing techniques.
Questions for authors

Can the authors provide independent out-of-period evaluation results for the proposed tokenizer and its matched-context baselines, using a fixed checkpoint selection protocol? This would help determine whether the reported improvements generalize beyond the repeatedly used 2024 validation period.

Reviewer_UtNrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • Study representations of intraday market and order-book data for predicting stocks’ relative next-day close-to-close returns.
  • Develop a method for tokenizing a stock’s daily time-series for use in a causal sequence model. Enter a competition, which supplies 240 minute records per stock-day.
  • While equal-width partitioning assigns a vector of the same dimension to every 5-min interval, this work learns variable temporal coverage. High-rate intervals therefore get more tokens
  • Contribution has two parts: shared network for sample-dependent, monotone, differentiable temporal allocation under a fix 48-token budget, and integration with causal backbone and cross-sectional prediction objective.
  • Each sample contains 240 one-minute records for a stock at the close of trading day (market and order-book prices, quantities, and order counts. Model predicts score for stock’s relative return on the next trading day.
  • 18M params. Model is an after-close predictor (not streaming predictor). Normalization, mask concat, projection, three stride-one causal convolution produce 128-d context every minute.
  • Architecture assigns 48 token centres based on soft assignment using midpoint coordinates. My understanding is that pooled regions corresponding to individual tokens can overlap due to soft assignment, therefore no discrete cut points in the tokenization. Each minute’s 128-d context gives a positive rate; the normalized cumulative sum of rates gives a monotone coordinate normalized between 0 and 1. Then 48 centres are equally spaced in that coordinate; minutes are softly assigned to centres by distance.
  • Causal Mixer backbone alternates between temporal and channel mixing, using cyclical levels of convolutional dilation. Kernel-size 5 temporal depth wise convolution followed by linear projection to exchange channel info. Channel mixing is just MLP. Use residual connections, normalization in both branches Readout is weighted average of last token and all tokens with learned pooling. Primary head ranks next-day returns, auxiliary head (for training) averages three consecutive one-period labels.
  • Evaluation: 240K stock-date samples, 241 effective trading days, 1.2K stocks in 2024 public validation period of competition.
Critical strengths

Paper is clear, well-written, easy to understand, and well motivated. Numerical results appear to be consistent across seeds. Findings adaptive allocation beats uniform, position-only, and fixed-window controls (IC .065 vs. about .057, 3 seeds), and the proposed model won the competition. Straightforward with limitations.

Critical weaknesses
  • Heavy overfitting due to selection on evaluation year; IC peaks at epoch 20 and drops by epoch 40 for every configuration. No results from competition’s hidden test set, competition is unknown, no metrics, num candidates, etc.
  • No analysis of what the rates learn, would like to understand what the learned rules for allocation; how they correlated with volume, spread, volatility. How far is it from uniform?
  • Why never vary the budget beyond 48 tokens; even if it’s fixed, would be important to understand the effect of other budgets. Does adaptive allocation help more when budget is tight? Does it matter if we just use lots of tokens? Networks are quite small, and the best results are similar with only 4-blocks (similar to 11 and 6-block results), can tradeoff some of the depth budget for more tokens.
  • None of the adaptive tokenization baselines are included in the empirical evaluation. Sharpe proxy and IR… formal definitions?
  • 240 minutes with 1.2K stocks suggests Chinese A-shares, but market is never mentioned anywhere. Data is not public, no code release. For ICLR can apply the method and test on other intraday datasets or other time-series benchmarks.
  • Unabated: importance of 3-day auxiliary loss, label neutralization, bounded pooling gate (for mixing last token with global pooled rep to construct feature).
  • A hand-defined clock based on volume, trade count, or volatility is the standard finance approach and is cheap to test.
Questions for authors

See questions in weaknesses section. Happy to reconsider my rating based on additional information from authors.

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

This paper proposes a sample-adaptive tokenization framework for intraday financial time series, termed Monotone Information Time. The method predicts positive information rates for individual minutes, constructs a monotone cumulative temporal coordinate, and uses soft aggregation to compress 240 minute-level records into 48 adaptive tokens. These representations are processed by a causal convolutional Mixer for next-day stock return ranking. The authors report improvements over fixed temporal encoding and several controlled allocation variants on historical financial data, and additionally claim that a system based on the proposed method won an external quantitative modeling competition.

Critical strengths

S1. The paper addresses the limitations of uniform temporal partitioning, which may fail to allocate sufficient representation capacity to informative periods. The proposed adaptive allocation mechanism provides a reasonable approach to handling heterogeneous intraday dynamics.

S2. The method constructs a monotone temporal coordinate through positive information rates and uses differentiable soft aggregation to generate a fixed number of tokens. This design preserves temporal ordering while allowing sample-dependent temporal coverage.

Critical weaknesses

W1. Lack of independent out-of-sample evaluation. Most experimental conclusions rely on the same 2024 validation period, which is repeatedly used for checkpoint selection and architectural comparisons. Although the authors acknowledge this limitation, the absence of an independent evaluation period makes it difficult to establish whether the observed improvements generalize beyond the selected historical data.

W2. Missing computational efficiency evaluation. Although the method maintains a fixed output token count, its tokenizer requires full-resolution causal encoding, information-rate computation, and dense soft aggregation before the backbone. The paper does not report training time, inference latency, memory consumption, or throughput against matched fixed-token baselines, making its computational trade-offs unclear.

W3. Several sections and figures repeatedly explain the same tokenization mechanism or restate methodological limitations, while the most important experimental questions receive comparatively limited attention. The paper reads more like an extended engineering report than a focused machine learning research paper, making it difficult to identify the central research insights.

Questions for authors

Q1. Could the authors provide evaluation results on an independent chronological test period that was not used for model selection? In particular, does the advantage of sample-dependent temporal allocation persist under a strictly forward-in-time evaluation protocol?

Q2. Could the authors compare the proposed method with recent adaptive patching and temporal allocation approaches (such as APN: Rethinking Irregular Time Series Forecasting, AAAI26), or explain why these methods are not applicable to the current task?

Q3. Could the authors report end-to-end computational costs, including tokenizer overhead, training time, inference latency, and GPU memory usage, under matched model configurations?

1 comment

Thank the authors for the responses. However, my main concerns still remain. The authors say "The independent-test concern remains valid.", which makes the paper's improvements not robust. So I will retain my rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.