acceptodds
Under review as a conference paper at ICLR 2027

UQ-LOB: Uncertainty-Aware Limit Order Book Mid-Price Forecasting

Abstract

Forecasting short-horizon mid-price movements from limit order book (LOB) data is central to algorithmic trading, yet most deep LOB forecasters are point predictors: they output a direction or a displacement, but never indicate which of their forecasts can be trusted. We introduce UQ-LOB, a lightweight, encoder-agnostic uncertainty quantification module that attaches to any pretrained LOB encoder and, in the spirit of attentive neural processes, conditions each forecast on a context set of recently completed windows whose outcomes are already realised. The UQ-regression variant outputs a calibrated Gaussian over the future tick displacement, while the UQ-classification variant outputs a categorical distribution over down/up/stationary. Both expose a scalar confidence (predicted signal-to-noise ratio or class probability) that supports selective prediction. On 5.2 billion LOB events across seven cryptocurrency assets and horizons of 5, 10 and 15 seconds, UQ-regression attains near-nominal 68% interval coverage, and restricting to the most confident 10% of predictions raises directional macro F1 by 0.11–0.15 for UQ-regression and 0.05–0.11 for UQ-classification, at every horizon. On large, economically meaningful moves, the tightest confidence tier reaches a directional F1 of 0.88 (down) and 0.83 (up) at the 5-second horizon.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Uncertainty Quantification, Limit Order Book, Mid-Price Prediction, Algorithmic Trading, Cryptocurrency, Neural process, Attention mechanism

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ikLtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UQ-LOB, an uncertainty-aware framework for short-horizon mid-price forecasting using limit order book (LOB) data, addressing the lack of explicit uncertainty quantification in conventional point prediction methods. The framework introduces an encoder-agnostic uncertainty quantification head inspired by attentive neural processes, which incorporates recently completed historical windows with observed outcomes through self-attention and cross-attention mechanisms. Two variants are developed: UQ-regression, which predicts a Gaussian distribution over future price displacements using a combination of heteroscedastic likelihood, magnitude-weighted error, directional loss, and calibration loss, and UQ-classification, which predicts probabilities for upward, downward, and stationary movements. Experiments on approximately 5.2 billion LOB events across seven cryptocurrency assets and forecasting horizons of 5, 10, and 15 seconds demonstrate that UQ-regression achieves near-nominal 68% prediction interval coverage. Retaining the top 10% most confident predictions improves directional macro F1 by 0.11–0.15 for UQ-regression and 0.05–0.11 for UQ-classification. For large price movements at the 5-second horizon, the highest-confidence subset achieves F1 scores of 0.88 and 0.83 for downward and upward movements, respectively, demonstrating the potential of uncertainty-aware selective prediction to improve forecasting reliability.

Critical strengths

1.The paper addresses the important challenge of quantifying prediction reliability in limit order book forecasting, shifting the focus beyond predictive accuracy toward assessing the trustworthiness of individual predictions. This problem is practically relevant and provides a valuable direction for selective prediction and trading decision-making in high-frequency financial markets. 2.The paper proposes a flexible probabilistic forecasting framework that integrates historical outcomes through in-context conditioning with uncertainty quantification. The carefully designed composite loss function jointly considers prediction magnitude, direction, and calibration, enabling the model to provide both price displacement estimates and confidence information to support decision-making.

Critical weaknesses

1.The paper lacks direct experimental comparisons with existing methods from the literature. Although several LOB forecasting and uncertainty quantification approaches are discussed in the related work, none are included as competitive baselines. The evaluation primarily compares two variants of the proposed framework, making it difficult to assess its relative performance and incremental contributions over existing approaches. 2.The paper lacks systematic ablation studies to validate the effectiveness of its key components. In particular, the individual contributions of in-context conditioning, attention mechanisms, and the proposed loss functions are not evaluated. Without these experiments, it remains unclear which components are responsible for the reported improvements, weakening the empirical support for the proposed methodological innovations.

Questions for authors

1.Could the authors provide direct comparisons with existing LOB forecasting and uncertainty quantification methods, such as MC Dropout, Deep Ensembles, and the original D-TABL encoder with confidence-based selection? Evaluating these methods under the same experimental settings and prediction coverage rates would help clarify the relative advantages of UQ-LOB. 2.Could the authors provide ablation studies to quantify the contributions of the proposed components, particularly in-context conditioning, self-attention, cross-attention, and the individual loss functions? These results are important for determining whether the reported improvements can be attributed to the proposed architectural and training innovations.

3 comments
AuthorsOpenReview ↗

Response on baselines under the same settings and component ablations

Thank you for the review. The comparisons and ablations you requested are in the revised paper; the key results are below.

1. Baselines under the same settings (Sec. 5.3, Appendix "Baseline comparison"). All methods share the D-TABL encoder initialised from the same pretrained weights, the same test examples and the same retained fractions (confidence thresholds 10–90%). MC Dropout (30 passes), a 5-member Deep Ensemble and conformal prediction are trained with the Gaussian likelihood alone and without a context head. Mean directional F1 at 10 s, averaged over the seven assets:

MethodF1, top 90%F1, top 10%Cal. errorNLPD
MC Dropout0.400.510.0645.73
Deep Ensemble0.420.520.0665.79
Conformal0.390.480.057–
UQ-Regression0.430.550.0485.59

Regression. UQ-Regression has the highest mean F1 at every confidence threshold, and its lead over the strongest baseline, the Deep Ensemble, widens from 0.01 on the top 90% of predictions to 0.03 on the top 10%; against MC Dropout and Conformal the lead is 0.03–0.04 and 0.04–0.07. It also has the best predictive distribution: the lowest calibration error (0.048 against 0.057–0.066), the best NLPD (5.59 against 5.73 and 5.79) and a 95% coverage (0.904) closer to nominal than MC Dropout and the Deep Ensemble (0.886–0.888). Conformal prediction reaches 95% by construction (0.947) but gives no density. Because all baselines share the pretrained encoder, the gap is not an effect of pretraining, and UQ-Regression needs one trained model and one forward pass, against 30 passes for MC Dropout and five trained models for the Ensemble.

Classification. The D-TABL encoder with confidence-based selection is the softmax baseline: the pretrained D-TABL classifier, trained with the same class-weighted cross-entropy as UQ-Classification and gated by its maximum class probability. Its mean F1 moves only from 0.42 to 0.43 between the top 90% and the top 10% of predictions, whereas UQ-Classification rises from 0.45 to 0.50, mainly through down moves (0.47 to 0.55). The maximum softmax probability of the base classifier therefore says little about which directional calls are reliable, while the confidence learned by the context head does.

2. Ablations (Sec. 5.7, Appendix "Ablation studies"). Each variant is retrained from the same pretrained encoder with the same objective. Seven-asset means at 10 s:

VariantF1, top 10%F1, large movesNLPDCal. error
Full UQ-LOB0.5470.7245.5860.048
No context0.5250.6925.7490.063
Masked labels0.5310.7165.6740.054
No attention0.5360.7035.5950.051

Removing the context lowers both the calibration and the direction of large moves. Masking the context labels shows that the window representations carry directional information while the labels supply the scale of the uncertainty. Replacing attention by a mean leaves the uncertainty estimate nearly unchanged and lowers the top-10% and large-move F1 by 0.011 and 0.021. We replace the self- and cross-attention blocks jointly by a mean over the context embeddings, which is the aggregation of the original CNP (Garnelo et al., 2018).

3. Loss functions (leave-one-out, BTC-USD, 10 s; Sec. 3.4, Appendix "Ablation studies"). Without the WMAE term, the weighted R² falls from 0.054 to 0.038 and the top-10% F1 from 0.677 to 0.643. Without the directional term, the top-10% accuracy falls from 80.7% to 77.4% and the large-move accuracy from 87.4% to 84.8%. Removing the calibration-slope term leaves these aggregate metrics unchanged.

All new content in the revised paper is highlighted in green. In response to your questions, the revision adds comparisons with MC dropout, deep ensembles and the D-TABL encoder with softmax-based selection under the same settings and retained fractions, and ablations of the context, the attention blocks and the individual loss terms.

I thank the authors for the additional results and the detailed rebuttal. I will maintain my original score.

AuthorsOpenReview ↗

Follow-up on baselines and ablations

Thank you for reading our response and the additional results.

Following your suggestions, the revised paper now includes:

1. Comparisons under the same settings (Sec. 5.3). MC Dropout, Deep Ensembles, conformal prediction and the original D-TABL encoder with confidence-based selection were evaluated with the same pretrained encoder, test examples and retained fractions. UQ-Regression achieves the highest directional F1 at every confidence threshold, together with the best NLPD, using a single model and a single forward pass. For classification, the D-TABL softmax confidence barely improves F1 as coverage decreases (0.42→0.430.42 \rightarrow 0.43), whereas UQ-Classification rises from 0.450.45 to 0.500.50.

2. Ablations (Sec. 5.7). Removing in-context conditioning reduces top-10% F1 (0.547→0.5250.547 \rightarrow 0.525), large-move F1 (0.724→0.6920.724 \rightarrow 0.692) and calibration (0.048→0.0630.048 \rightarrow 0.063). Replacing the attention blocks with mean aggregation lowers top-10% and large-move F1. Leave-one-out loss ablations (Sec. 3.4, Appendix D "Ablation studies") show that the WMAE and directional terms each contribute to accuracy on confident and large-move predictions.

We hope these results clarify the relative advantages of UQ-LOB and the contribution of its components. We remain fully open to further discussion: if any point is unclear, or if there are additional comparisons or experiments that would help in assessing the work, we would be glad to provide them within the remaining discussion period and to include them in the final version. Thank you again for your time and constructive feedback.

Reviewer_QgZXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper adds an uncertainty head to a pretrained order-book encoder. Before each forecast, the head uses recent windows whose outcomes are already known to produce either a Gaussian distribution over the mid-price movement 5, 10, or 15 seconds ahead, or probabilities for downward, upward, and stationary movements, along with a confidence score. Using data from seven cryptocurrency pairs on Kraken, the authors report that, when pooled across assets, the 68% prediction intervals cover approximately 68% of outcomes, and that directional accuracy improves when only the most confident forecasts are retained.

Critical strengths

The paper addresses an important question for trading: not only where the price will move, but also when a short-horizon forecast is reliable enough to act on. This is what makes the paper worth engaging with.

Critical weaknesses

The main new idea is never tested in isolation. For an ML venue, the paper needs to demonstrate which new idea makes a difference and why it works for this problem. Here, the main contribution is to let the uncertainty head use the known outcomes of recent windows, but there is no experiment without this context (which the authors themselves acknowledge as a missing ablation). As a result, we cannot tell what it actually adds. A simpler explanation for any improvement is that the recent windows merely inform the model about recent market volatility and price direction, which a few simple features might capture just as well.

There are no external baselines. Apart from comparing the two proposed heads, the only reference points are random-guessing levels. There is no comparison with a simple model, an existing LOB forecaster, or a standard uncertainty estimation method such as MC dropout, ensembles, or conformal prediction. Without these baselines, it is difficult to assess how well the proposed approach actually performs.

Trading fees are mentioned but never accounted for, making it unclear in what sense the predicted large moves are “economically meaningful.” Based on my estimate from Table 7, a typical price movement over these horizons is comparable to, or smaller than, a single Kraken taker fee.

Overall, the paper addresses a worthwhile question, but for ICLR, it needs to demonstrate that its proposed ML contribution is responsible for the results and explain why it works. At present, the novel components remain untested in isolation, while the remaining methods are well established. The paper therefore reads more like an application of existing techniques to cryptocurrency order-book data than a validated ML contribution.

Questions for authors

Requested experiments:

What happens when the recent-outcome context is removed (i.e., the same model without context windows)? How does this compare with using simple features such as recent volatility and returns instead?

Clarifications:

None.

2 comments
AuthorsOpenReview ↗

Response to Reviewer QgZX: context ablation, simple features, baselines and trading costs

Thank you for the detailed review. It pushed us to test the central idea in isolation, and the revised paper does so.

1. Context removed, and simple features (Sec. 5.7, Appendix "Ablation studies"). Every variant is retrained from the same pretrained encoder with the same objective and evaluated on the same test examples. Seven-asset means at 10 s:

VariantF1 (all)F1, top 10%F1, large movesNLPDCal. error
Full UQ-LOB0.4100.5470.7245.5860.048
No context0.3930.5250.6925.7490.063
Label statistics0.3980.5180.6975.6410.052
Masked labels0.4030.5310.7165.6740.054
No attention0.4040.5360.7035.5950.051

The simple features are the mean, standard deviation and mean absolute value of the context outcomes (recent return and volatility), mapped by a linear layer. They recover most of the calibration gain (NLPD 5.641, calibration error 0.052, against 5.749 and 0.063 without context), so a large part of the value of context is a local volatility estimate. They do not recover direction: the top-10% F1 is 0.518, below the model without context (0.525). Masking the labels while keeping the windows shows where direction comes from: the window representations keep most of the large-move gain (0.716), while the labels supply the scale of σ. Context therefore provides a volatility estimate and directional information that a few summary features cannot.

2. External baselines (Sec. 5.3). MC Dropout, a 5-member Deep Ensemble, conformal prediction and a softmax baseline all share the encoder weights, test examples and gating of UQ-LOB, so the comparison is not an effect of pretraining. Mean directional F1 on the 10% most confident predictions is 0.55 for UQ-Regression against 0.52 (Deep Ensemble), 0.51 (MC Dropout) and 0.48 (Conformal); the calibration error is 0.048 against 0.057–0.066 and the NLPD 5.59 against 5.73–5.79. The softmax baseline stays flat (0.42–0.43) while UQ-Classification rises from 0.45 to 0.50.

3. Fees and "economically meaningful" (Sec. 5.5, Appendix "Trade simulation"). You are right that typical moves are comparable to a Kraken taker fee, and the revised paper reports this. We simulate SNR-gated trading with entry at the ask, exit at the bid, 0.5 ms latency and causal thresholds. For the three most liquid assets the executable edge is positive from the top 50% of signals down (+2.4 bps at the top 10%, +4.5 bps at the top 1%), but after a 1.25 bps taker fee per side the top 3% and top 1% stay profitable (+1.2 and +2.0 bps), and at 5 bps per side no retention level does; less liquid assets pay 4–5 bps of spread. We removed the phrase "economically meaningful" and now say that large moves are selected from the realised outcome, so these results are an upper bound. Our claim is a reliable confidence ranking, not profitability.

4. ML contribution. What the ablations validate is conditioning on causally realised labels (a constraint that excludes look-ahead) together with the loss terms that shape the output. In a leave-one-out study on BTC-USD (Sec. 3.4, Appendix "Ablation studies"), removing the WMAE term lowers the weighted R² from 0.054 to 0.038, and removing the directional term lowers the top-10% accuracy from 80.7% to 77.4%.

All new content in the revised paper is highlighted in green. In response to your questions, the revision adds an ablation without the context windows and a comparison with simple volatility and return features, external baselines (MC dropout, deep ensembles, conformal prediction and a softmax baseline), and a trade simulation with executable prices and taker fees; the phrase "economically meaningful" is removed.

AuthorsOpenReview ↗

Follow-up on context ablation, baselines and trading costs

Thank you again for your detailed review. As the discussion period closes soon, we would like to briefly summarise how the revised paper addresses your concerns:

1. Context tested in isolation (Sec. 5.7, Appendix "Ablation studies"). Removing the recent-outcome context reduces top-10% F1 (0.547→0.5250.547 \rightarrow 0.525), large-move F1 (0.724→0.6920.724 \rightarrow 0.692) and calibration (0.048→0.0630.048 \rightarrow 0.063). Simple volatility and return features recover much of the calibration gain but not the directional gain, so the context provides information beyond a few summary statistics.

2. External baselines (Sec. 5.3). MC Dropout, Deep Ensembles, conformal prediction and a softmax baseline were evaluated with the same encoder, test examples and gating. UQ-Regression achieves the highest directional F1 on the most confident predictions, together with the lowest calibration error and the best NLPD.

3. Trading fees (Sec. 5.5, Appendix "Trade simulation"). We added a trade simulation with executable prices and taker fees, and removed the phrase "economically meaningful".

4. ML contribution (Sec. 3.4, Sec. 5.7). The ablations isolate the contribution of conditioning on causally realised outcomes and of the individual loss terms.

We would be grateful to know whether these additions address your concerns. If any point remains unclear, or if further comparisons or experiments would help in assessing the work, we would be glad to provide them within the remaining discussion period and to include them in the final version.

Reviewer_U2WtOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UQ-LOB, an encoder-agnostic uncertainty quantification framework for short-horizon limit order book forecasting. Inspired by attentive neural processes, it conditions predictions on recent windows with observed outcomes and introduces regression and classification variants to support confidence-based selective prediction.

Evaluated on 5.2 billion events across seven cryptocurrencies and three forecasting horizons, UQ-LOB achieves near-nominal 68% interval coverage and improves directional macro F1 by 0.11 to 0.15 when retaining the top 10% most confident regression predictions.

Critical strengths
  1. The proposed encoder-agnostic UQ head incorporates recently observed outcomes through causal in-context conditioning, enabling confidence-based selective prediction. This addresses an important limitation of conventional point predictors in LOB forecasting.

  2. Experiments on 5.2 billion events across seven assets and three forecasting horizons demonstrate consistent improvements in directional macro F1 under confidence gating. The additional evaluation on large price movements further supports the practical relevance of the proposed approach.

Critical weaknesses
  1. The paper lacks comparisons with established uncertainty estimation methods, such as MC dropout, deep ensembles, and simpler confidence-gating strategies. Ablations of the context-conditioning mechanism and individual loss components are also missing, making it difficult to attribute the reported improvements to the proposed contributions.

  2. Despite near-nominal 68% coverage, the 95% prediction intervals substantially under-cover (around 90.5%), raising concerns about calibration. Furthermore, improvements under confidence gating are not evaluated against realistic trading costs or profitability, limiting the evidence for practical benefits.

Questions for authors

How does UQ-LOB compare against simpler uncertainty estimation approaches, such as MC dropout, deep ensembles, or confidence gating using the pretrained encoder's softmax outputs? Can the authors provide ablations isolating the contributions of in-context conditioning and the individual loss components?

3 comments
AuthorsOpenReview ↗

Response: baselines, ablations, 95% coverage and trading costs

Thank you for the careful and constructive review. The additional experiments it prompted are now in the revised paper, and the key results are below.

Baselines (Sec. 5.3, Appendix "Baseline comparison"). All baselines use the same D-TABL encoder initialised from the same pretrained weights, the same test examples and the same confidence gating, and are trained with the Gaussian likelihood alone (no context head). MC Dropout uses 30 passes, the Deep Ensemble 5 members, Conformal fixed-width split-conformal intervals. Mean directional F1 (average of down and up) at 10 s, averaged over the seven assets:

MethodF1, top 90%F1, top 10%Cal. errorNLPDCov95
MC Dropout0.400.510.0645.730.888
Deep Ensemble0.420.520.0665.790.886
Conformal0.390.480.057–0.947
UQ-Regression0.430.550.0485.590.904

UQ-Regression has the highest mean F1 at every confidence threshold and the lowest calibration error (|Cov68 − 0.68| + |Cov95 − 0.95|, computed per asset and averaged), with one trained model and one forward pass, against 30 passes for MC Dropout and five models for the Ensemble.

Softmax gating of the pretrained encoder. This baseline is the pretrained D-TABL classifier gated by its maximum class probability. Its mean F1 stays flat (0.42 to 0.43 from the top 90% to the top 10%), whereas UQ-Classification rises from 0.45 to 0.50.

Ablations (Sec. 5.7 and 3.4, Appendix "Ablation studies"). Context: retraining without it raises the NLPD from 5.586 to 5.749 and the calibration error from 0.048 to 0.063, and lowers the F1 on large moves from 0.724 to 0.692. Loss terms (leave-one-out, BTC-USD, 10 s): without the WMAE term the weighted R² falls from 0.054 to 0.038 and the top-10% F1 from 0.677 to 0.643; without the directional term the top-10% accuracy falls from 80.7% to 77.4% and the large-move accuracy from 87.4% to 84.8%; removing the calibration-slope term leaves these aggregate metrics mostly unaffected.

95% coverage. The 95% interval covers 90.4–90.7%, while the 68% interval covers 68.1–68.2%. This is the pattern of heavier-tailed residuals: a Gaussian predictive distribution that is calibrated at the one-sigma level under-covers at 95% when the residuals have heavier tails than a Gaussian. Among the Gaussian methods, the 95% coverage of UQ-Regression (0.904) is closer to nominal than that of MC Dropout (0.888) and the Deep Ensemble (0.886). Conformal prediction is closest (0.947), as its coverage guarantee implies, but it gives no predictive density, its fixed-width intervals cannot adapt to individual predictions, and it covers only 65.9% at the 68% level.

Trading costs (Sec. 5.5, Appendix "Trade simulation"). We simulate SNR-gated trading with entry at the ask, exit at the bid after 10 s, 0.5 ms latency, causal thresholds and one position at a time. For the three most liquid assets, the mid-to-mid edge rises from 1.0 bp (all signals) to 5.2 bps (top 1%), the spread costs 0.7–1.5 bps, and the executable edge is positive from the top 50% down (+2.4 bps at the top 10%, +4.5 bps at the top 1%). After a 1.25 bps taker fee per side, the top 3% and top 1% stay profitable (+1.2 and +2.0 bps); at 5 bps per side no retention level does, and less liquid assets pay 4–5 bps of spread. We therefore present confidence as a reliability signal and not as a standalone strategy, and we state that large moves are selected from the realised outcome, so these figures are an upper bound.

All new content in the revised paper is highlighted in green. In response to your questions, the revision adds comparisons with MC dropout, deep ensembles and softmax confidence gating (Sec. 5.3), ablations of the context and of the individual loss terms (Sec. 5.7 and the appendix), an explanation of the 95% coverage and its comparison with the baselines (Sec. 5.3), and a trade simulation with executable prices and taker fees (Sec. 5.5 and the appendix).

Thank you for the rebuttal. I have adjusted the scores.

AuthorsOpenReview ↗

Thank you for the updated assessment

Thank you for engaging with our response and for raising your score. We appreciate your constructive feedback, which helped us strengthen the paper with the baseline comparisons and ablations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.