DiffVolume: Diffusion Models for Volume Generation in Limit Order Books

Carmine Ventre; Zhuohan Wang

REVIEW 4 major objections 3 minor 1 cited by

DiffVolume: Diffusion Models for Volume Generation in Limit Order Books

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A conditional diffusion model can generate realistic limit order book volumes and improve liquidity forecasting.

desk verdict DiffVolume is a plausible diffusion-based LOB volume generator, but its counterfactual and downstream claims rest on a conditioning target that may not support true counterfactuals, and the abstract alone can't rule out leakage. read the letter →

arxiv 2508.08698 v1 pith:LMOZUEXZ submitted 2025-08-12 q-fin.TR cs.LG

Zhuohan Wang , Carmine Ventre This is my paper · ORCID

classification q-fin.TRcs.LG

keywords limitorderbooksdiffusionmodelsvolumegenerationcounterfactualliquidityforecastingmarketmicrostructuregenerativeconditional

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes DiffVolume, a conditional diffusion model for generating future limit order book volume snapshots. Conditioned on past volume history and time of day, it aims to reproduce the statistical texture of real order books—marginal distributions, spatial correlation across price levels, and autocorrelation decay—more faithfully than previous GAN-based generators. It further aims to support counterfactual generation by conditioning on a chosen future liquidity profile, so a user can ask what volumes would look like if liquidity evolved in a specified way. The load-bearing claim is that synthetic counterfactual volumes, when added to real training data, improve the accuracy of future liquidity forecasting models on real data. If correct, the paper turns a generative model into a controllable data-augmentation tool for market microstructure forecasting.

What carries the argument

The engine is a conditional denoising diffusion model. Diffusion models generate data by learning to reverse a gradual corruption process: the model is trained to remove noise added in small steps, and sampling starts from pure noise and iteratively denoises. The conditioning inputs—past volume windows, time-of-day indicators, and optionally a target future liquidity profile—steer this reverse process, so one trained model can draw realistic volumes from the learned distribution or push generation toward a prescribed liquidity scenario.

What would settle it

A direct test: train DiffVolume on one subset of stocks or time periods, generate counterfactual volumes under extreme target liquidity profiles, and measure whether their statistical properties (autocorrelation, cross-level correlation, consistency with price and order-flow dynamics) degrade relative to unconditional samples. Separately, evaluate the downstream forecast gain on a hold-out period completely disjoint from the generator's training window and compare it against augmenting with real historical snapshots; if the gain disappears or shrinks to that baseline, the counterfactual-data c

Watch

Extended reading notes

Core claim

The central claim is that volume snapshots in a limit order book can be generated by a single conditional diffusion model that is simultaneously realistic and controllable. The model is trained to reverse a noise-adding process, with conditioning on the past volume history and the time of day; for counterfactual use it is additionally conditioned on a target future liquidity profile. Evaluated on three axes, the paper reports that the generated volumes better match the marginal distribution, spatial correlation, and autocorrelation decay of real data than previous approaches, that the additional conditioning produces plausible counterfactual volumes, and that augmenting training data with th

Load-bearing premise

The load-bearing premise is that prescribing a future liquidity profile is a valid way to sample counterfactual volumes—that the diffusion model, trained on observed joint distributions, can be safely conditioned on a future state that did not occur in training and still produce volumes whose realism and downstream usefulness survive the forced conditioning.

Editorial extensions

If this is right

If the realism claims hold, DiffVolume provides a high-dimensional generative model of LOB volumes that preserves cross-level spatial correlation and temporal autocorrelation, which earlier GAN-based LOB generators reportedly missed.
If the controllability claims hold, researchers and traders can generate volume snapshots under hypothetical future liquidity conditions, enabling stress-testing and what-if backtests without waiting for such scenarios to occur naturally.
If the downstream claim holds, synthetic counterfactual volumes are a valid data-augmentation source: adding them to real data can improve the accuracy of future liquidity forecasting models.
The same conditional-generation setup could extend to other LOB fields (prices, depths, order flow imbalance) and to multi-asset order books, wherever a conditional distribution over a high-dimensional market state is needed.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The counterfactual claim is the one to watch: in a real order book the future liquidity profile is an outcome of order flow, not a knob a trader can set. Conditioning on a prescribed future profile can push the model off the learned joint distribution, so realism measured on unconditional samples does not automatically transfer to counterfactual samples.
The downstream improvement could be partly a data-augmentation effect rather than a counterfactual effect: adding any plausible volume samples, even historical ones, might help a forecaster. The paper's comparison should ideally isolate whether the 'counterfactual' part, not just the extra data, drives the gain.
A testable extension would be to condition on extreme or rare liquidity profiles (flash-crash-like depletion) and check whether generated volumes remain internally consistent with price and order-flow dynamics, and whether forecasters trained on such stress data generalize to real stress periods.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: I could only work from the abstract—my copy of the full text was corrupted and unusable. Based on the abstract, DiffVolume is a sensible conditional diffusion method for LOB volume snapshots, and the counterfactual/downstream angle is a step beyond most GAN-based generation papers, which usually stop at realism. That ambition is the most interesting thing here.

What's new: replacing GANs with conditional diffusion for LOB volumes is a natural and reasonable move, and conditioning on past volume history and time of day is commonsense. The three-axis evaluation—realism, counterfactual generation, downstream forecasting—is a good structure for a method paper.

Soft spots: the main concern is the counterfactual premise. Conditioning on a target future liquidity profile produces draws from the conditional distribution, not from an interventional distribution. In a real order book, the future profile is an outcome of the same order-flow process, so the model may simply learn correlations rather than counterfactual responses. The realism metrics don't test whether the conditional samples correspond to the hypothetical scenarios the authors claim. The downstream improvement claim also carries a leakage risk: no temporal split or baseline that controls for the generator's training data is described in the abstract. These are concerns, not refutations—a solid full paper would address them with held-out splits, ablations, and proper baselines. But I cannot judge whether the full text does that, because I couldn't read it.

Verdict: I would send this to a serious referee rather than desk-reject. The idea is plausible, the application is relevant, and the counterfactual issue is worth a rigorous test. I wouldn't cite it yet, and the authors should be pushed hard on the counterfactual framing and the leakage controls. A reading group might spend an hour on the causal semantics alone.

Referee Report

4 major / 3 minor

Summary. The paper proposes DiffVolume, a conditional diffusion model for generating future limit order book (LOB) volume snapshots. The model is conditioned on past volume history and time of day, and can additionally be conditioned on a target future liquidity profile for counterfactual generation. The abstract claims three evaluation axes: (1) realism, where DiffVolume outperforms prior approaches in reproducing marginal distributions, spatial correlations, and autocorrelation decay; (2) counterfactual generation under hypothetical liquidity scenarios; and (3) downstream prediction, where synthetic counterfactual data improve future liquidity forecasting. No quantitative results, baseline names, dataset description, or significance tests are provided in the supplied text, and the full text is largely corrupted/unreadable.

Significance. If substantiated, the contribution would be a flexible diffusion-based generative model for LOB volumes that combines realism, controllable counterfactual generation, and downstream utility. The counterfactual and downstream claims in particular would be notable: conditional generation on future liquidity profiles would enable scenario analysis, and augmentation of forecasting training data with realistic synthetic volumes is practically valuable. The paper also addresses a gap left by GAN-based approaches in LOB generation. However, the significance is currently prospective only: the supplied manuscript does not allow verification of any of the three main claims, and the counterfactual framing raises causal-identification issues that are not addressed.

major comments (4)

[Abstract, axes (1)-(3)] The central empirical claims are made without any numbers, baselines, dataset description, or statistical significance. 'Better reproduces' and 'improves the performance' are not operationalized: no metric values, confidence intervals, or error bars appear in the supplied text. Since these three axes are the paper's entire contribution, the manuscript as submitted does not support its abstract.
[Counterfactual generation (abstract, axis 2)] Conditioning on a 'target future liquidity profile' treats an endogenous outcome of order flow as an exogenously prescribable input. The sampling distribution is P(volume | past, target), not the interventional distribution P(volume | do(target)) needed for valid counterfactual reasoning. The realism metrics on axis (1) cannot validate axis (2) because they only check marginal/spatial/temporal statistics of observed data. The paper needs an explicit causal or structural justification, or a protocol (e.g., hypothetical interventions, propensity-based reweighting) that connects the conditional distribution to the intended counterfactual.
[Downstream prediction (abstract, axis 3)] The downstream improvement claim is vulnerable to leakage. There is no description of temporal splitting: whether the generator's training set overlaps the forecaster's training/evaluation period, how target liquidity profiles are chosen, and whether the forecaster has access to the target at test time. If the target future liquidity profile leaks information about the future, or if synthetic volumes are generated from the same periods on which the forecaster is tested, the reported gain would not reflect real predictive value. A clear holdout protocol is required.
[Full text] The supplied full text is largely corrupted and unreadable: most paragraphs are garbled or replaced by repeated placeholder strings, and sections, equations, and tables cannot be reliably identified. Key methodological components—model architecture, training objective, noise schedule, data preprocessing, evaluation metrics, hyperparameter choices, and the exact causal protocol for counterfactual generation—are absent. This is a load-bearing omission because the paper's claims cannot be checked in any form.

minor comments (3)

[Abstract] The abstract mentions 'prior approaches' without naming any. At minimum, the baseline(s) (e.g., GAN-based LOB models) should be identified, and the dataset (e.g., LOBSTER, NASDAQ, contract frequency) should be stated.
[Full text] There are no visible section headings, equation numbers, or table/figure captions in the supplied text. Even after repair, the manuscript would need a standard structure with labeled sections and numbered equations/tables.
[Full text, repetitive tables] The latter part of the supplied text contains large repetitive tables of hyperparameters or metrics; these appear to be the same experiments repeated with cosmetic variations. This presentation obscures the actual number of independent trials and any uncertainty quantification.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found in the abstract's evaluation claims.

full rationale

The supplied text consists of the abstract plus a heavily corrupted full-text rendering; the only analyzable content is the abstract. No load-bearing derivation chain is visible that reduces a claimed output to its own input. The realism axis compares generated volumes against external statistical properties (marginal distribution, spatial correlation, autocorrelation decay) of real LOB data, which are not fitted parameters of the model. The counterfactual axis is a stated conditioning protocol (past volume history, time of day, target future liquidity profile) using a standard diffusion-model conditional sampling formulation; no equation or fitted value is renamed as a prediction. The downstream axis reports that synthetic counterfactual data improves a future-liquidity forecaster; this is an empirical claim evaluated on real outcomes, not a quantity that DiffVolume's own training objective forces. There is no evidence in the abstract of the forecaster being trained and tested on overlapping data from the generator, no self-citation, and no uniqueness theorem invoked from the authors' prior work. Because the central claims are benchmarked externally and the conditioning inputs are stated explicitly, there is no circularity to flag at this level of evidence.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

With only the abstract readable, the ledger records the model's fitted parameters and the domain assumptions the abstract itself reveals. The diffusion model's trained weights and conditioning embeddings are fitted to LOB data, and the abstract does not report hyperparameters or architecture. The paper postulates that a future liquidity profile can serve as a conditioning input for counterfactual sampling, which is the main ad hoc assumption, and it equates realism with three summary statistics. No invented physical entities appear; DiffVolume is a named model, not a new force, particle, or dimension.

free parameters (2)

Trained diffusion network weights and conditioning embeddings (volume history, time of day, liquidity profile) = not disclosed in the abstract
The model is trained end-to-end on LOB data; the abstract reports no architecture, noise schedule, conditioning scale, or training details.
Realism evaluation statistics (marginal distribution binning, spatial correlation lags, autocorrelation lags) = not disclosed
The abstract claims each statistical property is reproduced 'better' without specifying the distance metric, lags, or thresholds used to measure reproduction.

assumptions (3)

domain assumption A denoising diffusion model can learn the high-dimensional joint distribution of LOB volume snapshots across price levels
The entire method rests on the diffusion model's capacity to represent the heavy-tailed, strongly correlated volume distribution; this is asserted by construction, not proven.
ad hoc to paper A target future liquidity profile is a valid conditioning input for counterfactual volume generation
Conditioning on a future outcome the model is supposed to generate inverts the causal flow of order book dynamics; the abstract provides no argument that forced future conditions stay on the learned manifold.
domain assumption Realism is adequately measured by marginal distribution, spatial correlation, and autocorrelation decay
The abstract equates realism with these three statistics; summary-statistic matching does not guarantee realistic joint dynamics or absence of mode collapse.

how reviews work

0 comments

Cite this review

Pith. "Pith review of DiffVolume: Diffusion Models for Volume Generation in Limit Order Books." pith.science (2026). https://pith.science/paper/LMOZUEXZ

@misc{pith2026250808698,
  author       = {Pith},
  title        = {Pith review of: DiffVolume: Diffusion Models for Volume Generation in Limit Order Books},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LMOZUEXZ}},
  note         = {Machine review of arXiv:2508.08698}
}

read the original abstract

Modeling limit order books (LOBs) dynamics is a fundamental problem in market microstructure research. In particular, generating high-dimensional volume snapshots with strong temporal and liquidity-dependent patterns remains a challenging task, despite recent work exploring the application of Generative Adversarial Networks to LOBs. In this work, we propose a conditional \textbf{Diff}usion model for the generation of future LOB \textbf{Volume} snapshots (\textbf{DiffVolume}). We evaluate our model across three axes: (1) \textit{Realism}, where we show that DiffVolume, conditioned on past volume history and time of day, better reproduces statistical properties such as marginal distribution, spatial correlation, and autocorrelation decay; (2) \textit{Counterfactual generation}, allowing for controllable generation under hypothetical liquidity scenarios by additionally conditioning on a target future liquidity profile; and (3) \textit{Downstream prediction}, where we show that the synthetic counterfactual data from our model improves the performance of future liquidity forecasting models. Together, these results suggest that DiffVolume provides a powerful and flexible framework for realistic and controllable LOB volume generation.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

ROD: RGB-Only Fast and Efficient Off-road Freespace Detection
cs.CV 2025-08 unverdicted novelty 3.0 of 10

An RGB-only ViT-plus-lightweight-decoder pipeline claims state-of-the-art off-road freespace detection on ORFD and RELLIS-3D at 50 FPS; the claim cannot be verified because the manuscript is unreadable.

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.