{"id":"8b5cc9f2-a2ca-4397-b9c1-258fe91e42c3","arxiv_id":"2505.08180","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A deep network trained jointly across S&P 500 stocks, using order-book predictors and volume commonality, roughly doubles out-of-sample predictability of intraday volume versus the CMEM benchmark.","lead":"Machine-learning models with order-book features forecast 15-minute U.S. equity trading volume with out-of-sample R2 of 0.624, versus 0.265 for the standard CMEM benchmark. A simulated VWAP execution study suggests these forecasts cut tracking error by an average of 28.7% on five sampled stocks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline R2 gain (0.265 to 0.624) compares CMEM against ML models evaluated only in a dynamic one-bin-ahead setting; static R2 is never reported for the ML models, so the claimed predictability may largely reflect same-day information rather than ex-ante forecasting skill.","rationale":"Both the forecasting headline and the VWAP application in Section 6 rely on the R2 values in Table 3. The reader's conditional verdict identified the static versus dynamic asymmetry as the weakest point, and I agree. Among the candidate concerns, including numeric inconsistencies, missing hyperparameters, and the five-stock economic evaluation, this asymmetry is the most load-bearing because it directly determines whether the central claim of high predictability is true in the ex-ante sense that the abstract implies. The numeric inconsistencies, such as Section 5.1 text reporting UAM linear R2 0.549 while Table 3 shows 0.459, and Appendix A.5 text reporting 0.420 and 0.408 while Table 8 shows 0.566 and 0.521, are serious but could be corrected in revision; the dynamic or static issue affects the interpretation of the headline number itself. A static evaluation of the ML models is a feasible and well-defined check that would settle the concern. I am not rejecting the paper: the dynamic setting is legitimate for execution algorithms, and within that setting the comparison between CMEM and ML is fair. However, the abstract and conclusion do not carry the necessary caveat, so the verdict remains conditional pending the static evaluation and the resolution of the numeric inconsistencies.","tokens_in":22030,"tokens_out":10397,"duration_ms":108475,"concrete_test":"Reproduce Table 3 under the static setting of Appendix A.2 for UAM XGB and UAM DeepLOBv: for each test day, train on data through the previous day, construct features using only information available before the open (no past_2, past_8, or other same-day realized values), predict all 26 bins, and compute mean out-of-sample R2. Compare with static CMEM (0.240) and with the reported dynamic values (0.622 and 0.624). If static ML R2 falls to roughly 0.3 to 0.4, or if the dynamic-minus-static gap for the ML models greatly exceeds CMEM's own gap of 0.265 minus 0.240 equals 0.025, the headline improvement is mainly a dynamic-updating artifact and the abstract's ex-ante framing should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.1 and Table 3 support the central claim with R2 values: CMEM 0.265 and UAM DeepLOBv 0.624. Appendix A.2 defines two settings: static (forecast all 26 bins of day t+1 using only information through day t) and dynamic (one-bin-ahead, using all information up to the immediately preceding bin). The paper reports static R2 only for CMEM (0.240) and then uses the dynamic CMEM (0.265) as the benchmark in Table 3; every ML model is evaluated in the dynamic setting only. This is not a cosmetic asymmetry. In dynamic prediction, the feature set includes realized same-day aggregates via operations past_2 and past_8 (Section 3.3, Table 2), so the model is effectively nowcasting the next bin conditional on the previous 30 to 120 minutes of realized volume. Given the strong autocorrelation documented in Appendix A.3, high R2 in this setting is expected. The gain from 0.265 to 0.624 therefore conflates the value of the ML model, nonlinearities, and commonality with the value of conditioning on same-day realized volume, and the paper provides no static R2 for the ML models to separate these. The abstract's claim that intraday stock trading volume is highly predictable is stated without this caveat. If static ML R2 is close to static CMEM R2 of 0.240, the central contribution shrinks from a large forecasting improvement to a within-day adaptive-scheduling improvement, which is a different and weaker claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a machine-learning framework for forecasting 15-minute intraday trading volumes of U.S. equities. Using LOBSTER order-book data for 469 S&P 500 constituents from July to December 2017, the authors augment CMEM decompositions with high-frequency features, train linear models (OLS/LASSO/Ridge), XGBoost, and a DeepLOB-based neural network under three schemes (single-asset, clustered, universal), and evaluate out-of-sample R2. They report raising R2 from 0.265 (CMEM) to 0.624 (UAM DeepLOBv) and further find that using these forecasts in VWAP replication reduces tracking error by 28.7% on average and improves passive-order fill ratios relative to CMEM.","tokens_in":22209,"tokens_out":5807,"duration_ms":52542,"significance":"If the reported forecasting gains survive a fair static-versus-dynamic comparison, the paper would provide substantial evidence that intraday volume is highly predictable given order-book features and cross-sectional commonality, and it would offer a practical execution tool. The study is carefully structured: genuine out-of-sample evaluation against an external benchmark, an explicit comparison of three training schemes, per-date and per-sector stability analysis, and a concrete economic application with a limit-order-book matching engine. The qualitative ordering across models in Table 3 is internally consistent, and the feature-importance analysis adds transparency. The dynamic-evaluation confound in the headline R2 comparison is, however, a load-bearing unresolved issue.","major_comments":[{"comment":"The comparison underpinning the headline result is confounded by the evaluation setting. Static out-of-sample R2 is reported only for the CMEM benchmark (0.240), while every machine-learning model in Table 3 is evaluated in the dynamic one-bin-ahead setting (0.265 for CMEM). Because the dynamic setting uses same-day realized volume through the past_2 and past_8 operations (Section 3.3, Table 2), the reported gain from 0.265 to 0.624 combines model improvement with the value of conditioning on already-realized bins of the target day. Please report static R2 for all models (or, at minimum, for the UAM DeepLOBv model) under identical feature sets, and restate the abstract's predictability claim in light of that comparison.","section":"Section 5.1, Table 3, Appendix A.2"},{"comment":"The out-of-sample R2 statistic is never defined. The paper reports means and standard deviations across dates, which suggests a per-day computation, but the formula, the benchmark used in the denominator, and the pooling convention are all unspecified. Without a precise definition (e.g., 1 - MSE_pooled/Var or a per-day average), the reader cannot determine what R2 = 0.624 measures or compare it with values in the literature. Please add the exact definition, ideally as an equation in Section 5.","section":"Section 5.1"},{"comment":"Several prose statements contradict the tables. Section 5.1 states that 'in UAM, the R2 of the linear model is 0.549,' but Table 3 reports UAM RIDGE as 0.459. Appendix A.5 states that CNN+LSTM achieves R2 of 0.420 and MLP+LSTM 0.408, while Table 8 reports 0.566 and 0.521, respectively. Please correct these numbers and scan the paper for other such inconsistencies.","section":"Section 5.1; Appendix A.5"},{"comment":"The economic value claims (tracking-error reduction and fill-ratio advantage) rest on the matching-engine replay of Frey et al. (2023), which assumes best-quote passive orders, converts unfilled quantity to market orders at the end of each bin, and ignores the price impact of the parent order. These assumptions are strong for a 1% participation rate; the paper should state them as limitations and, ideally, provide a sensitivity analysis, e.g., with alternative participation rates or queue-position assumptions, before claiming that 'precise intraday forecasting offers substantial advantages.'","section":"Section 6 and Appendix B"}],"minor_comments":[{"comment":"There are several typographical errors, including 'and and employing' in Section 2, 'thecross-sectional' in Section 3.1, and the heading 'Nonliner Models' in Section 4.3.2; these should be corrected throughout.","section":"Section 2; Section 3.1; Section 4.3.2"},{"comment":"The notation in Equations (13) and (14) is garbled: '\\volumt,i(s)' and '\\volume(d)t,i' should be written with consistent hat notation, e.g., \\hat{v}^{(s)}_{t,i} and \\hat{v}^{(d)}_{t,i}, and the definitions clarified.","section":"Appendix A.2, Equations (13) and (14)"},{"comment":"The caption of Figure 8 and the body text disagree on whether the baseline is CMEM or the Rolling Mean (RM); please align the description to avoid ambiguity about which benchmark the 'advantage' is measured against.","section":"Figure 8 and surrounding text"},{"comment":"The claim to 'pioneer the identification of commonalities within intraday volume forecasting' is too strong given that the training schemes are drawn from Djanga et al. (2023) and commonality is a theme in Zhang et al. (2024); please temper this statement or clarify the specific new contribution.","section":"Section 7, Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' own prior work (Djanga et al. 2023; Zhang et al. 2024; Frey et al. 2023) for the training schemes and the matching engine; while this is not disqualifying, the incremental novelty relative to those papers should be made explicit in revision. The sample covers only six months of 2017, and the robustness of the results to different sample periods or market regimes is not addressed; the editor may wish to request such an analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: I think this paper deserves a proper referee, but only after the authors resolve an evaluation asymmetry and a set of number discrepancies. The central empirical direction is believable: for intraday volume prediction, nonlinear models with order-book features and cross-sectional pooling beat the CMEM baseline. The paper does a real service by pulling together a wide set of predictors, testing SAM/CAM/UAM schemes, and adding a VWAP execution evaluation. The qualitative ordering across tables is consistent, and the appendix honestly notes the correlation-matrix question needs more work.\n\nThe soft spot that matters most is the dynamic-versus-static comparison. All ML models are evaluated one-bin-ahead, using features like past_2 and past_8 that carry same-day realized volume, while static R2 is only given for CMEM (0.240). So the jump from 0.265 to 0.624 mixes model quality with the extra information that comes from updating within the day. The paper's footnote explains why dynamic prediction was chosen, but it doesn't give the reader the static R2 for the ML models. Without that, the abstract's 'highly predictable' claim is too strong.\n\nThere are also plain number inconsistencies. Section 5.1 says UAM RIDGE is 0.549 and CAM RIDGE is 0.577, while Table 3 shows 0.459 and 0.557. Appendix A.5 text gives 0.420 and 0.408, while Table 8 shows 0.566 and 0.521. Appendix A.7 text and Table 9 disagree on the 50-cluster result. These may be copy-paste slips, but they make it hard to trust any specific number without checking.\n\nI'd also like code or data. The economic claim rests on five stocks and a simulated matching engine; the caveats are stated, but the fill-ratio results are less robust than the forecasting results.\n\nOn balance, I'd accept it for review. The problem is real, the experiments are extensive, and the qualitative conclusions are likely right. But the revision should include static R2 for the ML models and a full number audit. Not a desk reject, but a conditional accept with teeth.","headline":"Good empirical work on a neglected execution problem, but the headline predictability gain is conflated with same-day information, and the numbers need a careful audit.","tokens_in":22918,"tokens_out":2484,"would_cite":false,"duration_ms":24579,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that 15-minute equity trading volume is highly predictable once order-book features, cross-stock pooling, and nonlinear models are combined.","keywords":["intraday trading volume","machine learning","commonality","limit order book","VWAP","execution scheduling","out-of-sample R2","S&P 500"],"falsifier":"Re-estimate Table 3 under static prediction, forecasting all 26 bins of day $t+1$ from data up to day $t$ for every machine-learning model, and compare the resulting out-of-sample $R^2$ with CMEM's static value of 0.240; if the ML advantage largely disappears, the dynamic-updating information is the real driver.","tokens_in":21616,"feed_emoji":"📈","tokens_out":7935,"duration_ms":74111,"temperature":0.7,"pith_summary":"This paper tries to establish that the next 15-minute trading volume of a U.S. equity is highly predictable, and that the predictability is exploitably large. Starting from the standard component multiplicative error model with an out-of-sample $R^2$ of 0.265, the authors add high-frequency limit-order-book predictors, share information across stocks through clustered or universal models, and switch from linear to nonlinear learners; the best configuration, a DeepLOB-style network trained on all stocks, reaches $R^2=0.624$. They then show that these forecasts cut VWAP replication tracking error by 28.7% on average across five sampled stocks and bring passive-order fill ratios closer to what a perfect-volume oracle would achieve. The paper's claim is that intraday volume is no longer a hard-to-forecast univariate time series once this information is used.","feed_headline":"ML lifts 15-minute volume forecast R² from 0.265 to 0.624","feed_subtitle":"Order-book features, cross-stock pooling, and nonlinear models make the next 15-minute trading volume highly predictable.","key_machinery":"The load-bearing object is the feature-and-pooling stack built around the CMEM decomposition. Each bin is represented by 54 predictors: seven CMEM components (daily, periodic, non-periodic, and their products) plus high-frequency auxiliary predictors such as buy and sell notional, numbers of trades and shares, and compound sums over the previous two bins, eight bins, the same intraday interval, and the full day. These features are fed into models under three schemes: single-asset (SAM), cluster-of-similar-assets (CAM, with PCA plus K-means++ on volume or feature correlations), and universal pooled (UAM). The final stage is DeepLOBv, a convolutional-Inception-LSTM network that extracts nonlinear temporal and cross-sectional structure from the features. The machinery works by giving a flexible learner enough same-day and cross-sectional information to override the fixed U-shape that limits the benchmark.","core_discovery":"The central discovery is a measured progression in out-of-sample $R^2$ for 15-minute volume forecasts on 469 S&P 500 constituents. The CMEM benchmark scores 0.265; adding CMEM's decomposition components as linear features gives a Ridge model 0.379; adding the auxiliary order-book predictors raises Ridge to 0.491; and nonlinear models with the same features reach 0.532 for XGBoost and 0.566 for DeepLOBv in per-stock training. Pooling stocks through a universal model lifts XGBoost to 0.622 and DeepLOBv to 0.624, while a clustered model gives Ridge 0.557 and XGBoost 0.600. The authors attribute the gains to three mechanisms: auxiliary predictors correct CMEM's rigid U-shaped pattern on days when volume deviates from it, cross-sectional pooling exploits commonality in volume across stocks, and nonlinear learners capture herding, overreaction, and spillover effects that linear models miss. The headline comparison uses dynamic one-bin-ahead prediction, where forecasts consume volume realized earlier the same day.","pith_inferences":["A natural extension the paper leaves implicit is to test the same feature stack in the static setting, where all 26 bins must be forecast before the day opens; the result would separate genuine next-day skill from the real-time updating advantage.","The commonality result suggests the clustered and universal training schemes could transfer to smaller or less liquid universes where individual volume series are too noisy for per-stock models, because pooled training acts as a regularizer.","The matching-engine evaluation assumes the parent order does not move prices; in practice, larger order sizes would erode the fill-ratio benefit, so the 28.7% tracking-error reduction and fill-ratio improvements are likely upper bounds for real trading.","Because the sample covers only July through December 2017, an out-of-period replication under different volatility and liquidity regimes would show whether the $R^2=0.624$ result is stable or period-specific."],"forward_implications":["If the headline $R^2$ holds, execution desks can schedule 15-minute child orders with substantially more confidence than CMEM allows, because the model explains over 60% of out-of-sample variation rather than about a quarter.","Cross-sectional commonality is economically usable: training one nonlinear model on all stocks outperforms per-stock training, and cluster-level training outperforms the linear universal model, so volume forecasts can be improved without per-stock manual tuning.","Auxiliary predictors are a large lever on their own: adding them to a Ridge model raises $R^2$ from 0.379 to 0.491, so the information content of order-book features is valuable even before nonlinearity or pooling is introduced.","The VWAP application directly converts forecast skill into a cost metric: the best model reduces tracking error by an average of 28.7% on the five sampled stocks, with the largest single-stock drop being 57.0% for AEP.","Because the best model uses cross-stock pooled features, the method should also deliver stable forecasts for stocks that have short or noisy individual volume histories, a benefit of the commonality finding."],"supporting_citations":[{"why":"It supplies the CMEM benchmark and its daily-periodic-nonperiodic decomposition that the paper improves on.","marker":"Brownlees et al. (2010)"},{"why":"It provides the LOBSTER limit-order-book data from which all auxiliary high-frequency predictors are computed.","marker":"Huang and Polak (2011)"},{"why":"It introduces the SAM, CAM, and UAM training schemes used to exploit commonality across stocks.","marker":"Djanga et al. (2023)"},{"why":"It contributes the DeepLOB CNN-Inception architecture that DeepLOBv adapts for volume forecasting.","marker":"Zhang et al. (2019)"},{"why":"It motivates the logarithmic and additive treatment of CMEM components and supplies the VWAP tracking-error definition used in the backtest.","marker":"Chen et al. (2016)"},{"why":"It supplies the matching engine used to estimate passive-order fill ratios in the economic evaluation.","marker":"Frey et al. (2023)"},{"why":"It establishes the dynamic versus static prediction settings and the cross-sectional volume-matrix estimation that the experimental design follows.","marker":"Białkowski et al. (2008)"}],"fun_headline_variants":["ML lifts 15-min volume R² from 0.265 to 0.624","Order-book features and deep learning: R² 0.624 for volume","Cross-stock pooling boosts intraday volume R² to 0.624","Nonlinear ML hits R² 0.624 in volume forecasting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline improvement is measured with one-bin-ahead forecasts that use volume already realized on the target day, so if the models were required to forecast the whole day before it opens, the gain over CMEM, whose static $R^2$ is only reported as 0.240, could be much smaller.","fun_headline_variants_meta":{"raw":{"variants":["ML lifts 15-min volume R² from 0.265 to 0.624","Order-book features and deep learning: R² 0.624 for volume","Cross-stock pooling boosts intraday volume R² to 0.624","Nonlinear ML hits R² 0.624 in volume forecasting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000612,"raw_usage":{"total_tokens":2813,"prompt_tokens":880,"completion_tokens":1933,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":1849}},"tokens_in":496,"tokens_out":1933,"duration_ms":17516,"temperature":1.0,"reasoning_tokens":1849,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:02:57.593604+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-estimate Table 3 under static prediction, forecasting all 26 bins of day $t+1$ from data up to day $t$ for every machine-learning model, and compare the resulting out-of-sample $R^2$ with CMEM's static value of 0.240; if the ML advantage largely disappears, the dynamic-updating information is the real driver.","supporting_citations":[],"review_version":1}