{"id":"96b8b379-37cc-49ef-a201-923166eae167","arxiv_id":"2412.01062","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"A pruned lightweight CNN with dynamic feature selection is claimed to achieve 10 ms latency and better RMSE and R² on ETF data, but the evidence is incomplete.","lead":"This paper proposes a high-frequency trading pipeline that combines dynamic feature selection with a pruned lightweight neural network, reporting lower error and lower latency than five benchmark methods on ETF data. The claims are undercut by missing code, data, and profitability metrics, so the result is not yet verifiable or actionable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pruning rule in Eq. 7 is inverted relative to magnitude pruning, so the reported latency and accuracy advantages lack an internally consistent mechanism.","rationale":"The reader's verdict was REJECT with high correctness risk, and the reader's rationale did mention that 'the pruning rule appears inverted.' The weakest_assumption field, however, identifies Eq. 3's inverse-variance feature weighting as the load-bearing assumption. My stress-test agrees with the overall rejection, but I find the pruning inversion in Eq. 7 to be the single most load-bearing concern because it directly undermines the mechanism that is supposed to produce the speed advantage, which is the paper's main selling point. If Eq. 7 is a typo, then the manuscript is not a verifiable artifact; if it is not a typo, the described algorithm would remove exactly the weights that carry the convolutional signal and would not be a valid pruning method. Eq. 8's indicator penalty is also inconsistent with inducing sparsity. The paper provides no code, no clear dataset description (the text switches between '50 ETF datasets' and '50ETF dataset'), no training/test split, and no out-of-sample trading results, so there is no external support that could compensate for the internal inconsistency. A single computational check comparing the literal pruning rule to standard magnitude pruning would settle whether the concern lands: if the literal rule destroys accuracy or fails to reduce parameters, the central claim cannot stand. I therefore recommend no change to the reader's REJECT verdict, and I partially agree with the reader's weakest-assumption analysis because the pruning defect is at least as damaging as the Eq. 3 assumption, while the reader focused on the latter.","tokens_in":5024,"tokens_out":2847,"duration_ms":29247,"concrete_test":"Implement the reported pipeline on a public time-series benchmark (or the released 50ETF data if it becomes available) with two pruning rules: (A) standard magnitude pruning, removing weights with |w| ≤ τ; (B) the rule as written in Eq. 7, removing weights with |w| > τ. For each rule, retrain or fine-tune with the stated threshold τ = 0.01 and report parameter count, validation RMSE, R², and inference latency. If rule (B) does not both reduce the parameter count and preserve accuracy close to the unpruned model, the mechanism claimed in the paper is unsupported; if only rule (A) works, the manuscript's Eq. 7 must be corrected before any of the latency or accuracy claims can be evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the pruned lightweight CNN achieves the lowest execution time (35 ms) and latency (10 ms) while outperforming benchmarks in RMSE and R². For this to hold, the pruning step must actually remove low-information parameters and reduce computation. As printed, Eq. 7 defines the pruning set as P_pruned = {w_{i,j} ∈ W | |w_{i,j}| > τ}, i.e., the parameters with large magnitude are designated for pruning. This is the reverse of standard magnitude pruning, where the weights below a threshold are set to zero. If the implementation follows Eq. 7 literally, it removes the large convolutional weights that dominate the response in Eq. 6, destroying the learned features; if instead the intended set was |w_{i,j}| ≤ τ, then Eq. 7 is miswritten and the central mechanism is not reproducible from the manuscript. Equation 8 is also problematic as displayed: the Lagrangian term Σ 1(w_i ≤ τ) penalizes small weights, which would push weights away from the sparse/pruned region rather than toward it. Because the speed advantage rests entirely on this pruning operation, the paper's central experimental claim is not internally consistent as written. The inverse-variance feature weighting in Eq. 3 is likewise unsubstantiated, but the pruning inversion is a more direct defect: even granting every experimental number, the described method cannot produce the claimed lightweight network.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a machine-learning pipeline for real-time high-frequency trading data processing, consisting of a dynamic feature-selection module (k-means clustering with inverse-variance feature weighting, mutual information, and kernel density estimation) and a lightweight CNN with pruning. The authors report experiments on 50ETF data comparing their method ('Ours') with five baselines on RMSE, R², execution time, and latency, claiming the lowest average execution time of 35 ms and latency of 10 ms while maintaining accuracy across market conditions.","tokens_in":5311,"tokens_out":5744,"duration_ms":51317,"significance":"If substantiated, the claimed combination of adaptive feature selection and a pruned low-latency CNN would be a practically relevant contribution to HFT systems. The paper addresses a timely problem and deserves credit for comparing against five named baselines and for reporting both accuracy and latency metrics. However, the manuscript as written does not support these claims: the core equations are garbled, the pruning rule is internally inconsistent, and no out-of-sample evaluation or measurement methodology is presented. The contribution cannot be independently assessed or reproduced from the supplied text.","major_comments":[{"comment":"Equations (1)-(5) are rendered with placeholder '?' symbols in place of all mathematical notation, making the dynamic feature-selection procedure unreadable and unverifiable. Since every experimental claim depends on this procedure, the method as printed cannot be implemented or checked; this alone blocks reproducibility.","section":"III-A, Eqs. (1)-(5)"},{"comment":"Equation (7) defines the pruning set as P_pruned = {w_{i,j} ∈ W | |w_{i,j}| > τ}, which is the reverse of standard magnitude pruning: it selects large-magnitude weights for removal. If implemented as written, the operation would delete the dominant convolutional features in Eq. (6) and destroy the network's learned representation; if the intended condition was |w_{i,j}| ≤ τ, then the equation is miswritten and the mechanism is not reproducible. Equation (8) compounds the inconsistency by penalizing the indicator 1(w_i ≤ τ), which encourages weights to move away from the sparse region rather than toward it. The paper's central speed and lightweight-network claims rest on this defective step.","section":"III-B, Eqs. (7)-(8)"},{"comment":"No train/test split, temporal walk-forward validation, or any held-out period is described. Feature weights, cluster assignments, network weights, and hyperparameters (k=10, τ=0.01, λ=0.1, learning rate 0.001, batch size 128) are all selected on the same data, so the reported RMSE and R² values may simply reflect in-sample fitting. The abstract's claim of 'consistent performance in varying market conditions' is therefore not supported by the experiments as reported.","section":"IV-A and IV-B"},{"comment":"The execution-time and latency results in Table I are reported as single numbers with no hardware/software environment, no measurement repetitions, and no statistical uncertainty. The text refers to 'error range shading' in Figures 2 and 3, but the actual plots are not available in the reviewed version, so the claimed stability and error bounds cannot be checked. In addition, neither the abstract's nor the conclusion's mention of 'revenue enhancement' is backed by any profitability, return, or cost metric in the experiments.","section":"IV-B and Table I"},{"comment":"The feature-weight definition in Eq. (3) is the reciprocal of the weighted within-cluster variance, so features with low intra-cluster variance are treated as most important. The paper provides no theoretical justification, empirical validation, or ablation for this inverse-variance assumption; if it is wrong, the dynamic feature selection will discard predictive features and the reported advantages collapse. This assumption is load-bearing for the proposed method and is left completely untested.","section":"III-A, Eq. (3)"}],"minor_comments":[{"comment":"The introduction says machine learning can formulate predictions and act 'in a matter of seconds,' which is inconsistent with HFT's microsecond-scale latency requirements; rephrase to reflect the actual time scales.","section":"I"},{"comment":"'We selected 50 ETF datasets' and 'The 50ETF dataset contains...' are ambiguous; clarify whether 50 ETFs or the 50ETF contract is used, and describe the feature set precisely.","section":"IV-A"},{"comment":"The text attributes reference [9] to 'Li et al.,' but the reference list gives Xu, Ruohao et al.; correct the attribution.","section":"II"},{"comment":"The notation in Eq. (6) is garbled (e.g., '? represents the size of the convolution kernel' and the bias term expression); define all variables and rewrite the convolution formula cleanly.","section":"III-B, Eq. (6)"},{"comment":"The prediction target (next-tick price, direction, return, or volatility) is never defined; specify it and the exact input features used in the experiments.","section":"IV-B"},{"comment":"The conclusion claims the method offers 'advantages in terms of processing speed and revenue enhancement,' but no revenue-related metric appears in Table I or the text; this claim should be removed or supported.","section":"V"}],"recommendation":"reject","confidential_remarks":"The manuscript has placeholder symbols where mathematical notation should be, the central pruning equation is inverted, and the evaluation does not use any held-out data. The reported findings are not reproducible as written, and correcting these issues would require re-deriving the method and re-running the entire experimental study, which goes beyond a normal revision. I do not see a viable path to acceptance for the current submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper does not hold together as written. The high-level idea—dynamic feature selection feeding a pruned lightweight CNN for low-latency HFT prediction—is a reasonable engineering combination, and the authors cite the right neighborhood of prior work. But the manuscript is too broken to evaluate, and the key pruning rule is inverted.\n\nWhat is actually new: not much in principle. Clustering, mutual information, KDE, CNNs, and magnitude pruning are all standard. The only twist is weighting features by inverse within-cluster variance (Eq. 3), which is an ad hoc variant of variance-based feature selection. Credit where due: the design intent is clear, the component choices are defensible, and the latency/accuracy tradeoff is a real problem worth working on. The paper reads like a sincere attempt to combine known pieces.\n\nThe soft spots are not minor. Equations 1 through 9 have symbols replaced by question marks; as printed, Eq. 7 defines the pruning set as {w | |w| > τ}, which removes the large weights and destroys the network, and Eq. 8 penalizes small weights, pushing in the opposite direction. Even if these are typos, the experiments do not support the claims: no train/test split is described, the dataset is ambiguous, no code or data are shipped, hardware details are absent, and the abstract promises \"revenue enhancement\" but no profitability metric appears anywhere. The feature weighting in Eq. 3 rests on an unstated assumption that low within-cluster variance implies informativeness, which is false for near-constant features. The error shading in the figures has no described protocol behind it, so the \"Ours wins\" story is not independently checkable.\n\nI agree with the stress-test note: the pruning inversion is a load-bearing defect. Even granting every reported number, the described method cannot produce the claimed lightweight network. This is not a case of a small flaw in an otherwise solid paper; it is a central mechanism that is internally inconsistent.\n\nWho is this for? As it stands, no one. A careful reader cannot reproduce the method or verify the results. If the authors fixed the equations, released code and data, and provided an honest out-of-sample evaluation with a real latency measurement, the paper could be worth a second look—but that is a different submission.\n\nRecommendation: desk-reject this version. It does not deserve referee time until the technical writing and experimental protocol are repaired.","headline":"A well-intentioned engineering combo that is too garbled to evaluate; the pruning equation is inverted and the experiments lack any reproducible protocol.","tokens_in":5839,"tokens_out":1843,"would_cite":false,"duration_ms":18503,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A pruned convolutional network with dynamic feature selection keeps high-frequency trading forecasts accurate at 10 ms latency.","keywords":["high-frequency trading","real-time data processing","dynamic feature selection","lightweight neural network","convolutional neural network","model pruning","latency","machine learning"],"falsifier":"On the same 50-ETF data, replace the Eq. 3 reciprocal-variance weights with weights estimated from mutual information alone and re-run the reported comparisons; if RMSE and $R^2$ do not worsen, the reciprocal-variance assumption is not what carries the accuracy, and if they do worsen, that assumption is confirmed as load-bearing.","tokens_in":4815,"feed_emoji":"⚡","tokens_out":10669,"duration_ms":92425,"temperature":0.7,"pith_summary":"This paper is trying to establish that a machine-learning pipeline can satisfy the accuracy and latency demands of high-frequency trading at the same time. It combines a dynamic feature-selection stage, which re-picks input features as market conditions shift, with a deliberately small convolutional network whose weights are pruned to keep inference cheap. The empirical claim is that this pipeline maintains consistent prediction error and explanatory power across changing data windows and feature-selection thresholds, while running at 35 ms average execution time and 10 ms average latency, beating five baseline methods on both accuracy and speed. If correct, the contribution is a practical architecture for adaptive, low-latency forecasting in a setting where a few extra milliseconds can change the profitability of a trade.","feed_headline":"Dynamic features plus a pruned CNN cut trading latency to 10 ms","feed_subtitle":"The pipeline keeps errors low across shifting markets and beats five baselines on speed and accuracy.","key_machinery":"The central machinery has two parts. The first is the dynamic feature selector: it clusters data vectors (improved $k$-means++), assigns each feature a weight equal to the reciprocal of its probability-weighted within-cluster variance, and keeps the features whose mutual information with the prediction target is large, with the joint density estimated online by kernel density estimation. That component is what is supposed to track regime changes and revise the input set in real time. The second is the lightweight predictor: a modular convolutional network with small kernels ($3\\times3$ and $5\\times5$) whose multi-scale outputs are weighted and summed, plus a pruning step meant to remove parameters to cut computation while a Lagrangian term keeps training error low. Together these components carry the paper's claim that accuracy and speed can be achieved without a full-size deep model.","core_discovery":"The paper argues that a complete high-frequency trading forecasting pipeline can be both adaptive and fast when feature selection and model size are co-designed for speed. The dynamic feature selector clusters incoming market data with an improved $k$-means++ rule, scores each feature by the reciprocal of the probability-weighted within-cluster variance, and retains features with large mutual information with the target, estimated online by kernel density estimation. The retained features feed a lightweight network built from small convolutional modules with $3\\times3$ and $5\\times5$ kernels whose outputs are weighted and summed, and pruning keeps the parameter count down. On 50 ETF datasets the method is reported to outperform five comparison methods (DL-AN, VE-FTS, T-TS, DF-QA, RL-EMF) on RMSE and $R^2$ across window sizes and mutual-information thresholds, with the lowest average execution time of 35 ms and latency of 10 ms.","pith_inferences":["The reciprocal-variance weighting in Eq. 3 is an assumption, not a derived result; a direct comparison against mutual-information-only feature ranking would show whether the clustering step is actually responsible for the accuracy gain.","The same recipe of dynamic feature selection plus a pruned small network could be transferred to other latency-sensitive prediction settings, such as online advertising bidding or network traffic classification, where input distributions drift quickly.","The reported 10 ms latency is a system-level figure; decomposing it into feature computation, network inference, and queueing would reveal which component the design actually optimises.","Reproducing the parameter-reduction result requires resolving the pruning criterion as written: Eq. 7 puts weights with $|w| > \\tau$ into the pruning set, the opposite of the usual small-weight pruning, so the intended rule should be confirmed before comparing parameter counts."],"forward_implications":["HFT systems that need to update their feature sets online could run this pipeline inside the decision loop without waiting for a full retraining cycle.","Prediction quality that stays stable when data windows or feature-selection thresholds change would make the system less sensitive to hyperparameter tuning than the compared baselines.","The 35 ms execution / 10 ms latency figures provide a concrete budget: any comparable adaptive feature-selection method must meet this speed to be usable in the same setting.","The small convolutional kernels and pruning imply the model can run on commodity hardware, reducing the infrastructure cost of low-latency forecasting.","If the consistency across market conditions is real, the approach could reduce the frequency of strategy failures during volatile periods."],"supporting_citations":[{"why":"Supplies the adaptive-normalization deep learning baseline (DL-AN) that the proposed pipeline is compared against.","marker":"[6]"},{"why":"Supplies the volatility-estimation functional time series baseline (VE-FTS) used as a benchmark.","marker":"[7]"},{"why":"Supplies the non-stationary Transformer baseline (T-TS) used as a benchmark.","marker":"[8]"},{"why":"Supplies the Q-learning dynamic feature selection baseline (DF-QA) used as a benchmark.","marker":"[9]"},{"why":"Supplies the reinforcement-learning evolutionary metric filter baseline (RL-EMF) used as a benchmark.","marker":"[10]"}],"fun_headline_variants":["Adaptive features and pruned CNN trim HFT latency to 10 ms","Co-designing features and model cuts HFT latency to 10 ms","Lightweight CNN with adaptive features hits 10 ms HFT latency","Pruned CNN plus online feature selection: 10 ms trading latency","Speed-focused HFT pipeline: adaptive features, pruned CNN, 10 ms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that features with low variance inside each cluster are the informative ones, because Eq. 3 defines feature importance as the reciprocal of that weighted variance; the paper offers no independent justification for this weighting, so if it is wrong the dynamic selector can discard precisely the signals the network needs.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive features and pruned CNN trim HFT latency to 10 ms","Co-designing features and model cuts HFT latency to 10 ms","Lightweight CNN with adaptive features hits 10 ms HFT latency","Pruned CNN plus online feature selection: 10 ms trading latency","Speed-focused HFT pipeline: adaptive features, pruned CNN, 10 ms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1261,"prompt_tokens":937,"completion_tokens":324,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":227}},"tokens_in":553,"tokens_out":324,"duration_ms":3320,"temperature":1.0,"reasoning_tokens":227,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:43:18.914323+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the same 50-ETF data, replace the Eq. 3 reciprocal-variance weights with weights estimated from mutual information alone and re-run the reported comparisons; if RMSE and $R^2$ do not worsen, the reciprocal-variance assumption is not what carries the accuracy, and if they do worsen, that assumption is confirmed as load-bearing.","supporting_citations":[{"cited_title":"Deep adaptive input normalization for time series forecasting","cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive-normalization deep learning baseline (DL-AN) that the proposed pipeline is compared against."},{"cited_title":"Forecasting stock index futures intraday returns: Functional time series model","cited_arxiv_id":null,"evidence_quote":"Supplies the volatility-estimation functional time series baseline (VE-FTS) used as a benchmark."},{"cited_title":"Non-stationary transformers: Exploring the stationarity in time series forecasting","cited_arxiv_id":null,"evidence_quote":"Supplies the non-stationary Transformer baseline (T-TS) used as a benchmark."},{"cited_title":"Dynamic feature selection algorithm based on Q- learning mechanism","cited_arxiv_id":null,"evidence_quote":"Supplies the Q-learning dynamic feature selection baseline (DF-QA) used as a benchmark."},{"cited_title":"Reinforcement Learning based Evolutionary Metric Filtering for High Dimensional Problems","cited_arxiv_id":null,"evidence_quote":"Supplies the reinforcement-learning evolutionary metric filter baseline (RL-EMF) used as a benchmark."}],"review_version":1}