{"id":"e4187725-889d-4596-809e-6ded72a342c2","arxiv_id":"2412.16160","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"The paper combines two feature-importance methods with automatic k-means clustering to forecast limit order book mid-prices with a radial basis function network.","lead":"This paper tests a stock-price forecasting system for high-speed trading that automatically chooses which market features matter and how to group them, instead of using a human expert. If it works, it could speed up and simplify machine-learning trading models, but the study lacks a comparison against simpler alternatives.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that autonomous clustering 'enhances forecasting ability' is unsupported because no baseline comparison to a fixed-cluster or standard RBFNN is reported, a limitation the paper itself admits in Section IV.C.","rationale":"The reader's weakest-assumption pick (isotropic clusters) is a real technical limitation, but it is not the most load-bearing. The central claim is an enhancement claim, and the experiments contain no ablation or baseline that could establish enhancement. The paper itself states the missing benchmark in Section IV.C, which per the reviewing rule must be flagged. Even if the isotropic assumption were fixed, the headline claim would still be unverified without a control group. I therefore agree with the reader's REJECT verdict, but for a different primary reason. The concrete test I propose is a direct comparison against a fixed-cluster RBFNN, which would settle whether the autonomous selection actually improves forecasting. My recommendation of UNCHANGED reflects that the reader's rejection remains appropriate, with the missing baseline as the decisive weakness.","tokens_in":19661,"tokens_out":4043,"duration_ms":36812,"concrete_test":"Re-run the protocol on the same 20-stock dataset per month with three baselines per stock: (1) a standard RBFNN using a fixed number of clusters (e.g., K=2 and K=5) set a priori; (2) k-means with cluster count selected by the elbow method; (3) a simple linear regression or random-forest regressor on the same feature sets. Compare test RMSE/RRMSE distributions between the proposed autonomous method and each baseline using a paired Wilcoxon signed-rank test across the 20 stocks. If the autonomous method does not significantly outperform the fixed-cluster RBFNN baseline, the claim that automation 'enhances forecasting ability' is not supported; if it does, the isotropic-cluster and implausible-value concerns become the main issues to investigate next.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the dual feature-importance mechanism and silhouette-driven k-means 'enhances the forecasting ability of the RBFNN regressor' and removes the need for manual topological grid search (Abstract and Section V). The experiments, however, only compare two feature-importance methods (MDI vs GD) under two feature sets (Simple vs Extended). There is no comparison to a standard RBFNN with a fixed or manually chosen number of clusters, no elbow-method baseline, and no other forecasting benchmark. Without such a control, the reported MSE/RMSE values are purely descriptive; they cannot support the causal claim that automation improves forecasting. The paper explicitly acknowledges this in Section IV.C: 'A further limitation is the lack of an extensive benchmark modeling framework to provide a more rigorous challenge to the existing RBFNN topology.' This is the most load-bearing gap because it defeats the headline assertion. Secondary concerns (e.g., the isotropic-cluster assumption acknowledged in Section IV.C, or implausible BRK RMSE values around 10^4) would only matter if the enhancement claim were established, but it is not. A method can be internally consistent and yet fail to demonstrate the stated benefit if no comparison is made.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an autonomous machine-learning pipeline for high-frequency trading (HFT) mid-price forecasting. The pipeline combines two feature-importance methods, mean-decrease impurity (MDI) from random forests and a gradient-descent (GD) based importance vector, applies a correlation-to-distance transformation, selects the number of k-means clusters via silhouette scores and a quality ratio, and feeds the resulting centroids and spreads into a radial basis function neural network (RBFNN) regressor. The method is tested on Level 1 limit order book data for 20 S&P 500 mega-cap stocks over three months (September--November 2022), with predictions reported per trading event under an expanding-window online setting. The headline claim is that automating feature importance and clustering enhances RBFNN forecasting ability and removes the need for manual topological grid search. The empirical section compares MDI versus GD under two feature sets (Simple and Extended) and reports MSE, RMSE, and relative RMSE (RRMSE) tables.","tokens_in":19906,"tokens_out":4798,"duration_ms":43422,"significance":"If the central claim were established, the paper would contribute a practical, hands-off configuration method for RBFNN-based LOB forecasting, with potential value for HFT practitioners and researchers. The proposed pipeline is modular and reproducible in principle, and testing on 20 real stocks with nanosecond-resolution data is a useful scale of evaluation. However, the paper's own limitation statement in Section IV.C acknowledges the absence of a benchmark modeling framework, and the experiments do not include any comparison to a standard fixed-cluster RBFNN, an elbow-method baseline, or a naive forecasting rule. Consequently, the evidence presented does not yet support the causal claim that automation improves forecasting performance. The identification of a potential data inconsistency in the RRMSE results for BRK and the mis-specified RBF spread formula in Eq. (15) further weaken confidence in the reported numbers. With additional control experiments and corrections, the underlying idea could still be valuable, but the current manuscript is not ready for publication.","major_comments":[{"comment":"The central claim that the autonomous feature-importance and clustering mechanism 'enhances the forecasting ability of the RBFNN regressor' is not supported by the experimental design. The results in Tables III–VI compare only two feature-importance methods (MDI vs GD) under two feature sets (Simple vs Extended). There is no comparison against a fixed-cluster RBFNN, a manually selected number of clusters, an elbow-method baseline, or any standard forecasting benchmark such as a persistence model. The authors explicitly acknowledge this in Section IV.C: 'A further limitation is the lack of an extensive benchmark modeling framework to provide a more rigorous challenge to the existing RBFNN topology.' Without such control experiments, the reported MSE/RMSE/RRMSE values are descriptive only and cannot establish the causal benefit of automation. Please add at least a fixed-cluster RBFNN and a persistence baseline, and report performance differences with appropriate significance testing.","section":"Section IV.B and Section IV.C"},{"comment":"The RRMSE values reported for BRK are implausible and suggest a normalization error. For BRK September 2022, Table V gives an MDI Simple test RMSE of 8.758E+03, while Table VII reports the corresponding RRMSE as 8.010E+03. If RRMSE is defined as RMSE divided by the mid-price, this implies a mid-price of roughly $1.09, which is inconsistent with a $200B+ mega-cap stock such as Berkshire Hathaway (whether Class A or Class B). Similar inconsistencies appear across all BRK rows. Please verify the RRMSE calculation and clarify what mid-price value is used in the denominator; if the price scale is transformed, that transformation must be stated explicitly.","section":"Section IV.B / Table VII"},{"comment":"Equation (15) is mis-specified as a per-neuron spread. The right-hand side of the equation does not depend on the index j, so the formula assigns the same value to every σ_j. This contradicts the notation σ_j and undermines the construction of the RBF basis functions in Eq. (14), where each center should have its own scale. Either σ should be defined once as a global hyperparameter, or the formula should be corrected to depend on the specific center µ_j (e.g., distance from µ_j to its nearest neighbors). The current form will bias all RBF neurons identically and may degrade the regressor.","section":"Section III, Eq. (15)"},{"comment":"The isotropic-cluster assumption of k-means is load-bearing but untested. The RBF centers and spreads in Eqs. (14)–(15) are derived directly from k-means centroids, and the assignment rule in Eq. (12) uses Euclidean distance. The authors acknowledge in Section IV.C that 'we made the assumption of isotropic clusters within the k-means algorithm, which presumes that every cluster exhibits constant variance.' If the transformed feature clusters are anisotropic or elongated, the basis functions are mis-specified and the claimed autonomous-clustering benefit may not hold. Please provide a diagnostic, such as a comparison of within-cluster covariance eigenvalues, or evaluate an alternative that uses Mahalanobis distances or a Gaussian mixture model, to assess whether the assumption is empirically reasonable.","section":"Section III and Section IV.C"}],"minor_comments":[{"comment":"The block numbering is inconsistent: the text says 'The silhouette and quality scores are part of Block 4' after Block 3 has already been described as defining the optimal number of clusters via silhouette scores. Please clarify the relationship between Blocks 3 and 4.","section":"Section III"},{"comment":"There is a typo in Eq. (1): the average target value is written as '˜yt' but should be '˜yj' (the average target value for node j).","section":"Section III, Eq. (1)"},{"comment":"The text near Eq. (3) contains 'tress' instead of 'trees' in 'the total number of tress in RF'.","section":"Section III, Eq. (3) surrounding text"},{"comment":"The company name for BRK is listed as 'BERKSHIRE HATHAW AY INC.'; please correct the typo and clarify whether this refers to Class A or Class B shares, as the price scale matters for interpreting RRMSE.","section":"Table II"},{"comment":"The description 'cumulative five-fold setting' is unclear. Please specify how the training/test partitions are constructed, whether the same testing windows are reused across folds, and how the sliding-window overlap of 99 out of 100 events affects the independence of the reported test errors.","section":"Section IV.A"},{"comment":"The claim that this is the first autonomous feature-importance and clustering protocol in the HFT domain is stated without comparison to prior automated ML pipelines. Please soften the novelty claim or provide a more systematic literature comparison.","section":"Abstract and Section V"}],"recommendation":"major_revision","confidential_remarks":"The paper has a clear and potentially useful idea, and the data scale (20 stocks, three months, tick-by-tick) is respectable. However, the empirical evaluation is not yet at the standard required to support the headline claim. The missing baseline comparison is the most serious issue; without it, the paper is a description of a pipeline rather than a demonstration of benefit. The BRK RRMSE inconsistency suggests that the reported metrics may contain a calculation or normalization bug, and the Eq. (15) spread formula is technically wrong as written. These are fixable, but the authors should be asked to provide code/data or detailed derivations to verify all reported numbers. I would also encourage the editor to ensure that the authors address the self-cited unpublished reference [34] transparently, since the GD feature-importance method is central to the pipeline and is drawn from that preprint."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick read on arXiv:2412.16160. The paper combines MDI and gradient-descent feature importance, then uses silhouette-quality k-means to set the number of RBF centers for tick-level mid-price forecasting on 20 S&P 500 names. The arrangement is new as a package, and the empirical setup is real: three months of Level 1 LOB data, per-stock per-month tables, two feature sets. That is honest work and worth a look if you care about automating topology choices in this niche.\n\nThe problem is the headline claim. The abstract says the approach \"enhances the forecasting ability of the RBFNN regressor.\" The experiments only compare MDI against GD, under Simple vs Extended features. There is no baseline with a fixed cluster count, no elbow-method alternative, no standard RBFNN, no linear or MLP comparator. So we learn which of two feature-weighting tricks does better on a given stock, but nothing about whether the automation itself helps. The authors admit this in Section IV.C: \"the lack of an extensive benchmark modeling framework.\" That is the load-bearing gap, and it is acknowledged rather than fixed.\n\nTwo softer issues. First, the isotropic-cluster assumption inside k-means is stated in IV.C but never tested; since the RBF spreads in Eq. 15 come directly from those centroids, this is not a cosmetic detail. Second, the reported RRMSE for BRK is around 8,000 in Table VII, which cannot be right if RMSE is divided by mid-price for a stock trading near $400k. Something in the units or the calculation is off, and it is never addressed. That kind of unreconciled number makes me worry about the rest of the data handling.\n\nIs the paper worth refereeing? I think yes, with a clear mandate. The idea is sensible, the writing is mostly clear, and the admitted limitation is repairable. A revision that adds a fixed-cluster control, a non-RBF baseline, and fixes the BRK anomaly could turn this into a decent incremental contribution. As it stands, the central claim is unsupported, so my own verdict would be reject or major revision, not accept.\n\nRecommendation: send to peer review, but the reviewers should insist on the missing baseline rather than letting the current comparison stand as evidence.","headline":"A plausible but unproven automation recipe: the paper never compares its autonomous clustering against a fixed-cluster RBFNN, so the central 'enhances forecasting' claim rests on no control group, even though the authors openly admit this gap.","tokens_in":20463,"tokens_out":1722,"would_cite":false,"duration_ms":18183,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","68T07","91G60"],"pacs":[],"model":"deepseek-v4-flash","headline":"Automating feature importance and cluster count improves RBF neural-network forecasts of order-book mid-prices for 20 large US stocks.","keywords":["high-frequency trading","limit order book","mid-price forecasting","radial basis function network","k-means clustering","feature importance","online learning","mean decrease impurity"],"falsifier":"For the same 20 stocks, compute the ratio of the largest to the smallest eigenvalue of the within-cluster covariance matrix for the features that get clustered; if many clusters show large eigenvalue ratios, the isotropy assumption is violated. Then rerun the pipeline with per-cluster covariance matrices replacing the scalar $\\sigma_j$ (or with Gaussian mixture clustering) and check whether test RMSE changes. If anisotropic clustering does not alter the errors, the isotropy assumption is not load-bearing; if it improves them significantly, the paper's assumption masks the true mechanism.","tokens_in":19450,"feed_emoji":"📈","tokens_out":5480,"duration_ms":49284,"temperature":0.7,"pith_summary":"This paper claims that the two most manual steps in a machine-learning forecasting pipeline for high-frequency trading—deciding which order-book features matter and choosing how many clusters the input should form—can be fully automated. The proposed protocol runs two competing feature-importance mechanisms, mean-decrease impurity from random forests and a gradient-descent-based importance vector, weights the input matrix with the winner, converts it to a correlation-distance matrix, and lets silhouette scores pick the number of k-means clusters online. Those clusters feed the radial basis function network directly, supplying the centers and widths of its neurons, so no topological grid search is needed. Tested on tick-by-tick Level 1 order book data for 20 large S&P 500 stocks over three months, the claim is that this automation improves the RBFNN's mid-price forecasts and that each stock needs its own feature space.","feed_headline":"Automated feature ranking and clustering forecast stock mid-prices","feed_subtitle":"Dual MDI/GD importance plus silhouette-driven k-means removes manual RBF topology search for 20 S&P 500 stocks.","key_machinery":"The load-bearing object is the coupling between k-means clusters and radial basis function neurons: Eq. (14) places an RBF neuron at each k-means centroid $\\mu_j$, and Eq. (15) sets the width $\\sigma_j$ as a scaled average of pairwise centroid distances. That coupling is what makes the network topology autonomous, because the silhouette-quality ratio $q = E[S_i]/\\sqrt{V[S_i]}$ decides the number of clusters, and hence the number of neurons, per trading window. The dual feature-importance mechanism (MDI and GD) weights the raw features before clustering, and the correlation-distance transformation in Eq. (6) gives k-means the metric it needs. These pieces together claim to remove manual topological search for the RBFNN regressor.","core_discovery":"On the paper's own terms, the discovery is that an autonomous protocol can replace the manual choices in RBFNN forecasting for the limit order book. Each sliding window of 100 events (99 overlapping) is transformed into two weighted input matrices, one by MDI and one by GD. The matrix is expressed as a correlation-based distance matrix, k-means is run over candidate cluster counts, and the ratio of mean to variance of silhouette coefficients selects the number of clusters. The winning centroids become RBF centers and the average pairwise centroid distance becomes the RBF width, and the output weights are computed by least squares. On 20 mega-cap US stocks from September to November 2022, the protocol reports per-event MSE, RMSE, and RRMSE, with MDI on the simple four-feature set achieving the lowest relative RMSE in 36 of 60 stock-month cases, and with each stock alternating between two and three clusters while switching importance methods roughly every ten trading events.","pith_inferences":["If the per-stock feature spaces differ as reported, then any fixed feature set deployed by an HFT firm is leaving forecast accuracy on the table; the natural next step is to run the same dual-importance machinery on Level 2 or Level 3 order-book depth and watch whether the MDI/GD switch rate changes with market regimes.","The isotropic-cluster assumption is the part most worth attacking: replacing scalar widths with per-cluster covariance matrices (or swapping k-means for a Gaussian mixture) is a direct test the authors did not run, and it would tell whether the reported gains come from the clustering or from the RBF fit.","Because the method switches between two and three clusters and swaps MDI/GD roughly every ten events, the switch itself could be read as a cheap online signal of microstructure regime change, a use the paper does not explore."],"forward_implications":["Each stock requires a different input feature space, so a single fixed feature set cannot be optimal across stocks.","The number of RBF neurons is no longer a user choice: silhouette quality determines it online for every sliding window.","Because the method alternates between MDI and GD roughly every ten events, feature relevance in the limit order book is itself a high-frequency quantity.","The protocol's online design, with overlapping 100-event blocks and cumulative training folds, yields tick-by-tick forecasts without retraining on the full history."],"supporting_citations":[{"why":"Supplies the mean-decrease-impurity definition that forms one arm of the dual feature-importance mechanism.","marker":"[1]"},{"why":"Supplies the gradient-descent conversion into a feature-importance vector, the benchmark arm that competes with MDI.","marker":"[34]"},{"why":"Supplies the correlation-distance transformation and the silhouette-quality ratio used to choose the number of clusters.","marker":"[35]"},{"why":"Shows the k-means-with-RBFNN combination the paper extends by automating cluster count and feature weighting.","marker":"[31]"},{"why":"Establishes RBFNN forecasting of limit-order-book mid-prices, the task this protocol automates.","marker":"[33]"}],"fun_headline_variants":["Autonomous feature selection and clustering forecast HFT mid-prices","Self-tuning RBFNN replaces manual topology search for stock forecasting","Dual MDI-GD importance and silhouette k-means automate RBF stock prediction","No grid search: auto feature clustering and ranking for HFT forecasts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline stands on the premise that each cluster of order-book features has the same spread in every direction, because the RBF neurons use only one scalar width per cluster.","fun_headline_variants_meta":{"raw":{"variants":["Autonomous feature selection and clustering forecast HFT mid-prices","Self-tuning RBFNN replaces manual topology search for stock forecasting","Dual MDI-GD importance and silhouette k-means automate RBF stock prediction","No grid search: auto feature clustering and ranking for HFT forecasts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000355,"raw_usage":{"total_tokens":1921,"prompt_tokens":929,"completion_tokens":992,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":915}},"tokens_in":545,"tokens_out":992,"duration_ms":9208,"temperature":1.0,"reasoning_tokens":915,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:05:21.681286+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the same 20 stocks, compute the ratio of the largest to the smallest eigenvalue of the within-cluster covariance matrix for the features that get clustered; if many clusters show large eigenvalue ratios, the isotropy assumption is violated. Then rerun the pipeline with per-cluster covariance matrices replacing the scalar $\\sigma_j$ (or with Gaussian mixture clustering) and check whether test RMSE changes. If anisotropic clustering does not alter the errors, the isotropy assumption is not load-bearing; if it improves them significantly, the paper's assumption masks the true mechanism.","supporting_citations":[{"cited_title":"Random Forest","cited_arxiv_id":null,"evidence_quote":"Supplies the mean-decrease-impurity definition that forms one arm of the dual feature-importance mechanism."},{"cited_title":"Detection of false investment strategies using unsupervised learning methods","cited_arxiv_id":null,"evidence_quote":"Supplies the correlation-distance transformation and the silhouette-quality ratio used to choose the number of clusters."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows the k-means-with-RBFNN combination the paper extends by automating cluster count and feature weighting."},{"cited_title":"Machine learning for forecasting mid-price movements using limit order book data","cited_arxiv_id":null,"evidence_quote":"Establishes RBFNN forecasting of limit-order-book mid-prices, the task this protocol automates."}],"review_version":1}