{"id":"8637aba0-efde-4764-b2ee-3c88815552b7","arxiv_id":"2411.18997","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GRU-PFG, a factor-only model, reports IC 0.134 on CSI300, surpassing the multi-source HIST model's 0.131.","lead":"This paper proposes GRU-PFG, a stock prediction model that uses only Alpha360 factor data and builds inter-stock correlations through softmax and Pearson correlation matrices over GRU hidden states. On CSI300 data it reports an information coefficient of 0.134, slightly above the multi-source HIST model at 0.131, suggesting factor-only models can compete with models that use extra data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 0.134-vs-0.131 IC comparison is not reproducible from the text: Eq. 8 defines MSE as the training loss but §5.1 says IC is the training loss, leaving two incompatible implementations and making the claimed edge over HIST unverifiable.","rationale":"The Reader's verdict is already CONDITIONAL, and I do not move it; this pass supplies a sharper basis for that condition. The strongest claim is the empirical superiority of GRU-PFG over HIST. That claim requires a well-defined, fair experiment. The paper provides two incompatible training objectives (MSE in Eq. 8, IC in §5.1), an unusual return definition (Eq. 2), and no code. These are internal inconsistencies, not mere deviations from consensus, so they are directly load-bearing. The Reader's chosen weakest assumption -- that the Pearson correlation graph learned from 2007-2014 generalizes to 2017-2020 -- is a plausible modeling risk, but it is secondary: if the loss-function or label ambiguity changes the experimental outcome, the correlation graph's stability is moot. The concrete test (clean reproduction under both loss protocols) would settle whether the reported 0.134 is real or an artifact of a mismatched training/evaluation protocol. I therefore keep the conditional verdict while pointing the condition at the protocol, not at the correlation graph.","tokens_in":11049,"tokens_out":5732,"duration_ms":91192,"concrete_test":"Ask the authors for the official code and run a clean reproduction on the qlib CSI300 benchmark. Train GRU-PFG under both protocols: (a) MSE loss exactly as in Eq. 8 and (b) IC as training loss as stated in §5.1, and retrain HIST, GRU, and ALSTM under the same protocol and the same return-label definition. If the reported IC=0.134 is obtained only in protocol (b) while baselines were trained under protocol (a), or if the Eq. 2 open-to-close label is not applied uniformly, Table 3 does not support the central claim. Report the daily IC series and test the 0.134-vs-0.131 difference for statistical significance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: GRU-PFG, using only Alpha360 factors, reaches IC 0.134 on CSI300 and beats HIST's 0.131. For this claim to hold, the experimental protocol must be unambiguous and fair. The paper undermines this condition internally. In Section 4, Eq. 8 defines the loss as MSE, while Section 5.1 states 'During training, IC is used as the loss function.' These define different training objectives; a model trained to maximize IC will typically report higher IC than one trained with MSE, so the comparison with HIST and the other baselines changes meaning depending on which sentence is correct. A second ambiguity is the target definition: Eq. 2 defines the daily return as (P_close - P_open)/P_open, an intraday open-to-close return, whereas standard qlib benchmarks label daily returns close-to-close; if labels differ across models, Table 3 is not a like-for-like comparison. No code, hyperparameters, or seeds are provided, so a reader cannot resolve either ambiguity. The 0.003 IC difference is also within the reported standard deviations (3.2e-3 and 2.2e-3), so even with a consistent protocol the 'outperforming' statement may be noise. The correlation-stability assumption identified by the Reader is real but secondary; if the comparison itself is invalid, no stability assumption can rescue the headline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GRU-PFG, a stock prediction model that uses only Alpha360 factors as input. The model first applies a GRU to obtain per-stock feature representations, then constructs two levels of inter-stock relationship graphs by applying softmax along rows and columns and computing Pearson correlations, and finally combines the resulting features through learned weights. The model is evaluated on CSI100 and CSI300 over a 2007-2020 split and compared with factor-only baselines (MLP, LSTM, GRU, SFM, GATs, ALSTM, Transformer) and multi-source baselines (HIST, ALSTM+TRA). The headline claim is that on CSI300, GRU-PFG reaches an IC of 0.134, exceeding HIST's 0.131 despite using fewer input sources. The paper also includes an ablation study and a comparison between Pearson correlation and cosine similarity for building the graph.","tokens_in":11501,"tokens_out":3720,"duration_ms":35465,"significance":"If the empirical claims are reproducible, the result would be interesting: it would show that a factor-only model can extract inter-stock correlations from Alpha360 data and match or exceed a multi-source model like HIST on CSI300. The paper reports standard deviations, uses standard qlib baselines, and includes an ablation study, which are positive features. However, the central empirical claim is currently unverifiable because of an unresolved contradiction in the training objective, an ambiguous label definition, and the absence of code, hyperparameters, and seeds. The reported 0.003 IC advantage over HIST is also within the reported standard deviations, so the headline claim needs a statistical test rather than a point comparison. The contribution is incremental but potentially useful if the protocol is clarified and the code is released.","major_comments":[{"comment":"The training objective is specified inconsistently: Eq. (8) defines the loss as MSE, while §5.1 states \"During training, IC is used as the loss function.\" These two objectives are not equivalent, and the reported IC values depend on which one was actually optimized. Please state the true objective, provide the training loss formula actually used, and release code or at least all hyperparameters and seeds so the 0.134 result can be reproduced.","section":"§4, Eq. (8) and §5.1"},{"comment":"The label definition is ambiguous relative to the qlib baselines. Eq. (2) defines the daily return as (P_close - P_open)/P_open, an open-to-close intraday return, whereas the standard qlib daily label used by the cited baselines is a close-to-close return. If the labels fed to GRU-PFG differ from those used for HIST and the other baselines, Table 3 is not a like-for-like comparison. Specify the exact label used for all models and confirm that it matches the qlib convention.","section":"§3.1, Eq. (2)"},{"comment":"The headline CSI300 IC advantage of GRU-PFG (0.134) over HIST (0.131) is 0.003, which is smaller than the reported standard deviations (3.2e-3 for GRU-PFG and 2.2e-3 for HIST). Please report the number of runs, the seeds, and a paired significance test (e.g., a paired t-test over test dates or a Diebold-Mariano test) before claiming outperformance.","section":"Table 3"},{"comment":"The correlation matrix construction is not fully specified: Eq. (4) refers to Fx and Fy without indicating whether the Pearson correlation is computed across the 64 hidden dimensions of each stock or across stocks, and the dimensions of R1 and R2 are not stated even though they multiply X and Xhid in Eq. (6). Please define all dimensions and the exact normalization steps so the graph aggregation is reproducible.","section":"§4, Eq. (4) and Eq. (6)"}],"minor_comments":[{"comment":"The abstract's statement that GRU-PFG achieves \"results better than the second category models\" is contradicted by Table 3 on CSI100, where HIST's IC is 0.120 versus GRU-PFG's 0.116; please qualify the claim to the CSI300 dataset.","section":"Abstract and Table 3"},{"comment":"Please correct typographical errors: \"T able 1/2/3\", \"Fundmentals\", \"Stock F eature\", \"F last\" in Eq. (7), and \"the others models relies\" in the abstract.","section":"Throughout"},{"comment":"The comparison between Pearson correlation and cosine similarity is described only qualitatively; please add the corresponding IC, Rank IC, and Precision@N values so the claimed superiority of Pearson correlation can be assessed.","section":"§5.3"},{"comment":"Reference [23] is incomplete (missing authors and title), and several references lack page numbers or DOIs; please complete the bibliography.","section":"References"},{"comment":"Figures 3 and 4 are difficult to read at the current resolution; please provide larger panels, clearer legends, and axis labels.","section":"Figures 3 and 4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nGRU-PFG is a factor-only stock predictor that builds a graph by taking GRU hidden states, softmaxing them twice, and using Pearson correlation as edge weights. That specific combination is new, and the paper does a few things right: the architecture is described step by step, the ablation at least compares Pearson against cosine similarity, and the empirical section reports standard deviations for every metric, which is more than many q-fin papers do.\n\nThe problem is the headline. The claim that GRU-PFG beats HIST on CSI300 (IC 0.134 vs 0.131) is not verifiable from the text. Section 4, Eq. 8 defines the loss as MSE; Section 5.1 says IC is used as the training loss. Those are different objectives, and the gap between the two models is 0.003, smaller than both reported standard deviations (3.2e-3 for GRU-PFG, 2.2e-3 for HIST). So even if the loss ambiguity were a typo, the 'outperforming' statement is likely noise. The return definition is also non-standard: Eq. 2 labels open-to-close returns, whereas qlib baselines typically predict close-to-close; if that's what the model was trained on, Table 3 isn't a like-for-like comparison.\n\nThe reader's concern about correlation stability is real but secondary. If the comparison itself is invalid, no stability assumption can save the headline. The lack of code or hyperparameter details means a reader can't resolve any of this.\n\nOn the positive side, the idea that inter-stock structure can be extracted endogenously from factors, without external industry or sentiment data, is worth testing. The paper doesn't prove it, but it's a legitimate research direction. The writing is rough but readable.\n\nI'd send this to a referee only if the editor is willing to require the authors to clarify the loss function, the return definition, and provide code or at least exact experimental settings. Without that, the empirical claim is uninterpretable. As it stands, I wouldn't cite it, and I wouldn't build on it. For a reading group it might be a good case study in why standard deviations and consistent protocols matter.\n\nRecommendation: conditionally engage, but only with a request for code and clarification before any further consideration.","headline":"Plausible architecture, unreproducible headline: the 0.134-vs-0.131 IC edge over HIST is inside the noise and the paper contradicts itself on the training loss.","tokens_in":11884,"tokens_out":3499,"would_cite":false,"duration_ms":33835,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A factor-only stock predictor extracts inter-stock correlations and beats a multi-source model on CSI300.","keywords":["stock trend prediction","inter-stock correlation","Alpha360 factors","graph neural network","GRU","CSI300","information coefficient","factor-only models"],"falsifier":"Retrain GRU-PFG and HIST on training data that stops in 2014 and test separately on market regimes with known breaks such as the 2015 crash, the 2017-2018 trade tensions, and the 2020 COVID sell-off; if GRU-PFG's IC advantage over HIST is confined to one regime and reverses in another, the claim of stable learned inter-stock correlations is not supported. A second check is to shuffle stock identities in the test set: if the graph component still yields high IC when the correspondence between features and stock labels is broken, the correlation module is not extracting genuine inter-stock information.","tokens_in":10860,"feed_emoji":"📈","tokens_out":5803,"duration_ms":51311,"temperature":0.7,"pith_summary":"The paper claims that a stock prediction model fed only the standard Alpha360 factor set can learn inter-stock relationships directly from those factors, without any extra data such as industry labels, financial reports, or sentiment. The proposed architecture, GRU-PFG, first processes each stock's factor sequence with a GRU, then builds explicit stock-to-stock correlation matrices from the learned features and uses them as graph-style weights. On the CSI300 dataset it reports an information coefficient of 0.134, above the 0.131 of the multi-source HIST model, while using fewer inputs. If the result holds, it would mean that inter-stock correlation information is already latent in price-and-volume-derived factors, and that standardization-friendly factor-only models need not sacrifice accuracy to multi-source models.","feed_headline":"With only stock factors, GRU-PFG beats HIST on CSI300","feed_subtitle":"A model using Alpha360 factors alone learns cross-stock links, matching models fed extra data.","key_machinery":"The load-bearing mechanism is a two-stage correlation construction on GRU feature outputs. The GRU treats the 360-dimensional Alpha360 vector as 60 time steps of 6 factors and compresses it to a 64-dimensional representation; softmax over rows and columns then produces two views, one capturing relationships within a stock's own features and one capturing relationships across stocks. Pearson correlation over these views yields an aggregation matrix R1, and a residual matrix computed by subtracting weighted row and column views gives a second aggregation matrix R2; the final stock feature is a learnable weighted sum of original, primary, and secondary graph-weighted features. The graph operation is what lets factor-only inputs carry inter-stock information.","core_discovery":"The central discovery the paper argues for is that inter-stock correlations needed for prediction can be extracted from Alpha360 factors alone, rather than supplied from external sources. GRU-PFG projects the factor-derived GRU hidden states into a graph by computing row-wise and column-wise softmax normalizations and then applying Pearson correlation to form aggregation matrices; these matrices weight the features of every stock by their relation to every other stock. Combining the primary and a secondary residual relationship extraction, the model reaches IC 0.134 and rank IC 0.128 on CSI300, outperforming HIST (0.131 and 0.126) and all factor-only baselines. The authors read this as evidence that the factor set contains shared inter-stock structure, and that a graph-based projection can recover it.","pith_inferences":["Editorial inference beyond the paper: If inter-stock correlation is recoverable from Alpha360 alone, then a similar graph-projection layer could be attached to other sequence encoders such as LSTM or Transformer, and it may yield comparable improvement without their external data.","Editorial inference beyond the paper: The result suggests that part of the advantage of multi-source models may be redundant with information already encoded in cross-sectional factor patterns; a direct test would be to feed HIST's extra inputs to GRU-PFG and see whether the gain over factor-only inputs is smaller than the gain HIST obtains.","Editorial inference beyond the paper: The correlation matrices R1 and R2, being learned from training-period data, could be inspected over time; if their entries track known industry or sector transitions, the model would provide an interpretable, data-driven alternative to fixed industry classifications.","Editorial inference beyond the paper: Since GRU-PFG benefits from more stocks (CSI300 over CSI100), its performance might improve further on larger universes such as CSI500 or the full A-share market, provided the factor computation remains standard."],"forward_implications":["On CSI300, GRU-PFG reports IC 0.134, beating multi-source HIST's 0.131 while using only Alpha360 factors.","It outperforms all factor-only baselines tested, including GRU (0.113), ALSTM (0.115), and Transformer (0.106).","Its advantage over the factor-only baselines grows from CSI100 to CSI300, suggesting that the graph component benefits from a larger pool of stock relationships.","Ablation shows performance improves as the primary and secondary relationship extraction blocks are added, and Pearson correlation outperforms cosine similarity for this aggregation.","Because the model needs no industry labels, reports, or sentiment data, it is easier to apply to new stocks and new markets."],"supporting_citations":[{"why":"HIST, the multi-source model whose 0.131 IC on CSI300 is the main baseline GRU-PFG claims to beat.","marker":"[12]"},{"why":"Defines and implements the Alpha360 factor set and its computation, supplying the only input the model uses.","marker":"[26]"},{"why":"TRA, the other multi-source model in the comparison table, showing that a factor-only model can match or beat it.","marker":"[28]"},{"why":"The GRU cell used for preliminary extraction, producing hidden states whose correlation structure drives the graph.","marker":"[31]"},{"why":"The cosine-similarity alternative tested when designing the correlation aggregation; replacing Pearson with it lowers performance.","marker":"[42]"}],"fun_headline_variants":["GRU-PFG extracts stock links from factors alone, beats HIST","Factor-only graph model GRU-PFG tops HIST on CSI300","New GRU-PFG: inter-stock links from Alpha360 factors only","Graph from factors: GRU-PFG outperforms HIST on CSI300"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's edge depends on the assumption that the correlation matrix built from softmax-normalized GRU states during 2007-2014 training captures stock relationships that remain predictive on the 2017-2020 test period; if that correlation structure does not generalize across market regimes, the reported advantage over baselines would not persist.","fun_headline_variants_meta":{"raw":{"variants":["GRU-PFG extracts stock links from factors alone, beats HIST","Factor-only graph model GRU-PFG tops HIST on CSI300","New GRU-PFG: inter-stock links from Alpha360 factors only","Graph from factors: GRU-PFG outperforms HIST on CSI300"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000149,"raw_usage":{"total_tokens":1182,"prompt_tokens":924,"completion_tokens":258,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":180}},"tokens_in":540,"tokens_out":258,"duration_ms":3196,"temperature":1.0,"reasoning_tokens":180,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:38:12.825560+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain GRU-PFG and HIST on training data that stops in 2014 and test separately on market regimes with known breaks such as the 2015 crash, the 2017-2018 trade tensions, and the 2020 COVID sell-off; if GRU-PFG's IC advantage over HIST is confined to one regime and reverses in another, the claim of stable learned inter-stock correlations is not supported. A second check is to shuffle stock identities in the test set: if the graph component still yields high IC when the correspondence between features and stock labels is broken, the correlation module is not extracting genuine inter-stock information.","supporting_citations":[{"cited_title":"Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining (2021)","cited_arxiv_id":null,"evidence_quote":"TRA, the other multi-source model in the comparison table, showing that a factor-only model can match or beat it."},{"cited_title":"In: 2017 IEEE 60th International Midwest Symposium on Circuits and Systems (MWSCAS), pp","cited_arxiv_id":null,"evidence_quote":"The GRU cell used for preliminary extraction, producing hidden states whose correlation structure drives the graph."},{"cited_title":"In: 2016 4th International Conference on Cyber and IT Service Management, pp","cited_arxiv_id":null,"evidence_quote":"The cosine-similarity alternative tested when designing the correlation aggregation; replacing Pearson with it lowers performance."}],"review_version":1}