{"id":"c70fcc04-b5e8-4725-81d4-509e2501fdbd","arxiv_id":"2509.24144","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"An LSTM-GAT model with news sentiment, trained end-to-end to maximize the Sharpe ratio, beat equal-weight and CAPM-MVO benchmarks on a nine-stock US portfolio from early 2024 to mid 2025.","lead":"The paper builds a stock-trading model that combines three signals at once: price history, relationships among stocks, and financial news sentiment, and it lets the model choose portfolio weights directly instead of first forecasting returns. On nine large US stocks tested from early 2024 to mid 2025, the model reported higher returns and better risk-adjusted performance than simple equal-weight and CAPM-based portfolios.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Universe selection may have used full-sample correlation data, undermining the out-of-sample claim; the paper never discloses when or on what data the 50-to-9 stock selection was made.","rationale":"The reader's verdict is CONDITIONAL, and I agree with that overall disposition. However, I elevate the universe-selection concern above the frictionless-execution concern. Frictionless execution is a real limitation, but the authors explicitly frame their results as 'upper-bound estimates' (Section 3.3), so the central claim is implicitly conditional on zero costs; the reader's own summary acknowledges this. The universe-selection issue, by contrast, threatens the very meaning of 'out-of-sample.' If the nine stocks were chosen using full-sample data, the test period is not a clean holdout: the assets themselves were selected to be diversifying over the very window in which performance is reported. This is not a matter of cost magnitude; it is a potential invalidation of the experimental design. The paper provides no timeline for the selection step, so the concern cannot be dismissed without additional information. A concrete, low-cost test is to ask for the selection date and data window, or to re-run selection on pre-2024 data and observe whether the reported margin persists. The reader identified this as a secondary premise; my analysis suggests it should be the primary check. The verdict remains CONDITIONAL because the issue is plausibly fixable via disclosure or a re-run, and the paper's other strengths (honest limitations, clear architecture, internal consistency) do not warrant rejection or unverdictability on this basis alone.","tokens_in":11499,"tokens_out":4920,"duration_ms":38628,"concrete_test":"Ask the authors to disclose (or, if code/data are released, inspect) the exact procedure and date of the 50-to-9 universe selection. In particular, state the start and end dates of the data used to compute 'low pairwise return correlations.' If that window includes any date on or after the test-period start (roughly January 2024), rerun the entire pipeline with the universe selected using only data through December 2023. Then compare the resulting v3 annualized return and Sharpe ratio against Table 2. If the margin over equal-weight and CAPM-MVO shrinks or reverses, the reported headline is partly an artifact of selection bias. If the selection was already restricted to pre-2024 data, the test simply documents that the out-of-sample claim is clean.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 states that nine S&P 500 stocks were chosen from a pool of 50 based on (i) low pairwise return correlations, (ii) sector diversity, and (iii) news coverage. The paper does not disclose the time window or date on which this selection was performed. If the low-correlation criterion was evaluated over the full 2021–2025 sample, the selected universe is precisely the set of stocks that were mutually diversifying during the test window (early 2024–mid 2025). This is a form of lookahead/data snooping: the 'out-of-sample' test is then conducted on a universe selected with information from the test period, which can mechanically inflate the Sharpe ratio and cumulative return of any dynamic allocation strategy, including the proposed LSTM-GAT. The authors' 'upper-bound estimates' caveat (Section 3.3) does not address this: the issue is not friction but the validity of the test set itself. This is the most load-bearing concern because, if true, the central claim of superior out-of-sample risk-adjusted performance is not established regardless of transaction costs. The reader mentioned this only as a secondary, unstated premise; I argue it is the primary threat to the paper's conclusion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an end-to-end deep-learning pipeline for daily portfolio allocation that combines LSTM temporal encoding, a Graph Attention Network over stock relationships, and news-sentiment features, with portfolio weights trained directly by maximizing the Sharpe ratio. Five model variants (v1–v5) are evaluated on a fixed nine-stock U.S. universe over an out-of-sample period spanning early 2024 to mid-2025, including the April 2025 tariff shock, and compared against equal-weight and CAPM-MVO benchmarks. The authors report that all LSTM-GAT variants beat both benchmarks on cumulative/annualized return and Sharpe ratio, with the sentiment-augmented v3 model achieving the best results (31.23% annualized return and 1.15 Sharpe ratio vs. 20.85%/0.83 for equal-weight and 18.58%/0.84 for CAPM-MVO). The paper frames the results as a proof of concept and repeatedly acknowledges the zero-transaction-cost, small-universe, and API-sentiment limitations.","tokens_in":11663,"tokens_out":3848,"duration_ms":32706,"significance":"If the results hold, the paper's contribution is a coherent architecture that unifies temporal modeling, relational graph structure, and sentiment in a single differentiable portfolio objective, avoiding the error compounding of two-step prediction-then-optimization pipelines. The inclusion of the April 2025 stress period is a strength, and the use of external benchmarks and a held-out test window is methodologically sound in spirit. The reported tables are internally consistent, and the authors are candid about several limiting assumptions. However, the evidence supporting the central claim is currently fragile: the universe-selection procedure is not time-stamped and may leak test-period information; the evaluation rests on one random seed and one test window; and the zero-cost daily-rebalancing assumption is acknowledged but never stress-tested. These issues are fixable but require additional experiments.","major_comments":[{"comment":"The stock selection procedure is not time-stamped. The nine stocks are chosen from an initial pool of 50 based on 'low pairwise return correlations,' sector diversity, and news coverage, but the paper never states the time window or data period used to compute those correlations. If the correlation criterion was evaluated over the full 2021–2025 sample, then the test window (early 2024–mid 2025) has been used to construct the universe, which is a lookahead bias that can mechanically inflate the out-of-sample performance of any allocation strategy. This is not covered by the 'upper-bound' caveat in Section 3.3, which concerns frictions rather than test-set validity. Please disclose when and on what data the selection was made, and ideally re-run the selection using only pre-test information or show robustness to alternative universe-construction windows.","section":"Section 3.1"},{"comment":"All reported results come from a single random seed (42) and a single out-of-sample window, with no error bars, confidence intervals, or significance tests. The performance gaps in Table 2—for example, the Sharpe ratio difference between Model v1 (0.91) and Model v2 (0.95), or between Model v4 (1.06) and Model v3 (1.15)—could easily be within the noise of one initialization and one market episode. The paper's central claim that all LSTM-GAT variants outperform the benchmarks needs statistical support: report results over multiple seeds (e.g., 5–10) and, if feasible, multiple test windows or a bootstrap of the test-period return series.","section":"Section 3.5 / Table 2"},{"comment":"The evaluation assumes zero transaction costs, zero slippage, and daily rebalancing, which the authors themselves describe as producing upper-bound estimates. This assumption is load-bearing because the reported advantage over the equal-weight benchmark is about 10 percentage points of annualized return, and daily rebalancing across nine stocks would incur nontrivial costs in practice. Please add a transaction-cost sensitivity analysis: for example, show how the cumulative-return and Sharpe-ratio gaps change under round-trip costs of 5, 10, and 25 basis points, and report the break-even cost at which the best LSTM-GAT variant no longer beats the benchmarks. Without this, the practical claim of 'smarter portfolio decisions' is not established.","section":"Section 3.3 / Section 4.1"},{"comment":"The Sharpe-ratio loss in Eq. (1) involves a covariance matrix Σ, but the paper never specifies how Σ is computed during training and evaluation. Is it the empirical covariance of the batch's realized returns? A rolling window estimate? A shrinkaged estimator? Is it recomputed at test time? This is essential for reproducibility and for understanding whether the end-to-end gradient flow is well-defined. Please provide the exact definition and estimation procedure for Σ.","section":"Section 2.4 / Section 3.5"},{"comment":"The comparison across Model v1–v5 is presented as an ablation of features, graph type, and dimensionality reduction, but each variant is independently tuned with Optuna over 50 trials. Consequently, differences between variants (e.g., v3 vs. v4 for static vs. dynamic graphs, or v4 vs. v5 for PCA) are confounded with differences in hyperparameters: Table 8 shows, for instance, that v3 uses LSTM hidden size 32 and 2 layers, while v4 uses 80/1 and v5 uses 32/1. Attributing performance differences to sentiment or dynamic graphs is only valid if the architectural component is varied while holding the optimization procedure and hyperparameter search fixed. Please either run controlled ablations with a shared hyperparameter budget or soften the causal language in Section 4.1.","section":"Section 3.4 / Table 8"}],"minor_comments":[{"comment":"The text in Section 3.4 says v3 adds only 'sentiment variance and weighted sentiment,' but Table 1 shows that v3 also adds 'News Count' and 'Average Sentiment.' Please reconcile the description with the table.","section":"Section 3.1 / Table 1"},{"comment":"The hyperparameter 'gat alpha' for Model v5 is listed as 0.35, but the Optuna search space in Table 7 specifies a range of 0.05–0.30. This inconsistency should be corrected.","section":"Table 8"},{"comment":"The note says 'Values highlighted in blue represent the best performance,' but the manuscript as rendered has no color. Please indicate the best values by boldface or another unambiguous marker.","section":"Table 2"},{"comment":"The phrase 'Building the future set required collecting both price and news data' appears to be a typo; presumably 'feature set' was intended.","section":"Section 3.2"},{"comment":"The limitations section acknowledges the fixed nine-stock universe, zero transaction costs, static GICS sectors, and opaque third-party sentiment APIs. These are appropriate, but the paper does not mention the lack of code/data availability; adding a reproducibility statement would strengthen the contribution.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a well-structured proof of concept, but the central claim of out-of-sample superiority currently rests on an undisclosed universe-selection window, one seed, one test period, and zero-cost trading. None of these issues are fatal in themselves—they are all addressable with additional experiments and disclosures. However, the universe-selection question in particular should be treated as a correctness risk: if the low-correlation filter was run on the full sample, the 'out-of-sample' test is compromised. I would encourage the editor to request the authors to clarify the selection timeline and, if needed, re-run the selection on training data only before considering publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on arXiv:2509.24144. The integration is new: an end-to-end LSTM-GAT that outputs portfolio weights directly, trained on the Sharpe objective, with news sentiment features. The authors test it on nine large-cap US stocks over a window that includes the April 2025 tariff shock, and the best variant (v3) beats equal-weight and CAPM-MVO on return and Sharpe. The methodology is described well enough that the work is understandable even without code. The paper also does several things right: the limitation section is candid, calling the results upper-bound and noting the sparse graph, opaque sentiment API, static sectors, and heuristic threshold. The ablation design is sensible—v1 price-only, v2 adds features, v3 adds sentiment, v4 dynamic graph, v5 PCA—and the finding that sentiment helps and dynamic graphs reduce volatility is coherent.\n\nThe soft spots are real and, in one case, more serious than the reader's note suggested. The biggest issue is the universe selection. Section 3.1 says nine stocks were chosen from a pool of fifty for low pairwise correlations and sector diversity, but it never says when or on what data the selection was made. If the correlations were computed over the full 2021–2025 sample, the test window is effectively in-sample for the universe, which can inflate the Sharpe of any dynamic strategy. That is lookahead, not a friction issue, and the authors' upper-bound caveat does not address it. The second issue is the uneven benchmark comparison: the model rebalances daily with zero costs while the CAPM-MVO rebalances every 21 days. Realistic frictions could erase much of the reported advantage. The third is statistical: the evaluation rests on a single seed and a single test window with no error bars or significance testing.\n\nNone of these is individually fatal, and the central architecture story holds up. But together they mean the headline outperformance is not established beyond this specific backtest. The fixes are straightforward: disclose the universe-selection timing, add multi-seed or bootstrap intervals, include a cost scenario, and consider a non-deep baseline using the same features. I'd want those before citing the results, but the paper deserves a serious referee—it is clearly written, the integration is novel, and the flaws are addressable. I'd bring it to a reading group to discuss the design and the lookahead trap.","headline":"A genuinely new end-to-end LSTM-GAT sentiment portfolio model, clearly written, but the empirical headline is not yet credible because of an undisclosed universe-selection step and a frictionless daily-rebalance benchmark; still worth refereeing.","tokens_in":12317,"tokens_out":2714,"would_cite":false,"duration_ms":21197,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91G10","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"An end-to-end LSTM-GAT network that directly optimizes the Sharpe ratio, using price, graph, and news-sentiment features, beats equal-weight and CAPM-MVO benchmarks out-of-sample.","keywords":["Portfolio optimization","Asset allocation","Graph neural networks","Graph attention networks","Long short-term memory","Financial news sentiment","Sharpe ratio","End-to-end learning"],"falsifier":"Re-run the same models on the same nine stocks with a per-trade transaction cost of 10 basis points plus slippage and compare net-of-cost Sharpe ratios to the equal-weight and CAPM-MVO baselines; if the LSTM-GAT variants no longer dominate, the paper's central claim is empirically falsified. A second decisive check is to repeat the experiment on a larger, independently selected universe with the stock selection made before the test window begins.","tokens_in":11217,"feed_emoji":"📈","tokens_out":7895,"duration_ms":49366,"temperature":0.7,"pith_summary":"This paper claims that an end-to-end deep network can replace the traditional forecast-then-optimize pipeline in portfolio management. The network combines an LSTM for each stock's price and sentiment history, a graph attention layer that lets stocks share information through correlations and sector ties, and a final layer that outputs daily portfolio weights; it is trained directly to maximize the Sharpe ratio. On a nine-stock universe that includes the April 2025 tariff shock in the out-of-sample period, all five model variants beat equal-weight and CAPM-based mean-variance benchmarks, with the sentiment-augmented version reaching a 31.23% annualized return and 1.15 Sharpe ratio. The authors read this as evidence that integrating temporal, relational, and sentiment signals in a single objective yields more robust risk-adjusted returns than two-stage methods. They caution that results are upper-bound estimates because trading is assumed frictionless.","feed_headline":"Deep-learning portfolio beats classic benchmarks by 68%","feed_subtitle":"Daily weights from price, graph, and sentiment beat equal-weight and CAPM-MVO, stress-tested through 2025 tariffs.","key_machinery":"The load-bearing object is the end-to-end weight generator: a shared LSTM encodes 30-day feature windows for each stock, a Graph Attention Network refines those embeddings using a graph whose edges encode sector membership and correlations of returns and sentiment (static or updated every five trading days), and a tanh linear layer produces raw weights that are normalized to sum to one. The entire system is trained with Adam to minimize the negative Sharpe ratio of the resulting portfolio, so the model directly optimizes the investor's risk-adjusted objective rather than a proxy forecast. The graph construction and the Sharpe loss are what distinguish this from a simple price predictor.","core_discovery":"The central claim is that an LSTM-GAT network trained end-to-end to maximize the Sharpe ratio—rather than first predicting returns and then optimizing—can learn daily portfolio weights that outperform equal-weight and CAPM-MVO benchmarks. The best configuration, which adds sentiment features to price and volume inputs, achieves a 31.23% annualized return and a 1.15 Sharpe ratio out-of-sample versus 20.85%/0.83 and 18.58%/0.84 for the baselines. The authors attribute the gains to three design choices: direct Sharpe optimization avoids compounding forecast errors; the GAT captures evolving cross-asset dependencies; and news sentiment carries information not fully reflected in prices. They also","pith_inferences":["A decisive next test is to add realistic trading costs (e.g., 5–10 basis points per trade) and slippage; if the sentiment variant's Sharpe ratio falls below the benchmarks, the practical claim collapses even if the frictionless claim stands.","The nine stocks were chosen from 50 S&P 500 names using low-correlation and news-coverage filters; since the paper does not state whether this selection was made before or after seeing the test window, a replication with a pre-registered universe is needed to rule out selection bias.","The third-party sentiment scores are black-box; replacing them with an openly documented sentiment model (e.g., a fine-tuned financial transformer) would identify how much of the reported gain depends on the specific sentiment source.","Because the graph has only nine nodes, the GAT's relational learning is severely constrained; extending to a larger universe is the natural stress test for whether relational information, rather than the Sharpe objective alone, drives the gains."],"forward_implications":["If the central claim is correct, daily allocations can be produced by a single network trained on portfolio-level objectives, bypassing the unstable two-step forecast-then-optimize loop.","Adding financial-news sentiment to price-based features improved the annualized return and Sharpe ratio by nearly 50% and 39% over the equal-weight benchmark in the tested universe, suggesting news tone carries information not already in prices.","Switching from a static to a weekly-updated correlation graph reduced volatility and drawdown at a small cost in raw return, indicating that adaptive relational structure helps in changing markets.","PCA-based feature compression produced the smallest maximum drawdown during the April 2025 tariff shock, hinting that noise reduction improves stress resilience.","Because the backtest assumes zero transaction costs and daily rebalancing, the reported outperformance is an upper bound; under realistic frictions the gap over baselines would narrow."],"fun_headline_variants":["Sharpe-optimized deep learning beats equal-weight and CAPM","LSTM-GAT with news sentiment lifts portfolio returns","End-to-end model learns weights, not returns, outperforms","Deep learning plus sentiment tops classic portfolio models","AI daily stock weights from price, graph, and news"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the backtest reflects frictionless trading at official prices with no costs, slippage, or market impact, and that the nine-stock universe was selected without peeking at test-window data; if either fails, the reported edge could vanish.","fun_headline_variants_meta":{"raw":{"variants":["Sharpe-optimized deep learning beats equal-weight and CAPM","LSTM-GAT with news sentiment lifts portfolio returns","End-to-end model learns weights, not returns, outperforms","Deep learning plus sentiment tops classic portfolio models","AI daily stock weights from price, graph, and news"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000243,"raw_usage":{"total_tokens":1344,"prompt_tokens":704,"completion_tokens":640,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":448,"completion_tokens_details":{"reasoning_tokens":560}},"tokens_in":448,"tokens_out":640,"duration_ms":8627,"temperature":1.0,"reasoning_tokens":560,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T14:39:03.995929+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same models on the same nine stocks with a per-trade transaction cost of 10 basis points plus slippage and compare net-of-cost Sharpe ratios to the equal-weight and CAPM-MVO baselines; if the LSTM-GAT variants no longer dominate, the paper's central claim is empirically falsified. A second decisive check is to repeat the experiment on a larger, independently selected universe with the stock selection made before the test window begins.","supporting_citations":[],"review_version":1}