{"id":"4ef62ff3-6e33-47d8-a075-cc99b142d58c","arxiv_id":"2505.06864","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A neural stochastic discount factor that reads news text through a pretrained language model and macro data through an LSTM reports out-of-sample Sharpe 2.80 on US equities from 2000 to 2022.","lead":"This paper builds a stock pricing model that combines financial news text, macroeconomic data, and firm characteristics using deep learning. The authors report a Sharpe ratio of 2.80, far above standard factor models, but details of the news pipeline and code are missing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The out-of-sample Sharpe ratio hinges on an unspecified news pipeline; if PCA or temporal alignment uses test-period information, the 2.80 result is an artifact.","rationale":"The reader's weakest assumption correctly identifies temporal alignment as the load-bearing point. My read of Sections 3.1, 4.1, and Algorithm 1 adds a specific mechanism: even if article timestamps are correctly sorted, the PCA projection of news embeddings is a global linear transform whose loadings, if estimated on the full 1980-2022 corpus, encode future article statistics into past features. The paper does not state that this projection is train-only. This is exactly the kind of detail that can silently create look-ahead in text pipelines, and it is not covered by the chronological train/validation/test split of the downstream model. The inconsistency of Figure 5's 2000-2024 window and Figure 3's 1970 start makes the reported evaluation harder to trust. On the positive side, the architecture is coherent, the moment-condition objective is a reasonable SDF estimator, and a chronological split is attempted. The right verdict is unchanged: conditional acceptance pending a clear causal-data specification and, ideally, code/data release. I agree with the reader rather than escalating to rejection because the concern is about missing verification, not a demonstrated contradiction.","tokens_in":10818,"tokens_out":7914,"duration_ms":87044,"concrete_test":"Run a strict-causality replication: fit the news PCA and attention parameters only on 1970-1994 articles; for each test month t, admit only articles whose publication timestamps precede the first day of the return month; and encode firm-months with no news as the zero vector. Recompute the 2000-2022 Sharpe ratio and MSPE. If the Sharpe ratio drops materially from 2.80 (e.g., below 2.0), the headline gain is inflated by leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that NewsNet-SDF earns an out-of-sample Sharpe ratio of 2.80 (Section 4.4, Table 2). For that number to be meaningful, every news embedding N_{t,i} must use only information available before the return month begins. The paper does not establish this. Section 3.1 says only 'We then apply PCA to project this embedding to a lower-dimensional space' — it never states that the PCA loadings and the attention aggregation parameters are fit on the training period (1970-1994) alone. Section 4.1 promises 'careful temporal alignment' but gives no timestamp-to-month matching rule, no rule for splitting multi-firm articles, and no handling for firm-months without news. Because the ablation attributes 41% of the Sharpe ratio to news text, a leak in this pipeline would directly manufacture the headline result. The exhibits compound the problem: Figure 5 plots beta-sorted decile returns through 2024 even though the stated test window ends in 2022, and Figure 3 begins in 1970, which includes the training sample. Without a causal news pipeline or released code, the 471% improvement over CAPM cannot be distinguished from look-ahead.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NewsNet-SDF, a stochastic discount factor estimator that fuses pretrained language model news embeddings, LSTM-encoded macroeconomic state, and cross-sectionally ranked firm characteristics, and trains an SDF through an adversarial moment-condition objective. On U.S. equity data from 1970 to 2022, with approximately 2.5 million New York Times articles, the authors report an out-of-sample Sharpe ratio of 2.80 for the 2000-2022 period, claiming large improvements over CAPM, Fama-French factor models, and neural SDF baselines. The paper also presents ablations, feature-importance analysis, beta-sorted decile portfolio tests, and a COVID-period analysis. The central empirical contribution is the claim that news text, processed through GTE-multilingual embeddings and attention aggregation, is the dominant source of the model's performance.","tokens_in":11067,"tokens_out":5866,"duration_ms":62589,"significance":"If the reported out-of-sample performance is valid, the paper would be a significant contribution to the growing literature on machine learning and text-based asset pricing: it offers a concrete way to incorporate unstructured news text into a theoretically grounded SDF framework, and the 2.80 Sharpe ratio is far above typical factor-model benchmarks. The architecture is coherent and follows the moment-based SDF logic of Hansen and Singleton, and the inclusion of three data modalities with an adversarial training objective is a sensible design. However, the significance is conditional on empirical credibility. The paper does not release code or data, does not report uncertainty around the headline Sharpe ratio, and leaves the news-to-return temporal alignment underspecified. Because the ablation attributes 41% of the Sharpe ratio to news text, these omissions are load-bearing for the main claim.","major_comments":[{"comment":"The temporal alignment between news articles and monthly return periods is not specified, which is critical because the headline result, a Sharpe ratio of 2.80 in Table 2, depends on the news features N_{t,i} containing no information realized after the start of month t+1. The paper says \"careful temporal alignment\" in Section 4.1 but does not state: how NYT publication timestamps are converted to monthly features; whether articles published during month t are included in N_{t,i} before the return R_{t+1} is realized; how multi-firm articles are split; how firm-months without news are handled; and whether the PCA projection and the attention parameters W, b, v in Eqs. (4)-(6) are estimated on the training sample (1970-1994) only or on the full sample. If any test-period information enters the news embeddings, the 2.80 Sharpe ratio and the 41% ablation effect are artifacts. Please provide an exact pipeline description, including the cutoff rule, and a robustness check that drops the most recent week of news before each month-end.","section":"Section 3.1 and Section 4.1"},{"comment":"The headline performance metrics are reported as point estimates with no measures of uncertainty. A Sharpe ratio of 2.80 estimated over 276 monthly observations, from 2000 to 2022, needs a standard error or a confidence interval, for example from a block bootstrap or Newey-West correction. Without such intervals, the statement that NewsNet-SDF exceeds CAPM by 471% and FF5 by 460% is not statistically interpretable, since the difference could be within sampling noise, especially under time-series dependence and estimation error in the SDF weights. Please report confidence intervals for SR, EV, XS-R2, and MSPE, and for the differences relative to the baselines.","section":"Section 4.4, Table 2"},{"comment":"The baseline implementations are not documented. Table 1 lists model names and one-line descriptions, but the paper does not state how each baseline is estimated on the same sample, what hyperparameters are used, whether the same 56 firm characteristics are used for all models, or how IPCA factors, GAN-SDF architectures, and TF-IDF/BERT text features are constructed. Without this information, the relative improvements in Table 2 are not reproducible, and the 471% improvement over CAPM is not a well-defined empirical statement unless the CAPM implementation is described in detail. Please provide pseudo-code, hyperparameter settings, or a released code repository.","section":"Section 4.2, Table 1"},{"comment":"There are inconsistencies between the stated test window and the figures. Table 2 and Section 4.1 define the test period as 2000-2022, but Figure 3 plots cumulative returns from 1970-2022, which includes the training and validation samples, making it impossible to visually isolate out-of-sample performance. Figure 5 is captioned \"Cumulative Returns of Beta-Sorted Decile Portfolios (2000-2024)\", with text in Section 4.6 also referring to 2000-2024, even though the data section states the sample ends in 2022. These discrepancies need to be resolved, and the figures should either be restricted to the out-of-sample window or clearly marked with training, validation, and test regions.","section":"Section 4.4, Figures 3 and 5"},{"comment":"The beta-sorted decile monotonicity is an internal consistency check, not an independent validation of the model's risk-return predictions. Since M_{t+1} = 1 - sum_i w_{t,i} R^e_{t+1,i} in Eq. (14), the model-implied beta in Eq. (25) is a deterministic transformation of the same weights w_{t,i} that produce the model's predicted expected returns; sorting on predicted beta is therefore essentially sorting on the model's own forecast. The perfect monotonicity and R^2>0.95 in Section 4.6 are consequence of this construction and do not provide independent evidence that high-risk assets earn higher returns. Please validate the model on external test assets, or on characteristic-sorted portfolios formed independently of the model, and report whether the monotonicity survives controlling for standard factor exposures.","section":"Section 4.6, Eqs. (14), (24), (25)"}],"minor_comments":[{"comment":"The definition of XS-R2 in Eq. (21) appears to use the model predictions in the denominator, whereas a cross-sectional R-squared should be based on the variance of realized mean returns; please clarify the formula and its intended interpretation.","section":"Section 4.3, Eq. (21)"},{"comment":"The algorithm computes M_{t+1} = 1 - sum_i w_{t,i} R^e_{t+1,i} over a mini-batch, but the SDF is defined over the full cross-section in Eq. (14); please specify how the batch-level SDF is formed and how the unbalanced panel is handled in the loss.","section":"Algorithm 1, step 7"},{"comment":"The loss in Eq. (18) averages firm-level moments, while Algorithm 1 step 9 computes a loss over individual observations; please align the notation and explain the aggregation rule.","section":"Eqs. (17)-(18)"},{"comment":"The claims of \"14.2% annual excess returns\" and \"18-32% reduction in pricing errors\" are presented without standard errors, sample sizes, or a description of the statistical test; please add the necessary detail.","section":"Section 4.7"},{"comment":"The Shapley value method in Eq. (26) is attributed to reference [5], a paper on predicting returns with text data; please cite the original Shapley value source and define the notation more precisely.","section":"References"},{"comment":"The table entry for Elasticnet-SDF is attributed to \"Kelly et al., 2019\" in the table body, which appears inconsistent with the reference list entry [42] for Kozak, Nagel, and Santosh; please correct the attribution.","section":"Table 1"},{"comment":"Figures 1 and 2 present very similar architecture diagrams; please consider merging them or clearly distinguishing the information each figure is intended to convey.","section":"Figures 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's main contribution is potentially interesting, but the empirical core is not currently verifiable. The underspecified news pipeline and the absence of any uncertainty quantification mean that the 2.80 Sharpe ratio cannot be assessed on the information provided. I would encourage the editor to request the code and data pipeline as part of the revision, and to ask for a clear statement of what was fit in-sample and what was used out-of-sample. The paper would also be stronger if the authors addressed the circularity of the beta-sorted decile exercise directly, since that section currently overstates what is shown."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the NewsNet-SDF paper. The combination is genuinely new: pretrained language model news embeddings fused into an adversarial SDF estimator, with macro conditioning and characteristic ranking. That is a sensible next step after Chen-Pelger-Zhu and a natural extension of the Ke-Kelly-Xiu text predictor line. The architecture is clean, the GMM/adversarial framing is standard, and the ablation pointing to news as the main driver is at least internally coherent. Credit where due: the authors know the asset pricing machinery and the writing is mostly precise where it bothers to be. The paper deserves a serious referee.\n\nThat said, the empirical claim as stated is not supported. A 2.80 out-of-sample Sharpe ratio on U.S. equities is an extraordinary number, and extraordinary numbers require airtight leakage controls. The paper does not provide them. The \"careful temporal alignment\" in Section 4.1 is a promise, not a description: no rule for matching article timestamps to monthly return periods, no handling of multi-firm articles, no treatment of firm-months with no news. The PCA projection in Section 3.1 never states whether loadings are fit on training data only. The stress-test note is right that a leak in either place would directly manufacture the headline result. The absence of confidence intervals or any significance test matters too: point estimates of Sharpe ratios this far from 1.0 can still be noise, especially in a long-horizon panel with correlated shocks.\n\nThe figure inconsistencies are minor individually but symptomatic. Figure 3 starts in 1970, so it shows training-period fit, not out-of-sample performance. Figure 5 runs to 2024 even though the stated test window ends in 2022. And the beta-sorted decile monotonicity in Figure 5 is largely internal consistency: with M = 1 - w'R, model-implied beta is proportional to the model's own predicted return, so sorting on beta is sorting on the forecast. That does not confirm the forecasts are right; it only confirms the model is not self-contradictory.\n\nNone of this is fatal. All of it is fixable with code and data release, explicit PCA-fitting and news-alignment documentation, and a robustness section with standard errors or block bootstrap. The method is plausible and the paper is honest about its building blocks. I would send it to review, but I would condition acceptance on those releases and on a rewritten empirical section that does not oversell what the current exhibits show.","headline":"A credible architectural novelty undermined by an under-specified news pipeline and no code; the 2.80 Sharpe is a promise, not yet a result.","tokens_in":11587,"tokens_out":1101,"would_cite":false,"duration_ms":13052,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B25","91G10","68T50"],"pacs":[],"model":"deepseek-v4-flash","headline":"NewsNet-SDF claims a stochastic discount factor built from news embeddings and adversarial networks prices U.S. equities out-of-sample with a 2.80 Sharpe ratio.","keywords":["Stochastic Discount Factor","Pretrained Language Models","Adversarial Networks","Financial News Embeddings","Asset Pricing","Multimodal Financial Analysis","Deep Learning","SDF Estimation"],"falsifier":"Re-run the test period using only news articles published strictly before the start of month $t$, and refit the PCA on training data alone; if the Sharpe ratio falls from 2.80 to near the no-news ablation value (1.65), the look-ahead-free alignment is doing the work. Alternatively, shuffle news dates across months by a random permutation; a surviving 2.80 Sharpe would indicate the text signal is not date-specific.","tokens_in":10591,"feed_emoji":"📰","tokens_out":9964,"duration_ms":85939,"temperature":0.7,"pith_summary":"This paper tries to establish that unstructured financial news can be folded directly into a stochastic discount factor (the pricing kernel that converts uncertain payoffs into today's prices) using pretrained language-model embeddings and adversarial training. On U.S. stock data from 1980-2022, the resulting NewsNet-SDF model reports an out-of-sample Sharpe ratio of 2.80, beating CAPM by 471% and the five-factor model by 460%, with pricing errors 74% lower than the five-factor benchmark. The paper argues that news text contributes unique systematic risk information: removing news cuts the Sharpe ratio by 41%, while removing macroeconomic features cuts it by 31%. If correct, the result would mean that narrative information, not just numerical characteristics, belongs inside the pricing kernel rather than as a separate sentiment overlay.","feed_headline":"News-driven pricing kernel hits 2.80 Sharpe, beating CAPM by 471%","feed_subtitle":"Language-model news embeddings cut pricing errors 74% versus the five-factor model on U.S. equities.","key_machinery":"The load-bearing mechanism is the adversarial minimax objective $\\min_\\phi \\max_\\psi \\frac{1}{N}\\sum_j \\| E[M_{t+1} R^e_{t+1,j} g_{t,j}] \\|^2 + \\lambda(\\|\\phi\\|_2^2 + \\|\\psi\\|_2^2)$, which converts the generalized-method-of-moments orthogonality condition $E[M_{t+1}R^e_{t+1}g_t]=0$ into a game: the SDF network minimizes pricing errors, while a conditional network generates instruments $g_t$ that challenge the SDF where it prices worst. Text enters through GTE-multilingual sentence embeddings aggregated by learned self-attention, projected with principal components, and concatenated with an LSTM macro state and ranked firm characteristics.","core_discovery":"The central claim is that a stochastic discount factor $M_{t+1} = 1 - \\sum_i w_{t,i} R^e_{t+1,i}$ can be learned adversarially from fused features consisting of an LSTM-processed macroeconomic state, cross-sectionally ranked firm characteristics, and self-attention-aggregated, PCA-projected news embeddings from a multilingual transformer. In the out-of-sample period 2000-2022, the model achieves a Sharpe ratio of 2.80, an MSPE of 0.56, and a cross-sectional $R^2$ of 0.11, with $\\beta$-decile portfolios monotonically increasing in returns ($R^2 > 0.95$). The paper further asserts that news-derived principal components rank among the most influential drivers of SDF dynamics, and that stocks most sensitive to news embeddings earned 14.2% annual excess returns after controlling for standard risk factors.","pith_inferences":["A testable extension is to apply the same adversarial instrument-generator mechanism to other unstructured data (analyst reports, regulatory filings, social media) or other pricing equations, which the paper leaves for future work.","The claimed 2-3 week lead time of news narratives suggests a real-time trading strategy could rebalance on news embeddings faster than monthly factor models; the paper does not test transaction costs or capacity.","An immediate robustness check is whether the 2.80 Sharpe survives the cost of translating the SDF weights into a tradeable long-short portfolio, since the paper reports the raw SDF portfolio's Sharpe.","If the lead-time result generalizes, text-based embeddings could serve as early-warning indicators for regime shifts in risk premia, which would be relevant to risk management systems."],"forward_implications":["If the model's pricing kernel is right, news text functions as a systematic risk factor, not just an alpha signal, and should be included in asset-pricing tests.","The adversarial network supplies test assets adaptively, so pricing errors and anomalies are evaluated against the hardest-to-price conditions the data can express.","The reported cross-sectional relation implies that firms with news-embedding sensitivity earn roughly 14.2% higher annual excess returns even after controlling for standard factors.","The ablation pattern implies that adding news helps more than adding macroeconomic state, suggesting text is not redundant with standard characteristics.","The monotone decile pattern implies the estimated SDF preserves risk-return ordering out of sample, including through the 2008 crisis and COVID-19."],"supporting_citations":[{"why":"Supplies the IPCA-SDF baseline and the sample restriction to stocks with complete characteristics; NewsNet-SDF is benchmarked against it.","marker":"[4]"},{"why":"Establishes the GMM orthogonality condition E[M R^e g]=0 that the adversarial training objective directly implements.","marker":"[10]"},{"why":"Provides the GTE-multilingual pretrained language model used to generate news sentence embeddings.","marker":"[33]"},{"why":"Supplies the machine-learning asset-pricing baseline (FFN) and evidence that neural networks outperform linear predictors.","marker":"[3]"},{"why":"Defines the deep-learning asset-pricing framework and the evaluation metrics used in the experiments.","marker":"[29]"},{"why":"Gives the text-based return prediction precedent and the attribution method used for news embedding importance.","marker":"[5]"},{"why":"The five-factor model that serves as the principal baseline for the 74% pricing-error reduction claim.","marker":"[41]"},{"why":"The elastic-net SDF baseline representing shrinkage-based modern SDF estimation.","marker":"[42]"}],"fun_headline_variants":["AI reads 2.5M news articles to hit 2.80 Sharpe in asset pricing","Multimodal SDF beats CAPM by 471% with news embeddings","News-aware pricing model slashes errors 74% vs Fama-French","Adversarial network fuses news and macros for 2.80 Sharpe","Language model news lift Sharpe to 2.80, top all benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 2.80 Sharpe ratio rests on the assumption that the news text assigned to each firm-month contains no information from after that month's returns are realized—that the temporal alignment has no look-ahead.","fun_headline_variants_meta":{"raw":{"variants":["AI reads 2.5M news articles to hit 2.80 Sharpe in asset pricing","Multimodal SDF beats CAPM by 471% with news embeddings","News-aware pricing model slashes errors 74% vs Fama-French","Adversarial network fuses news and macros for 2.80 Sharpe","Language model news lift Sharpe to 2.80, top all benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000471,"raw_usage":{"total_tokens":2386,"prompt_tokens":1028,"completion_tokens":1358,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":1254}},"tokens_in":644,"tokens_out":1358,"duration_ms":9592,"temperature":1.0,"reasoning_tokens":1254,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:31:09.407826+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the test period using only news articles published strictly before the start of month $t$, and refit the PCA on training data alone; if the Sharpe ratio falls from 2.80 to near the no-news ablation value (1.65), the look-ahead-free alignment is doing the work. Alternatively, shuffle news dates across months by a random permutation; a surviving 2.80 Sharpe would indicate the text signal is not date-specific.","supporting_citations":[{"cited_title":"Journal of Financial Economics134(3), 501–524 (2019)","cited_arxiv_id":null,"evidence_quote":"Supplies the IPCA-SDF baseline and the sample restriction to stocks with complete characteristics; NewsNet-SDF is benchmarked against it."},{"cited_title":"Econometrica50(5), 1269–1286 (1982)","cited_arxiv_id":null,"evidence_quote":"Establishes the GMM orthogonality condition E[M R^e g]=0 that the adversarial training objective directly implements."},{"cited_title":"The Review of Financial Studies33(5), 2223–2273 (2020)","cited_arxiv_id":null,"evidence_quote":"Supplies the machine-learning asset-pricing baseline (FFN) and evidence that neural networks outperform linear predictors."},{"cited_title":"Management Science 70(2), 714–750 (2024)","cited_arxiv_id":null,"evidence_quote":"Defines the deep-learning asset-pricing framework and the evaluation metrics used in the experiments."},{"cited_title":"NBER Working Paper No","cited_arxiv_id":null,"evidence_quote":"Gives the text-based return prediction precedent and the attribution method used for news embedding importance."},{"cited_title":"Journal of Financial Economics116(1), 1–22 (2015)","cited_arxiv_id":null,"evidence_quote":"The five-factor model that serves as the principal baseline for the 74% pricing-error reduction claim."},{"cited_title":"Journal of Financial Economics135(2), 271–292 (2020)","cited_arxiv_id":null,"evidence_quote":"The elastic-net SDF baseline representing shrinkage-based modern SDF estimation."}],"review_version":1}