{"id":"0fd7563f-9a8d-41d5-94f5-b6e1956c05ab","arxiv_id":"2608.09880","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"FinATOM uses a causal language model's token vocabulary to output stock forecasts and ETF allocations, and shows policy optimization improves out-of-sample Sharpe.","lead":"A single small language model, prompted with prices and news, emits three-day stock forecasts and five-ETF portfolio weights directly as text tokens, with no separate prediction layer. The paper reports higher out-of-sample Sharpe ratios after a reinforcement-learning stage, suggesting that language models can serve as complete financial decision interfaces.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 2025 multimodal edge may be an artifact of a post-hoc news summarizer: DeepSeek-V4 was trained after every test period, and the time-series-only ablation cannot distinguish genuine news signal from retroactive summarization.","rationale":"The reader's weakest assumption covers both the Llama base-model cutoff and the DeepSeek-V4 summarizer. I agree with the broad leakage concern but single out the summarizer as the load-bearing element because it coincides with the 2025 news contribution, the paper's clearest multimodal result. The base-model cutoff is less decisive for the central SFT-to-policy comparison: 2023 shows zero improvement, and 2024/2025 postdate Llama's disclosed cutoff, so the pooled gain does not primarily ride on base-model memorization. The summarizer, by contrast, was trained after every test period and can retroactively encode 2025 outcomes into the exact prompts where the multimodal advantage appears. The time-series-only ablation cannot settle this because it deletes the entire news channel rather than isolating hindsight. The paper is transparent about this residual channel and explicitly proposes re-summarization as a check, but until that check is run the +0.190 2025 news edge and the 1.653 Sharpe should be treated as unverified. This does not overturn the feasibility claim, which also rests on the time-series-driven policy gains, but it does mean the conditional verdict remains appropriate: the headline multimodal result needs an external-validity bound before it is cited as evidence that news helps in 2025.","tokens_in":16299,"tokens_out":4722,"duration_ms":47852,"concrete_test":"Replace DeepSeek-V4 summaries with summaries produced by a model with a documented pre-2023 training cutoff (or a non-parametric extractive summarizer over the same articles), keeping every other pipeline element fixed, and rerun the 2025 allocation test. If the multimodal 2025 Sharpe drops from 1.653 toward the time-series-only 1.463 while the time-series-only arm is unchanged, the news advantage is leakage; if it stays near 1.6, the concern is bounded and the conditional verdict can be upgraded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is causal validity of the news channel. Section 4 states that daily news is summarized by DeepSeek-V4, whose training postdates all test periods. The strongest standalone results--especially 2025 multimodal Sharpe 1.653 vs. time-series-only 1.463 in Table 4--could therefore reflect the summarizer encoding post-hoc knowledge of 2025 market outcomes into the prompt. This is not bounded by the time-series-only ablation: removing all news removes the legitimate signal and the hindsight channel together, so it cannot show that the +0.190 2025 news contribution is causal. The base-model cutoff is less critical for the main SFT-to-policy delta: 2023 contributes no improvement, and 2024/2025 postdate Llama's disclosed cutoff. But the summarizer channel sits exactly on the period where the news contribution is largest. The paper acknowledges this in Section 6 and Supplementary Q15, but it remains unquantified. If the channel is real, the 'clearest advantage in 2025' and the multimodal headline are not evidence of learned forecasting.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FinATOM studies whether a causal language model can emit financial forecasts and allocations directly through constrained token generation, without a task-specific regression or policy head. The forecasting contract autoregressively outputs volatility-standardized D1–D3 return bins and is trained with ordinal and ranking supervision plus a one-epoch policy stage; the allocation contract emits five normalized long-only ETF weights, is first fine-tuned to imitate a causal mean–variance anchor, and is then refined by DAPO-augmented GRPO against realized 21-day Sharpe with an anchor-consistency penalty. In 2023–2025 ETF tests the policy raises pooled gross Sharpe from 1.428 to 1.529 and net Sharpe from 1.394 to 1.494 under 5-bp costs, and the multimodal input has its clearest advantage in 2025. On FinTexTS the SFT and policy strategies achieve 73.52%/2.68 and 73.72%/2.69 cumulative-return/Sharpe. The paper frames the contribution as feasibility of the head-free interface, not superiority over head-based systems.","tokens_in":16545,"tokens_out":5577,"duration_ms":54272,"significance":"The paper is carefully framed and unusually honest about its limits: the causal mean–variance teacher is explicitly separated from future returns, the path-dependent anchor is recomputed sequentially within each experiment, and the discussion and supplementary material identify the base-model cutoff, single-seed scope, and missing significance tests. If the headline results hold, the work would be a useful demonstration that one tied-vocabulary token interface can support two different numerical contracts and that outcome-based RL can refine an imitated allocation without a separate policy head. The reported effect sizes, however, are not yet trustworthy as evidence of a learned edge because the strongest multimodal result sits exactly on an unquantified news-summarizer hindsight channel and all Sharpe improvements come from single runs. The contribution is therefore plausible and potentially useful, but the central empirical claim needs additional robustness evidence before it can be regarded as established.","major_comments":[{"comment":"The news channel is summarized by DeepSeek-V4, whose training postdates every test period, and the single ablation that removes news also removes any legitimate news signal, so the +0.190 2025 gap between news+time-series and time-series-only cannot distinguish genuine macro-news information from retroactive summarization. Because the 2025 multimodal contribution is the paper's 'clearest advantage' and is part of the pooled policy result, this is a load-bearing threat to the central feasibility claim. The paper should report a re-summarized ablation using a summarizer trained before the test periods, or equivalently bound the channel by comparing policies trained without news on 2025 and showing the pooled gain survives.","section":"Section 4, Table 4; Section 6 and Supplementary Q15"},{"comment":"The headline pooled Sharpe improvement of 1.428 to 1.529 (gross) and 1.394 to 1.494 (net) is based on a single random seed and a single walk-forward run per setting, with no confidence intervals or significance tests. The SFT and policy strategies produce paired daily return series on identical dates, so the paper should report a paired Sharpe-difference test (e.g., Jobson–Korkie with Memmel correction or a Ledoit–Wolf bootstrap) and, ideally, repeated seeds. Without this, the observed differences in 2024 and 2025—and the sharp 2025 news gap—could be within single-run noise; the paper's own Q11 concedes this, but the consequence for the main quantitative claim is not mitigated.","section":"Section 5, Tables 2 and 3"}],"minor_comments":[{"comment":"The caption says 'no completed forecasting policy-stage result is claimed,' which is confusing because the abstract and Section 5 report the FinATOM Policy result; please clarify that this refers to a fuller DAPO-style pipeline rather than the evaluated one-epoch policy stage.","section":"Figure 3 caption"},{"comment":"The 'Net 5 bp' column is empty for the Causal target and Equal weight rows; please either compute the same net values or state explicitly why the calculation is unavailable for those reference rows.","section":"Table 3"},{"comment":"The statement that the final 20 training dates are excluded because they lack a fully contained future reward window should clarify that this exclusion concerns the policy-stage reward computation, since the SFT target itself does not require future data.","section":"Section 4, Portfolio allocation"},{"comment":"The deployment setup conditions on the previous anchor state rather than the model's own previous allocation; this is an important qualification for deployability and should be stated in the main text, not only in the supplementary material.","section":"Supplementary Q19"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest and the Q&A is unusually thorough, but the two major issues interact: if the news-summarizer hindsight channel is real, the 2025 multimodal result—and the part of the pooled improvement it contributes—would not be evidence of learned forecasting. The single-seed issue compounds this because no statistical test can currently separate signal from noise. I would encourage the editor to ask for the re-summarization robustness check and a paired Sharpe-difference test before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely honest feasibility study of a head-free token interface for financial forecasting and allocation. The specific combination—one LM vocabulary for ordered return forecasts and constrained portfolio weights, SFT from a causal mean-variance teacher, then token-level GRPO/DAPO against realized Sharpe—is new as far as I know. The paper earns credit for transparency: it clearly separates within-protocol comparisons from contextual baselines, documents the leakage controls, and flags its own limitations, including the base-model cutoff and the post-hoc news summarizer.\n\nThe central structural result—that a 1B model can learn to emit legal allocations and that a one-epoch policy stage improves pooled gross Sharpe from 1.428 to 1.529—is plausible. The time-series-only arm still improves over the SFT baseline, so the head-free interface and policy optimization do real work even without news.\n\nThe soft spots are real but proportionate. First, single-seed, single-run walk-forward tests with no significance tests or confidence intervals. The authors admit this and even suggest the paired Sharpe tests that would apply. That is fine for a feasibility claim, but it caps how much weight the numbers can carry. Second, the news channel: DeepSeek-V4 was trained after all test periods, and the strongest 2025 multimodal result (1.653 vs. 1.463 time-series-only) could in principle reflect the summarizer encoding hindsight into the prompt. The time-series-only ablation cannot bound this because it removes the legitimate signal and the hindsight channel together. The paper acknowledges this in Section 6 and Q15, but it remains unquantified. This does not sink the central feasibility claim, but it should be fixed before anyone cites the 2025 news edge as evidence of learned forecasting.\n\nThe citation pattern looks fine: the authors cite the relevant LLM-time-series, quantized-output, policy-optimization, and financial-text work, and they are careful not to overclaim novelty. No circular derivation; the SFT teacher is causal, and future returns enter only through the RL reward.\n\nWho this is for: researchers working on LLM-based numerical decision-making, especially constrained token generation. It is a useful reference for the design space, and the honest limitation discussion is itself an example worth following.\n\nRecommendation: send it to review. It deserves a serious referee, but with the expectation of substantial revision: repeated seeds, paired significance tests on the Sharpe deltas, code/data release if possible, and re-summarization with a model whose training predates the test periods. I would not accept the current numerical claims at face value, but the paper's core feasibility question is legitimate and well-presented.","headline":"A transparent, honestly framed feasibility study of a head-free token interface for financial forecasting and allocation, but the headline Sharpe gains rest on single-seed runs and an unquantified news-summarizer leakage channel, so the numbers should not be taken at face value.","tokens_in":17031,"tokens_out":2395,"would_cite":true,"duration_ms":21666,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that a causal language model can emit both multi-step return forecasts and constrained portfolio weights through a single head-free token interface.","keywords":["financial forecasting","portfolio allocation","token generation","head-free interface","causal language model","mean-variance","GRPO","Sharpe ratio"],"falsifier":"Recompute the 2023 allocation Sharpe and the 2023 stock returns using a base model and a news summarizer whose training sets provably end before January 2023, keeping the pipeline otherwise identical. If those gains disappear, or if the original models can recite test-period prices or news they were not trained on, the causal-validity premise fails.","tokens_in":1879,"feed_emoji":"📈","tokens_out":2229,"duration_ms":77407,"temperature":0.7,"pith_summary":"FinATOM asks whether a causal language model can act as the numerical decision interface in finance, producing multi-step return forecasts and constrained portfolio weights by generating tokens, with no regression, ranking, or policy head added on top of the language model. The paper's central claim is that this is feasible, and it supports the claim with chronological ETF allocation tests in 2023-2025 plus an exploratory 100-stock forecasting test. In allocation, supervised fine-tuning on a causal mean-variance teacher followed by a DAPO-augmented GRPO policy stage raises pooled gross Sharpe from 1.428 to 1.529 and net Sharpe from 1.394 to 1.494 while barely increasing turnover. In forecasting, the same token interface sustains a top-20 strategy whose one-epoch policy refinement moves cumulative return from 73.52% to 73.72% and Sharpe from 2.68 to 2.69. A sympathetic reader would care because, if true, one auditable, parseable generation interface could replace several task-specific heads, and outcome-based training could improve risk-adjusted decisions directly through the language model's own tokens.","feed_headline":"Token generation lifts portfolio Sharpe from 1.43 to 1.53","feed_subtitle":"A causal language model emits ETF weights as tokens, and policy training improves net risk-adjusted return.","key_machinery":"The central object is the head-free token interface: forecast bins and allocation grid values are added vocabulary items scored by the tied LM readout, with no regression, policy, value, or reward head. Two training mechanisms carry the argument: (i) a causal mean-variance teacher that solves a 20-day-window long-only optimization with a turnover penalty, shrinkage toward equal weight, and 50-unit quantization, producing SFT targets that contain no future information; and (ii) a DAPO-augmented GRPO stage that samples legal allocation sequences, rewards their future 21-day Sharpe minus an $\\ell^1$ anchor-consistency penalty, and applies token-level clipped group-relative advantages. For forecasting, 41 train-only quantile bins with conditional-mean centers plus an ordinal and same-date ranking loss define the supervised contract, and a one-epoch token-PPO stage refines it.","core_discovery":"The paper's central claim is that a causal language model can emit both multi-step return forecasts and constrained portfolio allocations through one head-free token interface. For allocation, the model generates five long-only ETF weights as normalized token expectations; a strictly causal mean-variance teacher supplies supervised targets, and a DAPO-augmented GRPO stage optimizes realized 21-day Sharpe subject to anchor consistency. For forecasting, the model autoregressively emits D1-D3 volatility-standardized return buckets whose conditional means form a cross-sectional ranking score. The evaluation is deliberately framed as feasibility: the allocation policy improves pooled gross Sharpe from 1.428 to 1.529 and net Sharpe from 1.394 to 1.494 with turnover rising only 5.6 basis points per day, while the forecasting policy adds a small increment over its SFT initialization. The paper does not claim dominance over head-based systems and separates within-protocol evidence from contextual comparison with the published FinTexTS baseline.","pith_inferences":["A natural extension is to test whether the same grammar-plus-normalization interface enforces other constrained financial outputs, such as risk budgets, target volatility, or hedging ratios, without retraining the architecture.","The 2025 news advantage could be tested by re-summarizing the daily news with a model whose training predates the test periods; if the 2025 gap shrinks, part of that gain is summarizer hindsight rather than learned forecasting.","The anchor-consistency penalty coefficient of 8 is a fixed constant; a sensitivity sweep around it would show whether the policy gain is robust or contingent on that one design choice.","Because the causal-validity risk is concentrated in the 2023 allocation test and the 2023 stock test, a post-cutoff replication with a base model and summarizer trained strictly before the test window would cleanly separate learned forecasting from memorization."],"forward_implications":["Allocation policy optimization improves pooled gross Sharpe from 1.428 to 1.529 and net Sharpe from 1.394 to 1.494, with daily one-way turnover rising only about 5.6 basis points.","The same interface handles both ordered multi-step forecasts and jointly constrained portfolio weights, so a task-specific numerical head is not required for either contract.","Time-series context is the stable standalone signal, while news contributes most in 2025, suggesting the multimodal advantage is regime-dependent rather than uniform.","The policy can exceed its causal teacher in some periods, as in 2024 (1.564 vs. 1.407), while trailing it pooled (1.529 vs. 1.540), indicating outcome-based refinement rather than pure imitation.","Broader asset universes, longer post-cutoff evaluation periods, repeated seeds, and matched head-based baselines are the stated next steps needed to test generality."],"supporting_citations":[{"why":"Supplies the GRPO group-relative advantage estimator that the allocation policy stage builds on.","marker":"(Shao et al. 2024)"},{"why":"Provides the DAPO-style asymmetric clipping and token-level update adopted in the allocation policy stage.","marker":"(Yu et al. 2025)"},{"why":"Supplies the mean-variance portfolio-selection objective underlying the causal teacher.","marker":"(Markowitz 1952)"},{"why":"Establishes the equal-weight 1/N portfolio as the hard-to-beat reference used as a baseline.","marker":"(DeMiguel, Garlappi, and Uppal 2009)"},{"why":"Provides the Llama 3.2 1B backbone that both tasks adapt with LoRA.","marker":"(Dubey et al. 2024)"},{"why":"Supplies the LoRA low-rank adaptation method used to fine-tune the backbone.","marker":"(Hu et al. 2022)"},{"why":"Provides the FinTexTS dataset that pairs 100 stock histories with four-level news and filing text for the forecasting experiment.","marker":"(Lee et al. 2026)"},{"why":"Demonstrates serializing numerical sequences for language-model forecasting, the approach the forecasting interface extends.","marker":"(Gruver et al. 2023)"}],"fun_headline_variants":["Token generation lifts net Sharpe from 1.394 to 1.494","Causal LM emits weights as tokens; Sharpe hits 1.53","Head-free token generation: better returns, Sharpe 1.53","FinATOM: token-based forecasts and ETF allocation, Sharpe up","No heads, just tokens: LLM nets 1.49 Sharpe on ETFs"],"cache_read_input_tokens":19200,"weakest_assumption_plain":"The out-of-sample numbers assume that neither the pretrained Llama 3.2 base model nor the DeepSeek-V4 news summarizer secretly knows the test-period market outcomes, even though the 2023 tests overlap the base model's pretraining cutoff and the summarizer was trained after all test periods.","fun_headline_variants_meta":{"raw":{"variants":["Token generation lifts net Sharpe from 1.394 to 1.494","Causal LM emits weights as tokens; Sharpe hits 1.53","Head-free token generation: better returns, Sharpe 1.53","FinATOM: token-based forecasts and ETF allocation, Sharpe up","No heads, just tokens: LLM nets 1.49 Sharpe on ETFs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000637,"raw_usage":{"total_tokens":2962,"prompt_tokens":998,"completion_tokens":1964,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":1867}},"tokens_in":614,"tokens_out":1964,"duration_ms":13209,"temperature":1.0,"reasoning_tokens":1867,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:59:46.776410+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the 2023 allocation Sharpe and the 2023 stock returns using a base model and a news summarizer whose training sets provably end before January 2023, keeping the pipeline otherwise identical. If those gains disappear, or if the original models can recite test-period prices or news they were not trained on, the causal-validity premise fails.","supporting_citations":[{"cited_title":"The Review of Financial Studies , volume =","cited_arxiv_id":null,"evidence_quote":"Establishes the equal-weight 1/N portfolio as the hard-to-beat reference used as a baseline."}],"review_version":1}