{"id":"94d6570d-1b8c-4dd3-bb09-2d16ebbe9137","arxiv_id":"2508.14999","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LSTM-based variance-covariance estimation produced the best mean-variance portfolios in a backtest against DeepVAR, GPVAR, and classical estimators, with longer training windows and slower rebalancing favoring the deep learning models.","lead":"This paper tests whether deep learning forecasts of covariance matrices build better stock and crypto portfolios than classical methods. Its headline finding, that LSTM models generally win on risk-adjusted return, holds only under specific data windows and rebalancing schedules.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LSTM advantage is read off a parameter grid with no significance testing or multiple-comparison control; need a held-out matched-cell resampling check.","rationale":"The abstract is the only readable portion; the full text is mojibake. Treating every part as in-scope, the abstract's own hedge ('performance differed significantly between different combinations of parameters') is the sharpest evidence of the weak point. The central empirical claim is an inference from a grid search over two tuning dimensions. For that inference to be sound, the winner must be robust to selection over the grid and to the unequal/overlapping test windows implied by changing observation lengths. The reader identified the same primary weakness (configuration selection), so agreement_with_reader = agree. My proposed check is deliberately concrete: a fixed holdout plus paired bootstrap with multiple-comparison correction. If the check fails, the paper should be read as a set of descriptive backtests rather than a demonstrated LSTM advantage. Since no readable full text and no supporting artifacts are available, I would not move the reader's UNVERDICTED verdict: the appropriate disposition remains 'unverified until the protocol is clarified or recomputed.'","tokens_in":26244,"tokens_out":6868,"duration_ms":86169,"concrete_test":"Reproduce the backtest with a fixed held-out test period that is identical across all configurations. Use a validation slice of the training period only to pre-select one (window, rebalancing) cell per method. Then, on the held-out period, compute paired differences in information ratio and annualized return between the LSTM strategy and the best classical covariance strategy, using a stationary bootstrap over non-overlapping rebalance blocks and applying Holm correction across all cells. If the LSTM advantage does not remain significant (e.g., bootstrapped 95% CI for the paired difference excluding zero after correction), the abstract's 'generally best' claim is not supported. A secondary check: evaluate the same models on equal-length out-of-sample windows to test whether the longer-window trend survives controlling for test length.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim — that LSTM-RNN covariance forecasts 'generally' produce the best information ratio and annualized returns — requires that the advantage be structural, not a winner-takes-all reading of a configuration grid. The abstract concedes that performance 'differed significantly between different combinations of parameters,' and it describes no pre-registered selection rule, no multiple-comparison correction, and no significance test. When many (observation window, rebalancing period) cells are scanned, at least one deep-learning cell can top the table by chance; the more cells, the more likely. There is a second related confound if the total sample is fixed: increasing the observation window shortens and changes the out-of-sample period used to score each model, so the claimed 'longer observation windows help deep learning' trend may reflect unequal test lengths/regimes rather than model capability. The supplied full text is corrupted, so I cannot rule out that the paper does contain a holdout or resampling analysis; but the readable abstract gives no such evidence, and the central claim stands or falls on exactly this.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for estimating the variance-covariance matrix used in Markowitz portfolio optimization, using LSTM-RNN, DeepVAR, and GPVAR one-day-ahead multivariate forecasts. These forecasts are fed into a mean-variance optimizer to construct portfolios of stocks and cryptocurrencies. The authors compare performance across a grid of observation windows and rebalancing periods, and conclude that LSTM-RNN models generally achieve the best information ratio and annualized returns, that longer observation windows benefit deep learning models, and that less frequent rebalancing tends to improve performance.","tokens_in":26459,"tokens_out":3703,"duration_ms":47470,"significance":"If the central claim is correct, the paper provides useful evidence that a relatively simple LSTM-based covariance forecast can outperform classical covariance estimators and two deep generative baselines in a realistic multi-asset portfolio setting. The comparison across multiple deep learning architectures and classical methods is a strength, and the problem is of practical importance. However, the claim is entirely empirical and depends on the evaluation protocol being unbiased and statistically sound; the abstract, which is the only readable part of the submission, does not demonstrate that.","major_comments":[{"comment":"The headline conclusion that 'generally the best results... are obtained using the LSTM-RNN models' is read off a grid of observation windows and rebalancing periods. The abstract concedes that performance 'differed significantly between different combinations of parameters' but reports no significance tests, no multiple-comparison correction, and no pre-specified selection rule. With many cells, the best cell can arise by chance. To make the structural claim, the authors need a matched-cell analysis with a paired test (e.g., Diebold-Mariano on information-ratio differences) and a correction for multiple comparisons, plus reporting of all cell results rather than a summary 'generally.'","section":"Abstract"},{"comment":"The conclusion that 'longer observation windows translate into better performance of the deep learning models' may be confounded by unequal out-of-sample periods. If the total sample is fixed, increasing the observation window shortens the test period and changes the regime mix, so the apparent improvement may reflect test-length or time-period differences rather than model capability. A rolling-origin evaluation with a fixed test length per configuration, or at least reporting the exact out-of-sample dates for each window, is necessary to support this claim.","section":"Abstract: observation windows"},{"comment":"The abstract does not state how expected returns are estimated. A Markowitz optimizer requires a mean return forecast in addition to the covariance matrix; if the mean estimator differs across methods, the LSTM advantage could stem from mean estimation rather than covariance estimation. If the same mean estimator is used for all methods, this must be stated explicitly. In addition, transaction costs and rebalancing turnover are not mentioned; without a transaction-cost model, the information ratio and annualized return comparisons may not reflect realizable performance, especially for high-turnover deep learning strategies.","section":"Abstract: portfolio construction"},{"comment":"The supplied full text is unreadable due to encoding corruption (mojibake); equations, data descriptions, hyperparameter settings, and implementation details cannot be verified. I cannot determine whether significance testing, transaction-cost adjustments, or other important details are present. A resubmission must include a readable full text. Without it, the paper cannot be properly evaluated.","section":"Full text"}],"minor_comments":[{"comment":"The phrase 'differed significantly between different combinations of parameters' is ambiguous: it may mean 'substantially' or 'statistically significantly.' Please disambiguate in the text.","section":"Abstract"},{"comment":"The term 'variancecovariance' is missing a hyphen; should be 'variance-covariance.'","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The central issue is statistical inference on a configuration grid. Even with a readable text, the paper likely needs a substantial addition of significance testing and multiple-comparison control. The corrupted full text is a blocking issue for further review; please ask the authors to provide a properly encoded manuscript. The novelty is moderate, but the problem is relevant for a quantitative finance or machine learning audience."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate empirical comparison of deep-learning covariance estimators in a Markowitz optimizer, but the headline finding rests on a grid search with no visible inferential controls. I couldn't audit the full text because the supplied source is mojibake, so this take is driven by the abstract and the stress-test note.\n\nWhat is actually here: LSTM, DeepVAR, and GPVAR for one-day-ahead covariance forecasting, fed into mean-variance optimization, on stocks and crypto, across a grid of observation windows and rebalancing periods. That is a reasonable extension of an active line of work rather than a new technique. The abstract's statements that longer windows and less frequent rebalancing work better are concrete and actionable, and the study honestly reports that performance 'differed significantly between different combinations of parameters.'\n\nThe soft spot is the inference. The abstract reads a winner (LSTM) off the same grid that was swept, without stating a selection rule, a multiple-comparison correction, or a significance test. The stress-test note about unequal test lengths is real: if the total sample is fixed, longer windows shorten and shift the out-of-sample period, so the 'longer windows help' trend may be an artifact of which regimes each model gets tested on. I can't confirm that the full text lacks a holdout or resampling analysis, because the encoding makes it unreadable as supplied. But the abstract alone doesn't rule it out, and the central claim stands on exactly that.\n\nMinor, from the abstract: transaction costs and look-ahead bias controls aren't visible. That's not fatal; it's just something I can't score.\n\nWho this is for: practitioners using mean-variance optimization who want a quick read on whether LSTM covariance forecasts are worth trying. It's not a theoretical contribution. If the full text is intact, it deserves a serious referee - the question is practical and the comparison is informative even if the winner's margin is small. I'd want the referee to require significance testing or a matched-cell resampling check before publication. I would not cite this yet.","headline":"Legitimate empirical comparison, but the LSTM headline is read off a parameter grid with no visible inference; it needs a holdout or resampling check before I'd trust it.","tokens_in":26963,"tokens_out":1845,"would_cite":false,"duration_ms":21270,"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":"Deep-learning covariance forecasts, especially from LSTM recurrent networks, lift mean-variance portfolio performance on stocks and cryptocurrencies.","keywords":["portfolio optimization","variance-covariance matrix estimation","long short-term memory","deep learning","Modern Portfolio Theory","cryptocurrency","information ratio","rebalancing frequency"],"falsifier":"Fix the grid of observation windows and rebalancing periods before running the experiment, then apply a significance test for out-of-sample information-ratio differences between LSTM and classical-covariance portfolios; if the LSTM advantage shrinks to noise after multiple-comparison correction, the structural claim fails. A second check is to hold out the configuration-selection period and compare only on data never used to pick the winner.","tokens_in":26112,"feed_emoji":"📈","tokens_out":4970,"duration_ms":55720,"temperature":0.7,"pith_summary":"This paper tries to establish that the variance-covariance matrix used in Markowitz-style portfolio optimization can be improved by replacing classical estimators with one-day-ahead forecasts from deep learning models. It compares two probabilistic deep models, DeepVAR and GPVAR, and an LSTM recurrent neural network against classical covariance estimates, using the forecast matrices as inputs to a mean-variance optimizer for portfolios of stocks and cryptocurrencies. Across a grid of observation windows and rebalancing periods, LSTM-RNN portfolios generally deliver the highest information ratios and annualized returns. The paper also argues that deep learners need longer training windows to learn covariance structure and that less frequent rebalancing outperforms the shortest rebalancing intervals for every method tested.","feed_headline":"LSTM covariance forecasts beat classical estimators in portfolio tests","feed_subtitle":"Deep-learning covariance inputs, with longer training windows and slower rebalancing, produce the best stock and crypto portfolios.","key_machinery":"The carrying mechanism is the one-day-ahead variance-covariance matrix forecast. The paper trains three deep learners—an LSTM recurrent neural network, and two probabilistic sequence models, DeepVAR and GPVAR—on asset return histories, then feeds their forecast covariance matrices into a mean-variance optimizer. LSTM is a recurrent network that maintains a memory state across time steps, which the paper argues lets it capture long-term dependencies in the covariance structure. The classical sample-covariance estimators are the baselines the deep forecasts must beat.","core_discovery":"The central claim is that a long short-term memory recurrent neural network trained to forecast the multivariate return covariance matrix one day ahead is a better input to mean-variance portfolio optimization than classical covariance estimators, DeepVAR, or GPVAR. The evidence is the paper's out-of-sample comparison: portfolios built from LSTM forecasts usually rank first by information ratio and annualized return, with the advantage concentrated at longer observation windows and slower rebalancing schedules. The discovery is an empirical extension claim: deep-learning covariance forecasting, and LSTM in particular, can serve as the statistical engine inside the classical Markowitz optimiz","pith_inferences":["The winner is read off the same grid over which methods were compared, with no pre-registered configuration rule or multiple-comparison correction; a fair structural test would fix the grid in advance and test LSTM's edge on a hold-out sample.","Because Markowitz portfolios are highly sensitive to expected-return errors and the paper only replaces the covariance estimate, the LSTM advantage might not survive in a full pipeline where mean returns are also forecast.","A natural next test is to pit LSTM covariance forecasts against shrinkage estimators and dynamic conditional correlation models, which are standard covariance baselines not covered by this comparison.","The one-day-ahead design says little about longer-horizon investors; multi-day forecast covariance matrices may erase or reverse the LSTM edge."],"forward_implications":["If LSTM covariance forecasts genuinely beat classical estimates, practitioners can improve Markowitz portfolios without changing the optimizer—only the covariance input.","Longer observation windows should be used for deep covariance models, since the paper finds their performance degrades with shorter histories.","Daily rebalancing is the worst schedule tested; allowing positions to run longer improves information ratio and annualized return across all methods.","DeepVAR and GPVAR are viable but generally inferior to plain LSTM for this covariance-forecasting task, suggesting probabilistic sequence models are not automatically better than a well-tuned recurrent network."],"supporting_citations":[],"fun_headline_variants":["LSTM covariance forecasts beat classical in portfolio tests","LSTM outperforms classical covariance estimation in portfolios","Deep learning covariance beats classical in Markowitz portfolios","Longer windows, slower rebalancing boost LSTM portfolio gains"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The central claim rests on treating the best configuration found on the searched grid as evidence of a structural LSTM advantage, with no pre-registered selection rule, multiple-comparison correction, or significance testing, and on assuming that a one-day-ahead forecast covariance matrix is enough for the optimizer to beat classical estimates.","fun_headline_variants_meta":{"raw":{"variants":["LSTM covariance forecasts beat classical in portfolio tests","LSTM outperforms classical covariance estimation in portfolios","Deep learning covariance beats classical in Markowitz portfolios","Longer windows, slower rebalancing boost LSTM portfolio gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000797,"raw_usage":{"total_tokens":3330,"prompt_tokens":719,"completion_tokens":2611,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":2548}},"tokens_in":463,"tokens_out":2611,"duration_ms":20421,"temperature":1.0,"reasoning_tokens":2548,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:10:21.340069+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix the grid of observation windows and rebalancing periods before running the experiment, then apply a significance test for out-of-sample information-ratio differences between LSTM and classical-covariance portfolios; if the LSTM advantage shrinks to noise after multiple-comparison correction, the structural claim fails. A second check is to hold out the configuration-selection period and compare only on data never used to pick the winner.","supporting_citations":[],"review_version":1}