{"id":"16722807-13d9-4c73-b22a-f7649c223c35","arxiv_id":"2505.03760","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A volatility-guided deep reinforcement learning framework splits Dow 30 stocks into aggressive, moderate, and conservative tiers and claims higher risk-adjusted returns than benchmarks in a two-year out-of-sample test.","lead":"This paper combines volatility forecasting with deep reinforcement learning to build three stock portfolios matched to different investor risk profiles. The authors report that these portfolios outperformed simple benchmarks on Dow 30 stocks during 2023-2024.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing tier-matched equal-weighted baseline confounds asset selection with DRL allocation; Table 1 cannot support the central outperformance claim.","rationale":"The reader's weakest_assumption identified the same core issue: the benchmarks use the full 30-stock universe, not the selected 10-stock subset, so the improvement cannot be attributed to the DRL allocation. This is not a matter of disagreeing with consensus; it is an internal experimental design flaw. The paper's own text confirms the benchmark universe mismatch. The missing control is directly testable and would settle whether the contribution of the PPO policy is real or just an artifact of stock selection. The recommendation to reject the paper as written remains appropriate because the central empirical claim is not established, although the proposed concrete test could provide a path toward conditional acceptance if the tier-matched control shows DRL still adds substantial risk-adjusted performance. No additional concerns about author conduct or novelty need to be raised; the issue is purely the missing experimental control.","tokens_in":4791,"tokens_out":1626,"duration_ms":17091,"concrete_test":"Re-run the Table 1 evaluation with an equal-weighted portfolio constructed from the same ten stocks in each GARCH tier (Aggressive, Moderate, Conservative), rebalanced at the same frequency and with the same 0.05% transaction cost. Compare the resulting Sharpe ratios, annual returns, and drawdowns to those of the DRL portfolios. If the tier-matched equal-weighted Sharpe ratios are statistically indistinguishable from the DRL Sharpe ratios (no overlapping confidence intervals or bootstrap test), the central claim that the DRL allocation adds value is unsupported. As a secondary check, run the MVO strategy on the same 10-stock tiers as well.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that volatility-guided DRL portfolios outperform MVO, the DJI index, and equal weighting. However, Table 1 compares the three DRL portfolios, each invested in only 10 stocks from a single GARCH volatility tier, against benchmarks that invest across the full 30-stock DJI universe. The paper explicitly notes this: 'All the DRL-based portfolio models significantly outperformed the benchmark models, which considered the entire pool of stocks for constructing the portfolios.' Thus, the observed outperformance could be entirely due to the preselection step selecting stocks that happened to rise during the 2023-2024 test window, rather than to the PPO allocation policy. No equal-weighted portfolio of the same 10 stocks per tier, and no MVO portfolio restricted to the same 10 stocks, is reported. Without such a tier-matched control, the causal attribution to the DRL agent is not identified. This is the load-bearing weakness: if the same-tier equal-weighting already achieves the reported Sharpe ratios (e.g., 1.811 for Moderate-DRL), the contribution of the DRL allocation step is unproven. The absence of any dispersion or significance measure across the five training runs further prevents assessing whether the DRL advantage over such a control would be robust.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a portfolio optimization framework that first forecasts stock volatility with a GARCH(1,1) model, sorts the 30 Dow Jones stocks into aggressive, moderate, and conservative tiers of 10 stocks each, and then trains a PPO agent to allocate capital within each tier. The out-of-sample evaluation covers January 2023 through December 2024 and compares the three tier-specific DRL portfolios against mean-variance optimization (MVO), the DJI index, and an equal-weighted portfolio. The paper reports that all three DRL portfolios achieve higher returns and Sharpe ratios than all three benchmarks and concludes that volatility-guided preselection combined with DRL allocation delivers consistent risk-adjusted performance.","tokens_in":5051,"tokens_out":6626,"duration_ms":60544,"significance":"If the result held, the framework would be a practical and interpretable way to tailor portfolio construction to investor risk profiles, and it would provide evidence that GARCH-based tiering improves DRL portfolio performance. The paper's strengths are its clear problem formulation, the reasonable use of GARCH for volatility-tier selection, and the inclusion of several standard benchmarks. However, the contribution is purely empirical, and the experiment as reported does not identify the source of the outperformance: there is no control portfolio restricted to the same selected assets, no statistical significance testing, and only a single two-year test window. The paper also omits key implementation details and provides no code or data, so the central claim is not yet established.","major_comments":[{"comment":"The reported comparison is confounded: the three DRL portfolios each invest in a 10-stock GARCH volatility tier, whereas the MVO, DJI, and Equal-Weighted benchmarks use the full 30-stock universe. The paper itself acknowledges this in Section 4.2 when it states that the benchmarks 'considered the entire pool of stocks for constructing the portfolios.' As a result, the higher returns and Sharpe ratios could be driven entirely by selecting stocks that rose during the 2023-2024 bull market, rather than by the PPO allocation policy. The authors should add, at minimum, an equal-weighted portfolio of the same 10 stocks in each tier and an MVO portfolio restricted to the same 10 stocks; without such tier-matched controls, the central claim of DRL-driven outperformance is not identified.","section":"§4.2, Table 1"},{"comment":"The out-of-sample evaluation is a single two-year window (January 2023 to December 2024). No multiple test periods, rolling windows, or different market regimes are considered, so the conclusion in Section 5 that the proposed portfolios deliver 'consistent risk-adjusted returns' is unsupported. The authors should either report performance over several out-of-sample intervals, including non-bull-market conditions, or explicitly restrict the claim to the tested window.","section":"§4.1"},{"comment":"The statement that 'All the DRL-based portfolio models significantly outperformed the benchmark models' is not backed by any statistical analysis. The text says all DRL models were trained five times and average results recorded, but no standard deviation, confidence interval, or significance test is reported for any metric. 'Significantly' is therefore not justified. The authors should report dispersion across training runs and perform a paired comparison, such as a block bootstrap or a Diebold-Mariano test on the daily return series.","section":"§4.2"},{"comment":"Several implementation details essential for evaluating the design are omitted: the GARCH(1,1) estimation window and refitting frequency, whether tier assignment uses training-period data only or the full sample (full-sample assignment would introduce lookahead bias), the threshold used to select the top and bottom 10 stocks, the Bayesian-optimization search ranges for PPO hyperparameters, and how the 0.05% transaction cost is applied. These omissions prevent replication and make it impossible to rule out data leakage in the stock-selection step.","section":"§3 and §4.1"}],"minor_comments":[{"comment":"The label 'Conversative' should be 'Conservative', and the title contains an odd space in 'V olatility'.","section":"Figure 1"},{"comment":"Equation (1) defines r_t = mu_t + epsilon_t, but the text immediately says the mean equation is assumed to be zero; please clarify whether mu_t is set to zero or estimated.","section":"Equation (1)"},{"comment":"The sentence 'epsilon_{t-1}^2 represents past squared shocks and sigma_{t-1}^2 represents and influences the future volatility' is grammatically garbled and should be rewritten.","section":"Equation (2) discussion"},{"comment":"No data or code repository link is provided; for reproducibility the authors should release the implementation, including the GARCH and PPO training code.","section":"General"},{"comment":"The MVO baseline is not described in enough detail; please state the covariance estimation method, lookback period, and rebalancing frequency, and specify whether the DJI benchmark is a price-return or total-return index.","section":"§4.1"},{"comment":"The reference to 'Sutton, 2018' should be to Sutton and Barto, 'Reinforcement Learning: An Introduction', second edition.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a short conference-style paper, and its current evaluation is not at the level expected for a serious journal. The missing tier-matched control is the central issue, and I would make its addition, together with proper significance testing and multi-period evaluation, a hard requirement for any resubmission. If the authors cannot supply these experiments, I would not support publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is a modest, clearly-written extension of the DRL-for-portfolio-optimization line, but its headline result—that GARCH-tiered PPO portfolios beat MVO, the DJI, and equal-weighting—is not actually supported by the experiments as reported. The missing control is the whole story.\n\nWhat's new and what it does well: the specific pipeline of GARCH(1,1) volatility tiering into aggressive/moderate/conservative buckets followed by a PPO agent is not something I've seen in exactly this form. That combination is a legitimate extension, not a new principle. The paper is readable, the problem framing is sensible, and the choice to tune hyperparameters with Bayesian optimization and to report transaction costs is reasonable. Credit where due: the authors do not oversell the novelty, and the writing is clear about what they did.\n\nThe load-bearing weakness: Table 1 compares each DRL portfolio (10 stocks from a single tier) against benchmarks that invest across all 30 stocks. The authors explicitly say the DRL models 'significantly outperformed' benchmarks that 'considered the entire pool of stocks.' With no equal-weighted portfolio of the same 10 stocks, and no MVO restricted to those same 10 stocks, the observed outperformance could come entirely from the tier selection—picking volatile stocks during a 2023-2024 bull market—rather than from the PPO allocation. That is not a minor omission; it is the difference between a paper that demonstrates something and one that doesn't. The stress-test note is correct on this.\n\nOther soft spots, in order of severity: a two-year out-of-sample window in a single market regime; no standard deviation or significance testing across the five training runs; and vague details on GARCH parameter estimation and hyperparameter search ranges. These are secondary but would need fixing in a revision.\n\nWho this is for: someone working on DRL asset allocation who wants a quick read on one more GARCH-plus-RL combination. It does not deserve a desk reject—there is a real question here and the pipeline is coherent—but it needs a serious referee who will insist on tier-matched baselines and dispersion measures before the claim can be believed.\n\nRecommendation: send it to peer review, but expect heavy revision. If the authors add the missing controls and the DRL advantage survives, the paper becomes a solid applied contribution. If the tier-matched equal-weight already matches those Sharpe ratios, the paper shrinks to a negative or null result—which is still publishable, just not as this paper currently frames it.","headline":"A readable empirical paper whose central outperformance claim is not established: the missing tier-matched equal-weight baseline confounds asset selection with DRL allocation.","tokens_in":5520,"tokens_out":880,"would_cite":false,"duration_ms":9453,"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":"GARCH tiers plus PPO beat MVO, the Dow index, and equal weighting on 2023-24 data.","keywords":["deep reinforcement learning","portfolio optimization","GARCH","volatility forecasting","asset pre-selection","investor risk profile","PPO","Dow Jones index"],"falsifier":"Rerun the 2023-2024 test on the same ten-stock aggressive, moderate, and conservative sets with a simple equal-weight or buy-and-hold strategy and no rebalancing. If that equal-weight tier matches or beats the reported PPO returns and Sharpe ratios, the central claim of DRL-driven outperformance is not supported; the tier-selection effect alone explains the result.","tokens_in":4641,"feed_emoji":"📈","tokens_out":5814,"duration_ms":45363,"temperature":0.7,"pith_summary":"The paper tries to establish that portfolio construction improves when asset selection is aligned with an investor's risk tolerance before a deep reinforcement learning agent allocates capital. Using GARCH volatility forecasts, the thirty Dow Jones stocks are sorted into aggressive, moderate, and conservative tiers, and a PPO agent is trained separately on each tier. On out-of-sample daily data from 2023 to 2024, all three investor-specific portfolios report higher annual returns, cumulative returns, and Sharpe ratios than mean-variance optimization, the Dow index, and equal weighting. The moderate tier posts the best risk-adjusted result with a Sharpe ratio of 1.81. A sympathetic reading of the reported evidence is that volatility-guided preselection plus DRL allocation produces consistent risk-adjusted outperformance.","feed_headline":"Volatility tiers plus PPO beat MVO, Dow, equal weights","feed_subtitle":"GARCH sorts the Dow 30 into three risk tiers; PPO then allocates. All three tiers beat MVO, the index, and equal-weight.","key_machinery":"The load-bearing mechanism is a two-stage pipeline. First, GARCH(1,1), $\\sigma_t^2 = \\omega + \\alpha \\epsilon_{t-1}^2 + \\beta \\sigma_{t-1}^2$, estimates each stock's volatility and assigns the Dow 30 into three risk tiers of ten stocks each. Second, a Proximal Policy Optimization (PPO) agent receives a state built from prices, the covariance matrix, and technical indicators, outputs portfolio weights, and earns the portfolio return as reward with 0.05% transaction costs. The GARCH tier defines the investable universe; PPO handles allocation within it.","core_discovery":"The central claim is that sorting assets by GARCH-based volatility before applying reinforcement learning changes the realized risk-return outcome: the aggressive tier of the ten most volatile stocks reaches 26.7% annual return, the moderate tier reaches a Sharpe ratio of 1.81, and the conservative tier keeps annual volatility near 10%. Because each benchmark allocates across the full thirty-stock universe, the paper interprets the gap as evidence that investor-specific preselection, not DRL alone, is what lets the agent reach better risk-return trade-offs.","pith_inferences":["The clean test the paper does not run is an equally weighted buy-and-hold version of each ten-stock tier. If equal-weight tiers also beat MVO and the Dow, the outperformance is driven by tier selection rather than by PPO; if DRL beats equal weight on the same tier, the allocation adds value. This is an editorial extension.","A natural next probe is varying the tier size or using rolling GARCH reclassification instead of a one-time split, which would show whether the reported margins are robust to the cutoff.","Because the 2023-2024 window was broadly rising, the paper leaves open what the same pipeline does in a falling or choppy market; a drawdown-heavy test window is a direct extension.","A stricter out-of-sample test would retrain GARCH and PPO at each rebalance so volatility tiers are estimated only from data available at decision time, rather than from the full historical sample."],"forward_implications":["Investors with different risk tolerances can receive distinct portfolios from the same 30-stock universe, and each portfolio reports a risk-return profile consistent with its tier's label.","The out-of-sample design, trained on 2010-2022 and tested on 2023-2024, suggests the reported outperformance is not limited to the training period.","The moderate tier's highest Sharpe ratio implies that a middle-volatility stock pool, not the most or least volatile, can yield the best risk-adjusted allocation.","If the reported results hold, dynamic allocation plus volatility-guided preselection offers a practical alternative to static mean-variance optimization on a well-known index.","All three DRL portfolios beating the Dow index and equal weighting on Sharpe ratio suggests the framework captures some time-varying allocation benefit beyond simple diversification."],"supporting_citations":[{"why":"Introduces the GARCH model used to forecast volatility and build the three risk tiers.","marker":"Bollerslev, 1986"},{"why":"Provides the volatility-forecasting reference for applying GARCH in financial markets.","marker":"Satchell & Knight, 2011"},{"why":"Supplies the mean-variance optimization baseline the DRL portfolios must beat.","marker":"Markowitz, 1991"},{"why":"Contributes the state and environment formulation adopted for the Markov Decision Process.","marker":"Liu et al., 2020"},{"why":"Defines the Proximal Policy Optimization agent trained to allocate portfolio weights.","marker":"Schulman et al., 2017"},{"why":"Motivates the 0.05% transaction cost applied in the trading simulation.","marker":"Orra et al., 2024"},{"why":"Provides the Bayesian hyperparameter tuning used to calibrate the DRL models.","marker":"Bhambu et al., 2024"}],"fun_headline_variants":["GARCH tiers + DRL: Investor-aware portfolios beat baselines","Volatility-tiered DRL beats market benchmarks","Risk-tier preselection lifts DRL portfolio performance","GARCH-guided DRL outperforms benchmarks across risk profiles","GARCH asset sorting lifts DRL portfolio scores"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes the outperformance comes from the DRL agent's allocation, but the comparison does not include an equally weighted portfolio of the same GARCH-selected stocks, so the reported gains may instead come from picking volatile stocks that happened to rise in 2023-2024.","fun_headline_variants_meta":{"raw":{"variants":["GARCH tiers + DRL: Investor-aware portfolios beat baselines","Volatility-tiered DRL beats market benchmarks","Risk-tier preselection lifts DRL portfolio performance","GARCH-guided DRL outperforms benchmarks across risk profiles","GARCH asset sorting lifts DRL portfolio scores"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001264,"raw_usage":{"total_tokens":5135,"prompt_tokens":862,"completion_tokens":4273,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":4196}},"tokens_in":478,"tokens_out":4273,"duration_ms":29242,"temperature":1.0,"reasoning_tokens":4196,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:45:12.233645+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the 2023-2024 test on the same ten-stock aggressive, moderate, and conservative sets with a simple equal-weight or buy-and-hold strategy and no rebalancing. If that equal-weight tier matches or beats the reported PPO returns and Sharpe ratios, the central claim of DRL-driven outperformance is not supported; the tier-selection effect alone explains the result.","supporting_citations":[{"cited_title":"Forecasting volatility in the financial markets","cited_arxiv_id":null,"evidence_quote":"Provides the volatility-forecasting reference for applying GARCH in financial markets."},{"cited_title":"Foundations of portfolio theory","cited_arxiv_id":null,"evidence_quote":"Supplies the mean-variance optimization baseline the DRL portfolios must beat."},{"cited_title":"Portfolio formation with preselection using deep learning from long-term financial data","cited_arxiv_id":null,"evidence_quote":"Contributes the state and environment formulation adopted for the Markov Decision Process."},{"cited_title":"Dynamic reinforced ensemble using bayesian optimization for stock trading","cited_arxiv_id":null,"evidence_quote":"Motivates the 0.05% transaction cost applied in the trading simulation."}],"review_version":1}