{"id":"ffb4f95e-08ad-4b2d-a304-f5783f786e2a","arxiv_id":"2504.16635","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A Double Deep Q-Network that classifies days as low- or high-risk is used to scale GARCH-based Value-at-Risk, reducing violations and capital requirements on daily Euro Stoxx 50 data from 2008 to 2025.","lead":"This paper combines a GARCH volatility model with a reinforcement-learning classifier that predicts low- and high-risk days, then scales the Value-at-Risk number up or down accordingly. On 16 years of Euro Stoxx 50 data, the hybrid reports fewer VaR violations and lower capital charges while passing standard backtests in most configurations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Preprocessing-leakage ambiguity is the load-bearing weakness: Sections 3.1 and D.1 never confine Boruta, Min-Max, and ADASYN to the training split, and if any fit sees the test split, the 79.4% accuracy and VaR backtests are inflated.","rationale":"The reader's weakest assumption is identical to the concern I find most load-bearing, so I agree with the reader. I considered the alternative that the validation-set tuning of b1 and b2 is the bigger issue; that is real but less decisive because the test split remains independent. The preprocessing ambiguity is more severe because it threatens the independence of the test split itself, and the paper's own wording in D.1, 'The features selected using the Boruta algorithm were split chronologically', at least permits a full-sample selection. The backtest-table inconsistencies and the coherence proof are secondary; they would affect reporting and theory, not the empirical core. The proposed check is a standard leak-free rerun and would settle the question without requiring new data.","tokens_in":22929,"tokens_out":8650,"duration_ms":87790,"concrete_test":"Re-run the pipeline in strictly chronological order: (1) split the raw daily series into train/validation/test before any preprocessing; (2) fit Boruta feature selection, Min-Max normalization, and ADASYN on the training split only; (3) apply the fitted Boruta selector and scaler to validation/test and generate synthetic minority observations only for training; (4) recompute Table 5(b) and Tables 7(b)/8(b) on the untouched test split. If the 79.4% test accuracy, the RL recall of 0.54, or the RL violation counts move materially toward the GARCH baselines, the central claim is unsupported. Additionally, verify that the threshold c in Eq. (9) is computed with a trailing window that never includes the day being classified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim depends on the test split being genuinely untouched by every fitted preprocessing step. Sections 3.1 and D.1 do not establish this. Section 3.1 runs Boruta on the explanatory set and displays the feature ranking before describing the chronological split; Appendix D.1 says 'The features selected using the Boruta algorithm were split chronologically into training, validation and test sets to avoid anticipation bias', which reads as selection on the full sample, and separately states that all predictors were Min-Max rescaled to [0,1] and that ADASYN addressed class imbalance, without specifying that these fits are confined to the training split. If any of these steps is fit using validation or test observations, the 79.4% test accuracy and the RL-specific reduction in test violations (64 to 52 at 5% for GARCH, 64 to 55 for GJR) are no longer out-of-sample, and the claimed outperformance over GARCH/GJR-GARCH becomes an artifact of information leakage. This is the most load-bearing assumption because every headline result, including classification accuracy, backtest pass rates, violation counts, and capital-requirement comparisons, flows through the alleged independence of the test split.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid Value-at-Risk (VaR) framework that combines GARCH/GJR-GARCH volatility forecasts with a Double Deep Q-Network (DDQN) reinforcement learning classifier. The classifier assigns each day to a low-risk or high-risk state based on a threshold derived from past VaR violations, and the GARCH-based VaR is then scaled by factors (1-b1) or (1+b2) depending on the predicted state. Empirical validation on daily Euro Stoxx 50 data from September 2008 to March 2025 reports 79.4% test classification accuracy, reduced numbers of VaR violations relative to GARCH/GJR-GARCH and TCN baselines, passing Kupiec and Christoffersen backtests, consistency with Extreme Value Theory tail fits, and lower capital requirements while maintaining regulatory coverage. The paper argues that the adaptive, RL-based adjustment yields both statistically valid and capital-efficient risk measures.","tokens_in":23223,"tokens_out":5214,"duration_ms":51240,"significance":"If the empirical claims are correct, the paper offers a practically relevant way to combine econometric volatility modeling with deep reinforcement learning for regulatory risk measurement, with clear economic benefits through lower capital charges without sacrificing coverage. The paper has notable strengths: a long (over 16-year) test dataset spanning crisis episodes, comparison against several baselines (LR, SVM, ANN, MLP, TCN), use of class-imbalance metrics beyond accuracy (G-Mean, recall), a detailed appendix with hyperparameters, and explicit backtesting and EVT robustness checks. However, the current manuscript does not establish that the entire empirical pipeline is leak-free, and several central claims are either circular or contradicted by the paper's own tables. The significance is therefore conditional: the contribution would be substantial after the methodological gaps are resolved, but as written the key evidence is not reliable.","major_comments":[{"comment":"The manuscript does not establish that the Boruta feature selection, Min-Max normalization, and ADASYN oversampling are confined to the training split. Appendix D.1 states that 'The features selected using the Boruta algorithm were split chronologically into training, validation and test sets to avoid anticipation bias,' which reads as feature selection on the full sample followed by splitting; the same appendix then states that 'all predictors were rescaled to [0,1] using Min–Max normalization' and that ADASYN addressed class imbalance, without specifying that the normalization parameters and the ADASYN fit are training-only. Section 3.1 similarly describes the chronological split but then discusses Boruta selection without an explicit statement that the selection is performed only on the training portion. Because the 79.4% test accuracy and the test-sample violation reductions (e.g., from 64 to 52 in Table 9) are the principal evidence for outperformance, this ambiguity is load-bearing. The authors must state precisely which preprocessing steps are fit on which split, and if the pipeline as run used any full-sample information, the experiments must be rerun with strictly training-only preprocessing.","section":"Appendix D.1 and Section 3.1"},{"comment":"The validation-sample backtest results are not independent evidence of model quality because the adjustment parameters b1 and b2 are deliberately calibrated to minimize the number of VaR violations on the validation set (Section 4.1: 'the pair (b1 = 0.30, b2 = 0.20) produces the lowest number of exceedances'). Presenting the validation reduction in breaches (Tables 6-9) as evidence of outperformance is circular. At the same time, the text in Section 3.3 claims that 'RL-adjusted VaR models pass through validation and test samples' and 'show no systematic bias,' which is directly contradicted by Table 8(a): the RL GARCH 5% validation Kupiec test reports p = 0.032 with decision 'RejectH0'. Please correct the text, and either remove the validation-based claims or explicitly re-frame them as in-sample calibration diagnostics rather than out-of-sample evidence.","section":"Section 4.1 and Tables 7/8"},{"comment":"The classification threshold c is defined as the maximum of returns that breach VaR over a horizon H, but the manuscript never states whether H is restricted to the training period or whether c is recomputed dynamically in a way that avoids lookahead. If the threshold is computed using validation- or test-period returns (or using VaR estimates that incorporate those returns), then the target labels y_t(c) in the test sample are constructed with information from the test period itself, which would inflate the reported 79.4% accuracy and the out-of-sample backtest pass rates. The authors must specify the exact construction window for c, demonstrate that the labels in the validation and test splits are generated without using those splits' returns, and discuss the endogeneity that arises because the labels are defined through violations of the very GARCH VaR model that is later adjusted.","section":"Section 2.3, Eq. (9)"},{"comment":"The proof of Proposition 2.1 does not establish that Classification-Adjusted VaR preserves the fundamental properties of Value-at-Risk. In particular, translation invariance fails: for a portfolio with a deterministic cash position c, VaR_ML(X + c) = κ·(VaR(X) + c) = κ·VaR(X) + κ·c, which is not equal to VaR_ML(X) + c unless κ = 1. The claim that 'subadditivity and translation invariance are unaffected' is asserted without proof and is not a consequence of positive homogeneity. Either the proposition and proof must be corrected, or the authors should state the weaker and correct result that VaR_ML is a positive-homogeneous, monotonic transformation of the baseline VaR.","section":"Section 2.4, Proposition 2.1"}],"minor_comments":[{"comment":"The phrase 'captures non-nonlinearities' in the conclusion is a typo and should read 'nonlinearities.'","section":"Section 5"},{"comment":"The claim that the RL model's skewness and kurtosis are 'far closer to empirical distributions' would be more convincing if the empirical skewness and kurtosis of the test-period returns were reported next to the model values; please add these statistics.","section":"Section 3.3"},{"comment":"There is a formatting issue in the features table: 'sig gjr-GARCH0.013' is missing a space between the variable name and its value; please fix the alignment.","section":"Table 2"},{"comment":"The sentence beginning 'For time series models, ARIMA orders capture the optimal trade-off...' has a subject-verb agreement issue and is difficult to parse; please rewrite for clarity.","section":"Appendix D.1.1"},{"comment":"The paper reports percentages of low- and high-risk observations in each split but does not give the absolute numbers of observations in the training, validation, and test sets; please add these to Section 3.1 or Table 1.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The core idea is timely, but the manuscript is not yet publishable because the most load-bearing evidence depends on an unverified preprocessing protocol. The internal inconsistency between the text's claim that all RL models pass the Kupiec test and Table 8(a)'s rejection for RL-GARCH at 5% validation is the kind of error that should have been caught before submission; together with the circular validation calibration, it makes me doubt the reliability of the reported results until the pipeline is clarified and rerun if necessary. I would not recommend rejection if the authors can provide, or re-derive, the results with strictly training-only preprocessing and correct the reporting; otherwise the central claim collapses."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a genuine engineering attempt: it scales GARCH/GJR-GARCH VaR by a DDQN-predicted risk state, defines VaR_{ML} in Eq. (11), and shows on Euro Stoxx 50 data that violation counts drop and backtests mostly pass out-of-sample. That combination is new, and the authors deserve credit for a clearly described chronological split, detailed hyperparameters, training times, and an EVT tail check.\n\nBut the central claim is softer than the conclusion states. Three things bother me.\n\nFirst, the preprocessing leakage ambiguity is real. In Section 3.1, Boruta feature selection is presented after the split paragraph, but Appendix D.1 says 'the features selected using the Boruta algorithm were split chronologically' and separately states all predictors were Min-Max rescaled and ADASYN applied, with no explicit statement that these fits are confined to the training split. If any of those steps sees validation or test data, the 79.4% test accuracy and the reduced violations are inflated. The authors may well have done it correctly, but they need to say so.\n\nSecond, the backtest reporting is inconsistent. Table 8(a) shows the RL GARCH 5% Kupiec test on validation is rejected (p=0.032), yet Section 3.3 says RL models show no systematic bias and the conclusion claims all pass Kupiec. The test sample looks fine, but the overstatement should be fixed.\n\nThird, the calibration of b1 and b2 is fit to minimize validation violations (Section 4.1). So the validation 'reduction in breaches' is by construction. The test-sample results are still informative, but the narrative should not present the validation improvement as independent evidence.\n\nThere is also a flawed proof in Proposition 2.1: since the multiplier depends on the predicted class, positive homogeneity of VaR does not imply the adjusted measure is subadditive or translation invariant for portfolios. That claim should be dropped or properly conditioned. And the threshold definition in Eq. (9) is under-specified; a careful reader cannot tell whether it uses only past violations.\n\nOverall, this is a plausible applied paper, not a breakthrough. The test-set evidence is genuinely out-of-sample if the pipeline is clean, and the misreporting is fixable. I would send it to peer review, with the request that the authors clarify the preprocessing boundary and correct the backtest claims. It is a decent reading-group discussion for the leakage question, but I would not build on it until those issues are resolved.","headline":"A plausible hybrid GARCH-DDQN VaR model whose empirical gains depend on leak-free preprocessing; the paper needs revision to prove that and to fix overclaimed backtests.","tokens_in":23783,"tokens_out":3964,"would_cite":false,"duration_ms":35389,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Scaling GARCH-based Value-at-Risk by a reinforcement-learning risk classifier reduces breaches and capital charges while passing regulatory backtests on sixteen years of Euro Stoxx 50 data.","keywords":["Value-at-Risk","deep reinforcement learning","DDQN","GARCH","GJR-GARCH","directional prediction","imbalanced classification","Euro Stoxx 50"],"falsifier":"Re-run the pipeline twice, once with Boruta, Min-Max, and ADASYN fitted only on the training window and once with them fitted on the full sample, and compare the test-split accuracy and the Kupiec/Christoffersen statistics; if the second run beats the first, the reported results are look-ahead artifacts. As a control, replace the DDQN labels in the adjustment rule with random labels and check whether the violation counts and backtest decisions stay unchanged, which would show the classifier itself, rather than the $b_1,b_2$ scaling, is doing the work.","tokens_in":22745,"feed_emoji":"📉","tokens_out":15537,"duration_ms":122344,"temperature":0.7,"pith_summary":"The paper sets out to show that Value-at-Risk can be made both safer and cheaper by letting a reinforcement-learning agent decide when to trust the GARCH volatility model. Standard GARCH-type VaR tends either to overstate risk in calm periods, tying up regulatory capital, or to miss the rare extreme losses that matter most. The authors reformulate return prediction as an imbalanced classification of low- versus high-risk regimes, use a Double Deep Q-Network (DDQN) to label each trading day, and then scale the GARCH-based VaR down by 30% in predicted calm states and up by 20% in predicted stressed states. On daily Euro Stoxx 50 data from 2008 to 2025, they report that this adjusted VaR produces fewer and less clustered violations than GARCH or GJR-GARCH alone, passes the Kupiec and Christoffersen backtests that reject the plain models, and carries lower implied capital requirements. If the claim holds, the framework supplies a workable bridge between econometric volatility modeling and adaptive machine learning for regulatory risk management.","feed_headline":"Hybrid AI-econometrics VaR passes backtests and cuts breaches","feed_subtitle":"A DDQN classifier scales GARCH VaR up in stress and down in calm: fewer violations, lower capital, same coverage.","key_machinery":"Three objects carry the argument. First, the risk threshold $c = \\max\\{r_{k+1} : r_{k+1} < \\mathrm{VaR}_{k+1}(\\alpha),\\, k=1,\\dots,H\\}$, the mildest recent VaR violation over a horizon $H$, which converts returns into the binary label $y_t(c) = \\mathbf{1}\\{r_t \\le c\\}$ and ties the classification target directly to realized downside risk. Second, the Double Deep Q-Network classifier trained with an unbalanced reward scheme ($+1$ for true positives, $-1$ for false positives, $+\\rho$ for true negatives, $-\\rho$ for false negatives, with $\\rho$ the minority-to-majority class ratio) so the agent learns to detect rare high-risk days rather than optimizing majority-class accuracy. Third, the adjustment rule $\\mathrm{VaR}^{ML}(\\alpha) = \\kappa\\cdot\\mathrm{VaR}_{t+1}(\\alpha)$ with $\\kappa = 1-b_1$ for low-risk predictions and $\\kappa = 1+b_2$ for high-risk predictions, with $b_1 = 0.30$, $b_2 = 0.20$ fixed by grid search and corroborated by Bayesian posterior estimation. Because the adjustment is a positive scalar multiple of a coherent risk measure, the paper argues that $\\mathrm{VaR}^{ML}$ inherits VaR's homogeneity, monotonicity, and translation properties.","core_discovery":"The central claim is that the Classification-Adjusted Value-at-Risk, defined as $\\mathrm{VaR}^{ML}(\\alpha) = (1-b_1)\\cdot\\mathrm{VaR}_{t+1}(\\alpha)$ in predicted low-risk states and $(1+b_2)\\cdot\\mathrm{VaR}_{t+1}(\\alpha)$ in predicted high-risk states, strictly improves on the underlying GARCH or GJR-GARCH measure: it achieves the correct violation frequency, restores the independence of violations that the plain econometric models violate, and does so with less conservative capital buffers. The supporting record is the backtest comparison: on the validation window the RL-adjusted GARCH VaR produces 52 violations at the 5% level against 73 for the original model and passes both the Kupiec and Christoffersen tests that reject the original, while on the test window violations stay close to their expected counts (52 actual versus 50 expected at 5%) and a Wilcoxon rank-sum test against GARCH yields $p = 0.0039$. The authors also report that exceedances over the adjusted threshold remain consistent with a Generalized Pareto Distribution, which they read as evidence that tail behavior is preserved, and that the adjusted VaR lies significantly above the GARCH VaR under a Mann-Whitney test, which they interpret as lower capital charges at unchanged regulatory coverage.","pith_inferences":["The authors validate on a single index, Euro Stoxx 50; a natural test of transferability is to re-run the identical DDQN-plus-GARCH pipeline on US equities, FX, or commodities and check whether the 79.4% test accuracy and the violation reductions survive.","Their finding of a negative posterior correlation between $b_1$ and $b_2$ suggests the benefit comes from the balance of the two adjustments, not from either alone; a one-parameter variant would presumably sacrifice either coverage or capital efficiency, which could be tested directly.","The threshold $c$ is defined from realized VaR violations, which creates a feedback loop between the classifier's target and the backtest statistic; one could profit from this loop deliberately by training the agent with a reward proportional to breach size or to capital at risk, optimizing the economic objective rather than label accuracy.","One can separate the contributions of the classifier and the adjustment by setting $b_1 = b_2 = 0$ (no adjustment) and checking whether violation counts already improve; the paper does not report this ablation, but it would establish how much of the gain comes from the scaling rule versus the regime labels themselves."],"forward_implications":["If the framework is correct, a risk manager can report a VaR that is less conservative in calm periods and more conservative in stressed ones without failing the standard backtests, since the adjusted model is accepted by both the Kupiec and Christoffersen tests at the 5% and 1% levels.","The reported Mann-Whitney results imply the adjusted VaR sits systematically above the GARCH VaR distribution, so capital charges fall at unchanged coverage; this is the paper's economic payoff.","Because the paper argues that better VaR estimates directly improve the Expected Shortfall tail estimate, the same classification step carries over to the Basel III framework that replaced VaR with ES.","Because inference takes under 0.1 seconds per step while training is a one-off cost, the adjusted VaR can be recomputed daily in an operational risk-monitoring system."],"supporting_citations":[{"why":"Supplies the GARCH(1,1) model whose VaR forecasts form the baseline that the adjustment scales.","marker":"Bollerslev (1986)"},{"why":"Supplies the GJR-GARCH asymmetric variant used to model the leverage effect and as the second baseline.","marker":"Glosten et al. (1993)"},{"why":"Supplies the Double Deep Q-Network algorithm, the adaptive classifier that predicts the risk regimes.","marker":"Van Hasselt et al. (2016)"},{"why":"Motivates the refined multinomial threshold formulation for directional return prediction on which the binary risk labels are built.","marker":"Nevasalmi (2020)"},{"why":"Supplies ADASYN, the oversampling method applied to all supervised baselines to handle class imbalance.","marker":"He et al. (2008)"},{"why":"Supplies the Boruta feature selection procedure that fixes the explanatory input set.","marker":"Kursa and Rudnicki (2010)"},{"why":"Provides the unbalanced reward design that grades true positives and false negatives asymmetrically.","marker":"Firdous et al. (2023)"},{"why":"Grounds the Markov decision process formulation and value-function machinery the DDQN inherits.","marker":"Sutton and Barto (1998)"}],"fun_headline_variants":["RL tunes GARCH VaR: fewer breaches, lower capital","AI-adjusted VaR passes backtests, cuts capital","Deep RL refines GARCH for sharper VaR","RL sharpens VaR: backtest pass, capital down"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that all preprocessing, namely Boruta feature selection, Min-Max normalization, and ADASYN oversampling, is fitted only inside the training split so that the validation and test samples are never seen by any preprocessing step; the paper never states this, yet the reported 79.4% test accuracy and the out-of-sample backtest pass rates depend on it.","fun_headline_variants_meta":{"raw":{"variants":["RL tunes GARCH VaR: fewer breaches, lower capital","AI-adjusted VaR passes backtests, cuts capital","Deep RL refines GARCH for sharper VaR","RL sharpens VaR: backtest pass, capital down"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000276,"raw_usage":{"total_tokens":1675,"prompt_tokens":1005,"completion_tokens":670,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":603}},"tokens_in":621,"tokens_out":670,"duration_ms":6195,"temperature":1.0,"reasoning_tokens":603,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:58:43.498088+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the pipeline twice, once with Boruta, Min-Max, and ADASYN fitted only on the training window and once with them fitted on the full sample, and compare the test-split accuracy and the Kupiec/Christoffersen statistics; if the second run beats the first, the reported results are look-ahead artifacts. As a control, replace the DDQN labels in the adjustment rule with random labels and check whether the violation counts and backtest decisions stay unchanged, which would show the classifier itself, rather than the $b_1,b_2$ scaling, is doing the work.","supporting_citations":[],"review_version":1}