{"id":"2ab76a13-e024-4dd3-8afb-11b3c154ee6b","arxiv_id":"1908.07978","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A dilated causal CNN trained with quantile regression on multiple stocks forecasts one-day-ahead Value at Risk with exceedance rates close to nominal levels.","lead":"The paper uses convolutional neural networks with a quantile loss to forecast Value at Risk, and shows that training one network on many stocks at once calibrates forecasts better than training on a single stock. A generalist would read it to see whether multi-asset deep learning can solve the small-data problem of extreme-event forecasting.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The joint QCNN's calibration gain is not distinguished from the small-sample failure of the single-stock QCNN; the claimed cross-asset tail pooling rests on an untested homogeneity assumption.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing condition: the joint model's benefit depends on sufficiently similar tail dynamics across assets. My concern sharpens this by noting that the single-stock QCNN failure is a confound: the joint model's improvement may simply reflect more training data, not cross-asset pooling of tail information. This does not overturn the conditional verdict; it strengthens the need for a control experiment and for released code and data before the central claim is established. The paper honestly reports the single-stock QCNN's failure, which is evidence in its favor, but it makes the homogeneity assumption more salient rather than less.","tokens_in":4866,"tokens_out":8869,"duration_ms":97107,"concrete_test":"Run the same QCNN architecture and training protocol from Section 3.3 on 100 synthetic daily return series with known, deliberately heterogeneous tails, e.g., GARCH(1,1) processes with t-distributed innovations where the scale and degrees of freedom vary substantially across assets. Compare per-asset exceedance rates and Dynamic Quantile test rejections of the joint QCNN against constant-quantile and GARCH baselines. If per-asset calibration degrades for high-volatility or heavy-tailed assets, the real-data result is contingent on the particular composition of the stock sample; if per-asset calibration remains near nominal across all assets, the tail-similarity assumption is not the limiting factor.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that a QCNN trained jointly on 100 assets yields better-calibrated one-day-ahead VaR forecasts than the listed baselines (Section 3.5, Tables 1-3). The paper's own single-stock QCNN result, however, shows the same architecture is badly miscalibrated when trained on one asset alone: the mean 95% exceedance rate is 0.1269 versus the nominal 0.05. The joint model's improvement could therefore be explained entirely by the hundred-fold increase in training data, without any transfer of tail information across assets. The paper motivates joint training by asserting that stocks 'have experienced more extreme events, and we may expect them to have similar sources' (Section 3.1), but it provides no evidence that the 100 randomly selected stocks are tail-homogeneous enough for this assumption to hold. If tail dynamics are idiosyncratic, the joint QCNN may be fitting a pooled or average tail rather than an asset-conditional tail, making the reported aggregate exceedance rates and low Dynamic Quantile rejection rates an artifact of the particular stock sample. This assumption is load-bearing: it is the mechanism stated for why joint training should work, and it is untested by any split-sample or synthetic control varying tail heterogeneity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a quantile convolutional neural network (QCNN), a one-dimensional dilated causal CNN trained with the pinball loss, for one-day-ahead Value at Risk forecasting. The model is evaluated on daily log returns of 100 randomly selected US stocks over 2009-2018, with the last 30% as test set. The authors compare a per-stock QCNN and a joint QCNN trained on all stocks against a constant quantile estimator, a normal GARCH(1,1), and a linear quantile autoregression. The reported tables show that the joint QCNN yields exceedance rates closer to the nominal 95%, 99%, and 99.9% levels, lower standard deviations of exceedance rates, and lower Dynamic Quantile test rejection rates than all baselines, while the single-stock QCNN is badly miscalibrated. The paper concludes that training on multiple assets improves VaR forecasts.","tokens_in":5035,"tokens_out":8581,"duration_ms":79942,"significance":"The central empirical claim—that the joint QCNN improves quantile calibration relative to standard baselines—is plausible and, if confirmed, practically useful: VaR forecasts that are both accurately calibrated and less conservative than a historical quantile are valuable for risk budgeting. The authors are to be credited for evaluating out-of-sample with a standard backtesting test, for reporting exceedance rates, and for including simple baselines rather than only neural alternatives. However, the statistical support is incomplete: no significance tests, no confidence intervals, and no control experiment separate the effect of larger sample size from the claimed cross-asset tail pooling. The paper is clearly written and the QCNN idea is presented transparently, but the empirical evidence as reported is not yet at the standard needed to support the strong conclusions in Section 3.5.","major_comments":[{"comment":"The comparisons are based solely on point estimates (means, medians, standard deviations) with no significance tests or confidence intervals. Because the same 100 stocks are used for every method, paired tests are available; for example, a paired sign test or bootstrap on the per-stock exceedance-rate difference between Joint QCNN and GARCH would directly assess whether the reported gap (0.0433 vs. 0.0347 at the 95% level) is systematic. Likewise, the DQ rejection rates are proportions out of 100 stocks, so binomial confidence intervals should accompany them. Without such measures, the claim in the text that the joint QCNN 'produces the highest quality VaR estimates' is not statistically established.","section":"Section 3.5, Tables 1–3"},{"comment":"The rationale for joint training is that extreme returns across different stocks 'may ... have similar sources,' but this assumption is never tested, and the reported comparison of Joint QCNN with single-stock QCNN conflates cross-asset pooling with a roughly 100-fold increase in training data. The authors should add a control that holds sample size fixed while destroying the cross-asset structure, such as training the same architecture on a random permutation of the returns across the 100 stocks, or on the pooled returns with the asset index removed. If the permuted/pooled model matches Joint QCNN's performance, the improvement is attributable to sample size rather than to shared tail dynamics; otherwise the stated 'similar sources' mechanism receives support. This distinction is load-bearing for the paper's contribution, which is explicitly about learning from multiple assets.","section":"Section 3.1 and Section 3.5"},{"comment":"Assigning a p-value of 0 to zero-exceedance forecasts is not part of the Engle-Manganelli Dynamic Quantile test as defined by Eq. (6). When Hit has no variation, the regression in Eq. (5) is degenerate and the chi-square approximation is invalid, so a pre-specified replacement rule can bias the reported rejection rates. The authors should report the frequency of zero-exceedance forecasts for each method, justify the 0-p-value rule, and show that the DQ rejection-rate comparisons are robust to an alternative treatment (e.g., excluding such stocks or using an exact test).","section":"Section 3.4, Eq. (6) and Tables 1–2"},{"comment":"The joint QCNN training and evaluation protocol is underspecified. It is not stated whether the standardization ('subtracting the mean and dividing by the standard deviation') is computed per asset or globally, how training minibatches are composed from the 100 stocks, or whether the single trained network is applied to each stock's test sequence without any asset-specific conditioning. These details determine whether the network can in principle learn asset-specific tail behavior, and they are necessary for reproducing Tables 1–3. Please spell out the protocol and, if relevant, provide code.","section":"Section 3.3"}],"minor_comments":[{"comment":"The phrase 'slided local receptive fields' should be 'sliding local receptive fields'; there are also typographical spacing errors such as 'V aR' in the abstract.","section":"Section 1"},{"comment":"The quantile loss is written with summation subscripts that are hard to parse; use the standard pinball loss ρ_θ(u)=u(θ−1{u<0}) and sum over the batch.","section":"Section 2.2, Eq. (1)"},{"comment":"The statement 'The output sequences are the inputs shifted by 1 step' should clarify how the target value for the one-day-ahead VaR forecast is extracted from the output sequence.","section":"Section 3.3"},{"comment":"The robustness statement that 'the results were quite similar' for 1999–2008 is not accompanied by any numbers; include the corresponding table or remove the claim.","section":"Section 3.5"},{"comment":"State explicitly why DQ test results are not reported for the 99.9% level; presumably the expected number of exceedances is too small for the asymptotic test.","section":"Section 3.4"},{"comment":"No code or data-publication link is provided; for an empirical study of this kind, releasing the implementation would substantially increase confidence in the results.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"This is a straightforward empirical paper with a plausible central claim, but its current evidence is not strong enough for acceptance in a rigorous journal. The lack of significance testing and the untested sample-size-versus-pooling confound are fixable with additional experiments; I do not see a fundamental flaw in the method itself."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on Petneházi's QCNN paper. The genuinely new thing is the joint multi-asset training of a quantile CNN for VaR; the architecture itself is a known WaveNet-style dilated causal CNN with a pinball loss. The paper reports a real empirical result: a QCNN trained on 100 stocks at once gives better-calibrated one-day-ahead VaR exceedance rates and lower Dynamic Quantile rejection rates than constant, GARCH, and linear QR baselines, at lower average VaR. That is practically useful if it holds. Credit is due for honest reporting: the single-stock QCNN fails badly (95% exceedance rate 12.7% vs nominal 5%), and the paper says so rather than hiding it. The repeated experiment on 1999-2008 data is also a nice check.\n\nSoft spots are real but mostly about evidence quality, not fatal logic. There is no code or data release, no significance tests or confidence intervals on the differences between methods, and no sensitivity analysis around the chosen architecture (6 layers, 8 filters, kernel size 2, 128-step window). The paper cites LSTM quantile regression and QARNN in the introduction but does not compare against them—that is a missed baseline that matters for an empirical claim. The stress-test worry about 'similar sources' is somewhat overplayed: even if joint training works simply because 100 stocks give the network a hundred times more tail events, that is not an artifact; it is the mechanism. But the homogeneity assumption is untested, and the paper gives no split-sample or synthetic experiment varying tail heterogeneity, so we do not know whether the result transfers to a different stock universe. That is a legitimate limitation, not a load-bearing flaw.\n\nFor the right reader—an applied quant researcher or risk manager interested in simple deep learning tools for VaR—this is worth reading. It deserves a serious referee: the central claim is clear, falsifiable, and the method is simple enough to reproduce if code appears. I would send it to peer review with a request for code, significance testing, sensitivity analysis, and a direct comparison with a quantile LSTM. My own verdict is conditional, but the condition is standard empirical rigor, not a novel problem.","headline":"Joint multi-asset training helps quantile CNNs for VaR, but the evidence is thinner than the conclusion.","tokens_in":5589,"tokens_out":2137,"would_cite":false,"duration_ms":21717,"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":"A convolutional network trained jointly on many stocks' histories can forecast Value at Risk with exceedance rates close to nominal levels, beating constant, GARCH, and linear quantile regression baselines.","keywords":["Value at Risk","quantile regression","convolutional neural network","dilated causal convolution","time series forecasting","multi-asset learning","Dynamic Quantile test","tail risk forecasting"],"falsifier":"Train the joint QCNN on 100 stocks and evaluate it on a held-out set of stocks from a market or sector that was deliberately absent from training, for example train on US equities and test on European equities. The central claim predicts that exceedance rates stay near nominal and the Dynamic Quantile rejection rate stays low; if instead the rejection rate rises to the level of the single-stock QCNN, the pooling benefit depends on shared tail sources and does not generalize.","tokens_in":4610,"feed_emoji":"📉","tokens_out":7315,"duration_ms":64292,"temperature":0.7,"pith_summary":"This paper proposes a one-dimensional convolutional neural network that forecasts quantiles of daily stock returns directly, rather than forecasting the mean. The network is trained with a quantile (pinball) loss, so its output is an estimate of a Value at Risk level such as the 1% or 5% tail quantile. The central claim is that when the network is trained jointly on the price histories of many stocks, it produces one-day-ahead VaR forecasts with exceedance rates close to the nominal levels and with fewer Dynamic Quantile test rejections than constant, GARCH(1,1), and linear quantile regression baselines. If true, this means that pooling extreme events across assets can make deep quantile forecasting work despite the scarcity of tail data for any single stock, and that the resulting risk budgets are cheaper than those from a historical constant estimate.","feed_headline":"One CNN trained on 100 stocks beats GARCH at Value at Risk","feed_subtitle":"Sharing tail events across assets keeps VaR exceedance rates near target and passes the Dynamic Quantile test.","key_machinery":"Three components carry the argument. (1) The pinball quantile loss replaces mean squared error so that the network's output is a conditional quantile, i.e., a VaR estimate, without distributional assumptions. (2) The network is a WaveNet-style stack of six causal convolutional layers with eight filters of kernel size 2, ReLU activations, and exponentially increasing dilation rates ($2^{l-1}$), followed by a $1\\times 1$ convolution with a linear activation; causal dilated convolutions give an exponential receptive field so the 128-step input window can predict one step ahead. (3) Joint training feeds the overlapping 128-return windows from all 100 stocks into the same network, so the model learns from pooled tail events rather than from a single asset's history. The claim is that these three pieces together produce calibrated one-day-ahead VaR forecasts at 95%, 99%, and 99.9% confidence.","core_discovery":"The discovery is that quantile forecasting with a convolutional neural network works for Value at Risk when the network is trained jointly on multiple assets, and fails when trained on a single asset's history alone. The QCNN is a causal, dilated, one-dimensional convolutional network whose final layer outputs a $\\theta$-quantile of the next day's return, optimized with the asymmetric absolute loss $\\theta\\sum_{y_i \\ge \\hat q_i}(y_i-\\hat q_i) + (\\theta-1)\\sum_{y_i < \\hat q_i}(y_i-\\hat q_i)$. In the empirical study on 100 US-listed stocks, the single-stock QCNN overshoots the targeted exceedance rates by up to an order of magnitude, while the joint QCNN trained on all 100 stocks gives mean exceedance rates of 0.0433, 0.0115, and 0.0023 for the 95%, 99%, and 99.9% levels—close to the nominal 0.05, 0.01, and 0.001—and has lower average DQ-test rejection rates (0.05 and 0.16 at the 5% and 1% significance levels for 95% VaR) than all baselines, with lower average VaR values than the constant historical estimate.","pith_inferences":["A natural next test the paper does not run: train the joint QCNN on stocks from one sector or market and test on a held-out sector or market; if the calibration holds, the transfer is general, and if it degrades, the benefit depends on the training and test stocks sharing tail sources.","The paper's explanation for pooling—'more stocks have experienced more extreme events, and we may expect them to have similar sources'—suggests a data-economy interpretation: the effective sample size for tail estimation is the number of pooled extreme events, not the length of any one series. That interpretation could be tested by varying the number of training assets and observing whether DQ rej","The same pinball-loss architecture should apply to expected shortfall by swapping the loss for a coherent tail-loss function, and to intraday or high-frequency returns, though neither is evaluated here."],"forward_implications":["If the central claim is correct, VaR can be produced nonparametrically: no normality or t-distribution assumption is needed, because the network learns the empirical tail directly from returns.","Joint training is the load-bearing ingredient for deep quantile models: with only one stock's history the same architecture grossly overshoots exceedance rates, so pooling many assets is what makes the data volume sufficient.","Better calibration transfers into capital: the joint QCNN's average VaR values are lower than the constant historical estimate's while maintaining similar exceedance rates, meaning a bank or portfolio using it would set aside less capital for the same coverage.","The low Dynamic Quantile rejection rates indicate that the joint model's exceedances are not only correctly frequent but also correctly timed—past exceedances and VaR levels have little explanatory power for future hits."],"supporting_citations":[{"why":"Supplies the quantile regression objective (pinball loss) that the paper replaces MSE with, turning a CNN's output into a conditional quantile.","marker":"[Koenker and Hallock, 2001]"},{"why":"WaveNet introduced the dilated causal convolutional stack that QCNN adapts for time-series input.","marker":"[Oord et al., 2016]"},{"why":"Showed that a WaveNet-style network can be used for conditional time-series forecasting, the direct architectural antecedent of QCNN.","marker":"[Borovykh et al., 2017]"},{"why":"Provides both the CAViaR quantile-regression approach to VaR and the Dynamic Quantile test used as the main evaluation criterion.","marker":"[Engle and Manganelli, 1999]"},{"why":"Introduced dilated convolutions, which the paper uses to grow the receptive field exponentially with depth.","marker":"[Yu and Koltun, 2015]"},{"why":"Empirically argues convolutional sequence models can match or beat recurrent ones, motivating the choice of a CNN over an LSTM.","marker":"Bai et al. [2018]"},{"why":"Earlier neural-network quantile regression for VaR, establishing that the quantile approach to risk forecasting predates deep architectures.","marker":"Taylor [2000]"}],"fun_headline_variants":["Quantile CNN for VaR works only with multi-asset training","Joint training on 100 stocks makes CNN VaR beat GARCH","Single-asset CNN fails at Value at Risk; multi-asset succeeds","CNN quantile forecasts of VaR need cross-stock learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the daily returns of different stocks have similar enough tail behavior—common sources of extreme events—that one network trained on many assets can learn valid quantiles for any single asset; if tails are idiosyncratic, the joint model's calibration could be an artifact of pooling rather than a real transferable property.","fun_headline_variants_meta":{"raw":{"variants":["Quantile CNN for VaR works only with multi-asset training","Joint training on 100 stocks makes CNN VaR beat GARCH","Single-asset CNN fails at Value at Risk; multi-asset succeeds","CNN quantile forecasts of VaR need cross-stock learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000715,"raw_usage":{"total_tokens":3178,"prompt_tokens":875,"completion_tokens":2303,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":2228}},"tokens_in":491,"tokens_out":2303,"duration_ms":13801,"temperature":1.0,"reasoning_tokens":2228,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:52:01.618215+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the joint QCNN on 100 stocks and evaluate it on a held-out set of stocks from a market or sector that was deliberately absent from training, for example train on US equities and test on European equities. The central claim predicts that exceedance rates stay near nominal and the Dynamic Quantile rejection rate stays low; if instead the rejection rate rises to the level of the single-stock QCNN, the pooling benefit depends on shared tail sources and does not generalize.","supporting_citations":[{"cited_title":"Caviar: conditional value at risk by quantile regression","cited_arxiv_id":null,"evidence_quote":"Provides both the CAViaR quantile-regression approach to VaR and the Dynamic Quantile test used as the main evaluation criterion."},{"cited_title":"A quantile regression neural network approach to estimating the conditional density of multiperiod returns","cited_arxiv_id":null,"evidence_quote":"Earlier neural-network quantile regression for VaR, establishing that the quantile approach to risk forecasting predates deep architectures."}],"review_version":1}