{"id":"ba27c8d8-d9f7-43f4-bd2e-74820abefa21","arxiv_id":"2506.19894","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SHAP and gradient explanations of five day-ahead electricity price forecasting DNNs reveal that the most recent price dominates forecasts, and new SSHAP aggregations help visualize these patterns.","lead":"This paper applies explainable AI methods to deep neural networks that forecast electricity prices in five markets, and introduces grouped SHAP values and SSHAP lines to visualize what the models learned. The value is in making black-box price forecasters interpretable for market operators and in revealing cross-market patterns such as reliance on the most recent price.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper claims conditional SHAP but the cited [41] Monte-Carlo estimator is interventional/marginal; with correlated day-ahead price inputs, this can manufacture the reported recency and column effects, so the central market claims are not yet established.","rationale":"The strongest claim (Section 6) is a pattern-level statement about the DNNs' learned dynamics: the last available price dominates and recent variables are preferred over curve matching. Every piece of evidence for this is a SHAP heatmap or SSHAP aggregation. If the SHAP values are computed with an interventional rather than conditional expectation (or with an imprecise approximation), the attribution of credit among correlated inputs is not trustworthy; the most recent hour can act as a surrogate for the whole recent price curve, generating exactly the observed column effect. This is not an 'outside consensus' disagreement but an internal mismatch: Section 4.1 cites [41] as a conditional method, while [41] is a marginal-sampling method, and the SHAP library does not provide general conditional Monte Carlo for neural networks. The reader's weak assumption about SHAP accuracy is on the right track, but it stops at 'unspecified samples and background'; the deeper issue is the estimator definition itself. The training-set-only analysis is a valid secondary concern, but it would not settle the estimator issue; even on a test set, interventional SHAP can still produce spurious recency patterns. Thus the proposed check targets the root cause. If the conditional re-run preserves the patterns, the central claim is substantially strengthened; if not, the paper's main conclusions are unsupported and would require either a correction of the SHAP methodology or a reinterpretation of the findings. Because this is fixable with additional computation, a conditional-accept verdict is appropriate rather than outright reject.","tokens_in":17936,"tokens_out":8466,"duration_ms":99232,"concrete_test":"For the French market, recompute SHAP values with a genuine conditional estimator (e.g., sample complemented features from a fitted Gaussian copula or use the conditional permutation algorithm of Aas et al., 2021) on a documented 1000-instance background, using at least 1000 Monte-Carlo samples; also compute the same analysis with the SHAP library's `PermutationExplainer` in interventional mode. Compare the Price D-1 heatmap and the SSHAP feature rankings to Figs. 9(a) and 10. If the recency/column patterns vanish or reverse under conditional SHAP, the Section 6 market conclusions are artifacts of the estimator, not model behaviour.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 4.1 states that conditional Shapley values are estimated 'using the Monte-Carlo sampling method developed by [41]' from the SHAP library. However, [41] (Štrumbelj–Kononenko) estimates Shapley values by sampling coalition complements from the empirical marginal distribution—the interventional variant—not from the conditional distribution P(X_{\\bar S}|X_S). The SHAP library's corresponding `PermutationExplainer` (the general model-agnostic Monte Carlo implementation) is interventional; no general conditional-sampling Monte Carlo exists for arbitrary DNNs. Because the EPF inputs are strongly correlated (Price D-1 H23 is highly correlated with H22,...,H0; load/generation forecasts are cross-correlated), interventional SHAP can assign credit for a correlated block to whichever single hour is most correlated with the output—here the most recent hour. This would directly produce the 'column effect' (Price D-1 H23 influencing all forecast hours) and the Section 6 headline that the last available price dominates, even if true conditional SHAP would distribute credit across the curve. The number of Monte-Carlo samples and the background set are unspecified, so there is no convergence evidence. Since all market-behavior conclusions are drawn from these SHAP heatmaps and SSHAP lines, the central claim currently rests on an estimator-definition mismatch.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains the two-hidden-layer DNN day-ahead forecasting models from the epftoolbox benchmark for five electricity markets (Germany, France, Belgium, Nord Pool, ComEd) and explains the fitted models with SHAP values, gradients, and heatmaps. It introduces SSHAP values as sums of SHAP values over 24-hour super-variables and SSHAP lines as Gaussian-smoothed conditional averages of SSHAP values against actual price. On this basis, Sections 5 and 6 claim that the last available price is the most important input in all five markets, that the models prefer recency over a naive hour-matching 'curve' strategy, and that early-morning load forecasts push down later prices in France, Belgium, and ComEd.","tokens_in":18172,"tokens_out":9935,"duration_ms":102202,"significance":"If the explanations were computed with a valid conditional Shapley estimator on a representative evaluation set, the cross-market comparison would be a useful empirical contribution: it uses open benchmark data and models, covers five markets, and offers a clean visualization layer for high-dimensional tabular DNNs. The additive construction of SSHAP values (Eq. 2) is exact and convenient. However, the central market-behavior conclusions are not yet established because the estimator actually used appears to be the interventional permutation estimator rather than the conditional estimator claimed in Section 4.1. The paper's strengths are its reproducible data and models, the multi-market scope, and the new visualization tools; its main weakness is that the headline findings rest on an estimator-definition mismatch and on explanations computed on training data.","major_comments":[{"comment":"The paper claims to estimate conditional Shapley values but cites [41] and the SHAP library, whose model-agnostic permutation sampler is interventional: coalition complements are drawn from the empirical marginal distribution of the background data, not from the conditional distribution P(X_barS | X_S). For the strongly correlated hourly price and load inputs of these models, interventional SHAP can attribute the joint effect of a correlated block to the single feature most correlated with the output, which would mechanically produce the column effect and the 'last price available is the most important variable' conclusion reported in Sections 5 and 6. The manuscript does not state the SHAP library version, the number of Monte-Carlo samples, or the background dataset used, so there is no convergence or variance evidence. Please either implement a genuine conditional estimator (for example, Gaussian or nearest-neighbor conditional imputation) and show that the heatmaps are stable, or explicitly reframe all market conclusions as statements about interventional SHAP under an independence assumption, accompanied by a sensitivity analysis to the background distribution.","section":"Section 4.1; Section 6, first bullet"},{"comment":"The models are explained on their training sets, and Table 2 reports much better errors than the benchmark test-set results in [35]. A model that overfits four years of historical data can exhibit attribution patterns that reflect training noise rather than market fundamentals, yet Section 6 offers these patterns as insights into how the markets work. The authors should either recompute the explanations on a held-out period (or on the original test set) and demonstrate that the same conclusions hold, or explicitly restrict the claims to properties of the trained models and discuss the training-to-test performance gap as a limitation.","section":"Section 3.2; Table 2"},{"comment":"Equation (3) is presented as a prediction ('we can derive from Eq. (2) that the sum of the SSHAP lines should be almost a straight line with a slope of 1'). This is not an independent empirical test: it is an algebraic consequence of Eq. (2) whenever the SSHAP lines are weighted averages of SSHAP values over the same training instances and the model output is close to the target. The presentation should call Eq. (3) a self-consistency check of the efficiency property, not a prediction that validates the method.","section":"Section 4.3; Eq. (3)"}],"minor_comments":[{"comment":"The sentence 'a complex offsetting interaction, between.' is incomplete; it should end with an explicit reference to the two renewable forecast variables being offset.","section":"Section 5.1.3"},{"comment":"The Gaussian bandwidth of 5 used in the SSHAP line definition and the e0.5/MWh importance threshold in Table 3 are introduced without justification; please report sensitivity analyses for these hyperparameters.","section":"Section 4.3; Table 3"},{"comment":"The 'extreme SHAP values of a few hundred dollars per MWh' are labeled abnormal but are then averaged; reporting medians and robust dispersion measures would clarify whether the average absolute values are driven by a small number of unstable estimates.","section":"Section 5.5.1"},{"comment":"The discussion of [33] notes a conflicting conclusion (prices three days earlier are most important for Nord Pool) but does not attempt to reconcile it with the present finding that Price D-1 dominates; a comparison paragraph would strengthen the paper.","section":"Section 2"},{"comment":"The dataset link in the abstract contains a space in the URL ('4624 805') and should be corrected.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is the estimator mismatch in Section 4.1. If the authors cannot demonstrate that the reported SHAP values come from a valid conditional estimator, or cannot explicitly adopt an interventional framing with appropriate sensitivity analysis, the paper's central empirical claims should be substantially weakened before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful applied paper, but the headline market findings are not yet trustworthy because of an estimator mismatch.\n\nWhat's new: the SSHAP value is just a sum of SHAP values over a partition (efficiency, Eq. 1 to Eq. 2), so it is not a deep conceptual contribution, but it is the right tool for 24-output, 120+ input tabular models, and the SSHAP-line plots are a decent way to see price-conditional dependence. The real contribution is the empirical sweep across five markets with a consistent visual toolkit, extending the earlier Kernel SHAP papers [33,34] to a broader set of markets and to new aggregate visualizations. The paper is honest about its starting point (epftoolbox) and the data/models are open, so the experiments can be reproduced and extended.\n\nWhere it gets soft: the central claims in Sections 5 and 6 rest on SHAP values that the paper says are conditional Shapley values, but the cited Monte Carlo estimator ([41], Strumbelj and Kononenko) is the interventional, marginal one. With day-ahead prices across hours strongly correlated, interventional SHAP can assign a correlated block's credit to a single recent hour, which would manufacture the reported 'last price dominates' and 'column effect' findings. The paper cites [39,40] for the conditional versus marginal distinction but then uses an interventional estimator; that mismatch is load-bearing, not cosmetic. The Monte Carlo sample size and background set are also unspecified, so there is no convergence evidence. Second, the models are explained on the training set, as the paper openly states in Section 3.2; the good in-sample performance (Table 2) means the patterns could reflect overfitting, not market fundamentals. The cross-market consistency is suggestive but not conclusive. Third, the super-variable split points and Gaussian bandwidth are data-dependent and unjustified; that is minor if the results are robust, but the paper does not show robustness.\n\nBottom line: the paper deserves a serious referee and, with revision, could be a solid contribution. The visualization tools and cross-market documentation are worth having. But I would not cite the market-behavior conclusions until the SHAP estimator question is resolved and the training-set limitation is addressed. If you review it, ask for an explicit statement of what the estimator computes and, ideally, a conditional estimate or a sensitivity check with a different explainer.","headline":"A useful applied XAI study with a promising visualization toolkit, but the central market claims rest on an estimator-definition mismatch that should be resolved before the conclusions are cited.","tokens_in":18759,"tokens_out":3455,"would_cite":false,"duration_ms":37055,"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":"Why day-ahead electricity price AI favors recency over curve matching","keywords":["electricity price forecasting","explainable AI","SHAP","SSHAP","deep neural networks","day-ahead market","feature attribution","market dynamics"],"falsifier":"Retrain the same five architectures on price histories whose daily order is shuffled (or on a held-out period) and rerun the SSHAP analysis: if the 'last price available' dominance and the early-morning-load negative rectangle disappear or reverse, those explanations are artifacts of training-set memorization rather than market fundamentals. A second check is to recompute the attributions with many more Monte-Carlo samples or with exact conditional Shapley values on a subset of instances and see whether the reported SSHAP lines shift materially.","tokens_in":17659,"feed_emoji":"⚡","tokens_out":7649,"duration_ms":78225,"temperature":0.7,"pith_summary":"This paper sets out to show that explainable-AI attributions can expose what black-box neural networks have learned about wholesale electricity markets, turning accurate but opaque price forecasters into a source of market insight. Applying SHAP, gradient, and heatmap methods to deep networks for Germany, France, Belgium, Nord Pool, and ComEd, it claims that the five forecasters do not behave like naive curve matchers. The most recent price available is the most important variable in every market, and the models learn market-specific dynamics such as early-morning load driving later prices down in France, Belgium, and ComEd, and renewable generation dominating price moves in Germany. The paper introduces SSHAP values and SSHAP lines, super-variable aggregates of SHAP values, as a way to compress and visualize the high-dimensional 24-input-by-24-output attribution tables. The paper argues these explanations reveal genuine market fundamentals rather than arbitrary model quirks.","feed_headline":"Last price available drives day-ahead forecasts in five markets","feed_subtitle":"Explaining five neural-network markets shows recency beats hourly curve matching, with load and renewable effects.","key_machinery":"The engine of the analysis is Shapley-value attribution and its aggregation. SHAP assigns each of the 120 to 241 input variables a per-output-hour contribution to each of the 24 forecast hours, with the efficiency property that the sum of all SHAP values equals the model prediction minus its mean. The paper defines a super-variable as any partition of the input variables, typically grouping the 24 hourly lags of one underlying series such as Price D-1 or Load Forecast D, and defines the SSHAP value of a super-variable as the sum of the SHAP values of its members; because the partition is exclusive and complete, efficiency survives the aggregation. The SSHAP line is then the Gaussian-kernel-weighted average of SSHAP values plotted against the actual price, so that if the model is well calibrated the summed lines approximate the identity line. Heatmaps of SHAP and gradient values, arranged as input-hour by output-hour tables, make the difference between column effects (recent hours influencing the whole day) and diagonal effects (same-hour matching) visible.","core_discovery":"The central claim is that the trained day-ahead price forecasters, once explained, reveal real and interpretable market behavior. Across all five markets, the most recent observed day-ahead price is the most important input, and the models prefer that recent information over a price-curve-matching strategy that would pair each forecast hour with the same hour from yesterday. The explanations also show distinct market personalities: Germany's prices are driven heavily by renewable generation forecasts, Nord Pool is comparatively simple with the previous day's price dominating, and France, Belgium, and ComEd all contain a learned early-morning load effect in which a high load forecast in the first hours pushes prices down for the rest of the day. In addition, the paper claims the SSHAP value---the sum of SHAP values over a partition of variables into super-variables---and the SSHAP line, a Gaussian-weighted average of those values against actual price, are new explanatory tools that retain Shapley efficiency while reducing the dimensionality of the attribution problem.","pith_inferences":["SSHAP lines could serve as a calibration diagnostic: where the summed lines deviate from the identity line, the model is systematically over- or under-predicting in that price band, pointing to where features or training data are missing.","The recency result suggests adding even fresher information---for example intraday prices or the latest load and renewable updates---could push day-ahead accuracy further, a hypothesis the paper does not test.","The early-morning-load signature in France, Belgium, and ComEd is a simple checkable rule: unusually high overnight load forecasts should imply lower afternoon prices, independently of the neural networks.","The SSHAP construction is generic for any tabular model with grouped inputs, though its usefulness depends on choosing partitions that correspond to meaningful underlying variables."],"forward_implications":["In all five day-ahead markets, the most recent observed price is the strongest single predictor, with its influence concentrated on the early forecast hours and fading later in the day.","The models' preference for recency over same-hour curve matching means that 'last price available' style features deserve more weight in benchmark and production forecasters than the common naive same-hour benchmark suggests.","High early-morning load forecasts in France, Belgium, and ComEd are learned as a downward pressure on prices for the rest of the day, which is a concrete, market-specific mechanism worth testing outside the model.","Germany's model is renewable-dominated while Nord Pool's is price-dominated, so market complexity and generation mix are reflected in the attribution structure and in the proposed complexity measures.","SSHAP super-variable aggregation lets practitioners summarize 120 to 241 input by 24 output attributions into a small number of interpretable curves without losing the additive efficiency identity."],"supporting_citations":[{"why":"Supplies the five open-access day-ahead market datasets and the DNN architectures and hyperparameters that the paper explains.","marker":"[35]"},{"why":"Defines SHAP values and the efficiency property that the SSHAP aggregation relies on.","marker":"[24]"},{"why":"Provides the Monte-Carlo conditional expectation algorithm used to estimate Shapley values for the networks.","marker":"[41]"},{"why":"Justifies using conditional rather than marginal expectation when computing Shapley values for model explanations.","marker":"[39]"},{"why":"Surveys Shapley-value attribution estimators and supports the conditional-expectation choice.","marker":"[40]"},{"why":"Closest prior XAI application to one of the same benchmark markets (Nord Pool), used as a comparison for the paper's different conclusions.","marker":"[33]"},{"why":"Closest prior XAI application to the German market using the same benchmark platform, used to contrast the multi-market approach.","marker":"[34]"},{"why":"Provides background on generation mixes (nuclear flexibility, hydro, renewables) used to interpret the market-specific explanations.","marker":"[36]"},{"why":"Identifies load, solar, and wind as important features for the German market, which the paper's explanations confirm.","marker":"[31]"}],"fun_headline_variants":["XAI reveals last price dominates day-ahead forecasting","Neural price models lean on latest observed price","SHAP shows recency beats curve matching in power markets","SSHAP: new tool simplifies neural electricity forecasts","Explainable AI: renewables drive German price dynamics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the approximate SHAP values computed on the training sets faithfully reflect what the networks actually learned, so that the patterns they reveal are genuine market dynamics rather than estimation error or overfitting noise.","fun_headline_variants_meta":{"raw":{"variants":["XAI reveals last price dominates day-ahead forecasting","Neural price models lean on latest observed price","SHAP shows recency beats curve matching in power markets","SSHAP: new tool simplifies neural electricity forecasts","Explainable AI: renewables drive German price dynamics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000159,"raw_usage":{"total_tokens":1198,"prompt_tokens":885,"completion_tokens":313,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":238}},"tokens_in":501,"tokens_out":313,"duration_ms":4256,"temperature":1.0,"reasoning_tokens":238,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:06:20.366558+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same five architectures on price histories whose daily order is shuffled (or on a held-out period) and rerun the SSHAP analysis: if the 'last price available' dominance and the early-morning-load negative rectangle disappear or reverse, those explanations are artifacts of training-set memorization rather than market fundamentals. A second check is to recompute the attributions with many more Monte-Carlo samples or with exact conditional Shapley values on a subset of instances and see whether the reported SSHAP lines shift materially.","supporting_citations":[{"cited_title":"Algorithms to estimate Shapley value feature attributions","cited_arxiv_id":null,"evidence_quote":"Surveys Shapley-value attribution estimators and supports the conditional-expectation choice."},{"cited_title":"Electricity price forecasting on the day-ahead market using machine learning","cited_arxiv_id":null,"evidence_quote":"Closest prior XAI application to the German market using the same benchmark platform, used to contrast the multi-market approach."},{"cited_title":"The role of natural gas in setting elec- tricity prices in Europe","cited_arxiv_id":null,"evidence_quote":"Provides background on generation mixes (nuclear flexibility, hydro, renewables) used to interpret the market-specific explanations."}],"review_version":1}