{"id":"4d67c35b-e302-43f5-9a47-cb57cfd8f955","arxiv_id":"2412.01224","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Conv-KANs achieve the lowest MSE, RMSE, and MAE for CSI 300 option prices in this study, but the comparison is undermined by a possibly circular input feature and overlapping train/test splits.","lead":"This paper applies Kolmogorov-Arnold Networks and convolutional KANs to price CSI 300 index options, reporting that Conv-KANs give the lowest mean squared error among tested models. A sharp reader should care because the experimental setup appears to feed the models the same theoretical price formulas used as the comparison baselines, which would undermine the claimed advantage.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'theoretical price' input feature is undefined; if it is the B-S/B-S-M formula output, the central comparison is circular and the reported Conv-KAN advantage is not evidence of independent superiority.","rationale":"The reader's weakest_assumption correctly identifies the most load-bearing issue: the 'theoretical price' feature is never defined, and the paper's central comparison is only valid if that feature is independent of the B-S/B-S-M baselines. If the feature is the baseline formula's output, then the neural networks are given the parametric model's prediction as an input, so the reported MSE improvement (0.00790 vs 0.01552) is unsurprising and does not demonstrate that Conv-KAN is a superior pricing model in a head-to-head sense. I agree with the reader's REJECT verdict: this is a high-correctness-risk empirical claim with no code or data release, and the missing definition of a load-bearing input makes the result unfalsifiable as reported. I do not rely on the temporal-split leakage argument, because splitting by calendar date while retaining same-contract observations in both train and test is not obviously future leakage; the undefined 'theoretical price' feature is the sharper and sufficient concern. A concrete ablation removing this feature would settle whether the claimed advantage is genuine or an artifact, but as the manuscript stands, the evidence is inadequate.","tokens_in":9151,"tokens_out":8353,"duration_ms":79286,"concrete_test":"Obtain the dataset (or a sample) and recompute the Black-Scholes-Merton price from the listed S, K, T, r, q, and sigma for each row, then compare it to the 'theoretical price' column. If the values match to numerical precision, the feature is the baseline output. In that case, retrain all models (Conv-KAN, KAN, Conv-LSTM, LSTM) with the 'theoretical price' column removed, and also remove Delta if it is B-S-M-derived. If Conv-KAN no longer has the lowest MSE/RMSE/MAE, the paper's central claim collapses; if it still beats B-S-M, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Conv-KANs outperform the parametric B-S and B-S-M models on CSI 300 option pricing. For that comparison to be meaningful, the feature set given to the networks must not contain the baseline models' own outputs. Section 4.1 lists 'theoretical price' among the nine input variables but never defines how it is computed. In standard Chinese options datasets, including RESSET (the cited source), 'theoretical price' is typically the Black-Scholes or Black-Scholes-Merton formula price. If that is the case here, every neural network receives the exact quantity that the B-S/B-S-M baselines output, and the Conv-KAN result is an exercise in residual learning on top of the baseline rather than an independent model beating it. The same concern applies to 'Delta', which is also a model-dependent Greek usually computed from the same formula. The paper provides no ablation without these features and no definition, making the headline result unfalsifiable. The temporal split by calendar date, by contrast, is not obviously leakage; the load-bearing weakness is the undefined model-derived input feature.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies Kolmogorov-Arnold Networks (KANs) and Convolutional KANs (Conv-KANs) to price CSI 300 index options, comparing them with B-S, B-S-M, LSTM, and Conv-LSTM models. It proposes a data-processing strategy that splits each option's life across the training/test boundary and deliberately keeps raw noisy data to simulate a real trading environment. The headline empirical claim is that Conv-KANs achieve the lowest MSE/RMSE/MAE in Table 1 (0.00790, 0.08890, 0.07098), outperforming the parametric B-S-M baseline (0.01552, 0.12457, 0.08442). The text also claims that the B-S-M formula with a dividend-yield variable improves on the classical B-S formula.","tokens_in":9392,"tokens_out":5093,"duration_ms":43024,"significance":"The application of KANs and Conv-KANs to option pricing is novel and could be of interest to the computational-finance community. The proposed data-splitting strategy, if properly implemented, addresses a real concern about training and evaluating under conditions closer to actual trading. However, the current manuscript does not provide trustworthy evidence for its central claim: the undefined 'theoretical price' input feature is likely model-derived, the test split overlaps with training contracts, there are no repeated runs or error bars, and the MAPE results for the convolutional models are two orders of magnitude worse than those of the alternatives. No code or data is provided. These issues are load-bearing and cannot be fixed by minor editing.","major_comments":[{"comment":"The input feature 'theoretical price' is listed among the nine variables but is never defined. In the cited RESSET dataset, this field is conventionally the Black-Scholes or Black-Scholes-Merton formula price. If that is the case, every neural network receives the same quantity that the B-S and B-S-M baselines output, so the comparison in Table 1 is circular and the reported Conv-KAN advantage is an artifact of residual learning on top of the baseline rather than independent superiority. The authors must define how 'theoretical price' is computed and report results with and without this feature; the same concern applies to 'Delta', which is also model-dependent.","section":"§4.1, Table 1"},{"comment":"The proposed data-processing strategy splits individual option contracts across the training/test boundary: the text states that an option issued on July 1st and exercised on September 15th has its July 1-August 31 observations placed in training and the remainder in the test set. This means the same contract appears in both sets, so temporal patterns learned from the training portion can directly assist prediction of the test portion. Consequently, the test set is not an independent out-of-sample evaluation, and the claim of simulating a real trading environment is unsupported. A clean chronological split with no contract overlap should be used, or at least results under both split regimes should be reported.","section":"§4.1, temporal split"},{"comment":"All results correspond to a single run, with no repeated seeds, no confidence intervals, and no statistical tests. The differences between B-S and B-S-M (MSE 0.01612 vs. 0.01552) are small and could easily be within run-to-run variation, and the claimed advantages of Conv-KANs over Conv-LSTM also lack error bars. Given that the comparison involves different epoch counts and model architectures, the reported numerical ranking is not credible without repeated trials and a measure of variability.","section":"§4.2, Table 1"},{"comment":"The MAPE values for Conv-LSTM (12.36620) and Conv-KANs (15.11880) are roughly two orders of magnitude larger than those of B-S, B-S-M, LSTM, and KANs (0.61-0.79). The text dismisses this as due to 'the nature of the convolutional structures', but such extreme relative errors contradict the claim that the convolutional models are practically superior. The authors should explain this discrepancy and reconcile it with the headline MSE/RMSE/MAE results.","section":"Table 1, MAPE row"},{"comment":"The model comparison is not controlled: KANs and Conv-KANs are trained for 50 epochs while LSTM and Conv-LSTM are trained for 200, and the batch size and learning rate are fixed without any hyperparameter search. No details are given for the number of KAN layers, grid size, or the configuration of the fully connected layers in Conv-KANs. Without these details, the claimed 'superior fitting performance' of Conv-KANs cannot be separated from the choice of training budget and model capacity.","section":"§4.2, §3.4"}],"minor_comments":[{"comment":"Equation (16) labels a column 'MAP' with the formula 1/N sum |y_i - yhat_i|, which is the mean absolute error; the column should be renamed 'MAE' or the formula corrected.","section":"§4.2, Eq. (16)"},{"comment":"The competing-interest declaration ends mid-sentence: 'could have appeared to influence' should be completed (e.g., 'the work reported in this paper').","section":"§6"},{"comment":"The reference 'Ivas, cu C F' contains a spacing error; it should be 'Ivașcu C F'.","section":"References"},{"comment":"Several figures, including Figures 1 and 6, lack axis labels or adequate captions to interpret the plotted quantities.","section":"Figures 1, 6"},{"comment":"The 'white-noise / data-splitting rule' mentioned in the Introduction is not formally defined; Section 4.1 describes the split qualitatively but does not specify the exact rule for options that mature near the boundary, such as the treatment of the final day of an option in the training set.","section":"§4.1"}],"recommendation":"reject","confidential_remarks":"The central comparison is very likely circular because the 'theoretical price' feature is a standard RESSET field typically computed with the Black-Scholes formula. Even if the authors intended a different definition, the absence of any definition or ablation makes the headline result unfalsifiable. The temporal split that places the same contract in both training and test sets compounds the problem. Given that the main claim collapses without these fixes, the paper does not meet the bar for publication in its current form; a rejection is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is the first application of KANs and Conv-KANs to option pricing that I know of in its own cited literature, and the authors did real experimental work with CSI 300 data, a dividend-adjusted B-S-M baseline, and a deliberate noise-retaining split. But the headline result—Conv-KANs beating B-S/M on MSE—is not interpretable as reported. Section 4.1 lists 'theoretical price' as an input and never defines it. In the cited RESSET data, that field normally is a model price, typically B-S or B-S-M. If that is what was used, every network is handed the baseline's output as a feature, so the comparison is circular. The same goes for 'Delta', which is usually computed from the same formula. There is no ablation without these features, so the claim is unfalsifiable as written.\n\nThe split also leaks, contrary to the stress-test note: the paper's own example cuts a single option's life into a training segment and a test segment, so the same contract appears on both sides. That lets the networks memorize contract-specific patterns rather than learning to price new contracts. Combined with no error bars, no repeated runs, and no code/data, the evidence base for the central claim is thin.\n\nWhat deserves credit: applying KANs to option pricing is new, and the B-S-M dividend adjustment is a sensible improvement over the usual B-S comparison. The idea of retaining raw noise to test generalization is worth pursuing, even if the execution is flawed. The related work is relevant, and the appendix derivation is standard and correct in outline.\n\nThe MAPE column is a red flag the authors barely address: Conv-LSTM and Conv-KANs sit at 12.4 and 15.1 while B-S/M are around 0.6. The one-line 'nature of convolutional structures' explanation doesn't tell us whether these models fail systematically on low-priced options. Minor: equation (16) has a MAP/MAPE typo, and the d1 line in (14) is garbled.\n\nWho gets value: someone scouting whether KANs can be used for option pricing, but not as evidence of superiority. If the authors define the feature, run an ablation without model-derived inputs, and redo the split contract-by-contract, the result could become citable. As submitted, I would desk reject with an invitation to resubmit; it is not ready for a serious referee until the comparison is un-circular.","headline":"First KAN/Conv-KAN option-pricing application with a sensible B-S-M baseline, but the undefined 'theoretical price' feature and contract-level split make the headline MSE advantage uninterpretable.","tokens_in":9887,"tokens_out":4306,"would_cite":false,"duration_ms":42581,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91G20","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A convolutional Kolmogorov-Arnold network trained on raw CSI 300 option data prices options with lower error than Black-Scholes-Merton, LSTM, and Conv-LSTM baselines.","keywords":["option pricing","Convolutional Kolmogorov-Arnold Networks","Kolmogorov-Arnold Networks","Black-Scholes-Merton model","CSI 300 index options","LSTM","data generalization","neural network comparison"],"falsifier":"Retrain the Conv-KAN and B-S-M comparison on the same 2020 CSI 300 samples with the 'theoretical price' input removed or replaced by a neutral feature such as moneyness. If Conv-KAN's MSE no longer beats B-S-M's 0.01552, the reported advantage is an artifact of the setup rather than a property of the architecture.","tokens_in":8929,"feed_emoji":"📈","tokens_out":12280,"duration_ms":89847,"temperature":0.7,"pith_summary":"This paper tests whether modern neural architectures can beat classical parametric option-pricing formulas when trained on raw market data rather than carefully filtered data. It builds a feature set around the dividend-adjusted Black-Scholes-Merton formula and compares six models on 2020 CSI 300 Index options, using the first eight months for training and the last four for testing. The central claim is that Convolutional Kolmogorov-Arnold Networks (Conv-KANs) achieve the lowest mean squared error, RMSE, and mean absolute error of all six models, including the B-S-M parametric baseline. The authors take this as evidence that KAN-style nonlinear layers are a viable alternative to MLP-based networks for option pricing, and that their chronological train/test split better reflects how a trader actually encounters option data.","feed_headline":"Conv-KANs beat B-S-M on raw CSI 300 option prices","feed_subtitle":"It posts the lowest MSE, RMSE, and MAE across six models on 2020 CSI 300 option data.","key_machinery":"Two mechanisms carry the argument. The first is the Conv-KAN layer, in which each kernel entry is a learnable one-dimensional function $\\phi(x)=w_1\\,\\mathrm{spline}(x)+w_2\\,\\mathrm{silu}(x)$ rather than a scalar weight. For an input image $a$ and a $N\\times M$ KAN kernel $K$, the convolution output is $(i\\ast K)_{i,j}=\\sum_{k=1}^{N}\\sum_{l=1}^{M}\\phi_{kl}(a_{i+k,j+l})$, so the nonlinearity happens inside the kernel. The paper keeps a conventional fully connected output layer to isolate the effect of the nonlinear convolution. The second mechanism is the data-selection strategy: options are split by calendar date rather than filtered on moneyness, so the training set retains raw noisy observations and some options appear for only one day, forcing the network to generalize.","core_discovery":"On 2020 CSI 300 Index options, the paper reports that a Conv-KAN trained for 50 epochs reaches MSE 0.00790, RMSE 0.08890, and MAE 0.07098, the best among Black-Scholes, Black-Scholes-Merton, LSTM, Conv-LSTM, KAN, and Conv-KAN. It also reports that the dividend-adjusted B-S-M formula with dividend rate $q$ improves on the basic B-S formula (MSE 0.01552 versus 0.01612). These numbers are used to argue that KAN-based architectures are a new and effective option-pricing method and that training on deliberately noisy, unfiltered data improves generalization.","pith_inferences":["A separate ablation that removes the undocumented 'theoretical price' input would test whether the Conv-KAN advantage is intrinsic or inherited from the parametric baseline; this check is not reported in the paper.","Because the train/test split is chronological, the results also capture any market-regime shift between the first eight and last four months of 2020; a rolling-window or shuffled split on the same raw data would isolate the architecture's contribution.","The high MAPE of the convolutional models (15.12 for Conv-KANs versus 0.61 for B-S-M) suggests the improved MSE comes from the bulk of prices while small or near-zero prices are poorly predicted; stratifying errors by moneyness or time to maturity would show where the gains and losses are.","The same architecture and data strategy could be moved to other index-option markets with dividend-paying underlyings, which would indicate whether the result is specific to CSI 300 market microstructure."],"forward_implications":["If the reported errors hold, Conv-KANs can price options from raw data without pre-filtering on moneyness, lowering the data-preparation burden.","KAN and Conv-KAN can join the standard option-pricing benchmark set, giving future work a comparison point against MLP, LSTM, and parametric baselines.","The B-S-M feature set with dividend rate q is confirmed as a useful input encoding for dividend-paying underlyings.","Conv-KANs reach the best error in 50 epochs versus 200 for LSTM, suggesting the advantage is architectural rather than a matter of training budget.","The convolutional models' high MAPE shows the error reduction is not uniform across all option prices, pointing to uneven fitting across the price range."],"supporting_citations":[{"why":"Supplies the original no-dividend Black-Scholes formula that serves as the first parametric baseline.","marker":"[1]"},{"why":"Supplies the dividend-adjusted Black-Scholes-Merton formula, used as the improved baseline and as the source of the dividend-rate feature.","marker":"[2]"},{"why":"Provides the prior option-price prediction setup with LSTM and Conv-LSTM models that this paper compares against.","marker":"[3]"},{"why":"Cited for the empirical claim that machine-learning models can outperform parametric option-pricing models, motivating the benchmark comparison.","marker":"[4]"},{"why":"Cited for the generalization-and-noise argument that justifies retaining raw noisy data instead of filtering on moneyness.","marker":"[5]"},{"why":"Defines the Convolutional LSTM architecture used as the convolutional time-series baseline.","marker":"[6]"},{"why":"Introduces Kolmogorov-Arnold Networks and the learnable spline-activation layer that Conv-KANs build on.","marker":"[7]"},{"why":"Introduces Convolutional Kolmogorov-Arnold Networks, the architecture under test.","marker":"[8]"},{"why":"Refines Kolmogorov-Arnold convolution designs and supports the Conv-KAN construction choices.","marker":"[9]"}],"fun_headline_variants":["Conv-KAN beats five rivals on CSI 300 options","Conv-KAN posts lowest error on 2020 CSI 300 options","KAN-based networks improve option pricing accuracy","Conv-KAN beats Black-Scholes on real option data","Noisy data helps Conv-KAN generalize on options"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that the 'theoretical price' variable used as a network input is not itself generated by the Black-Scholes or Black-Scholes-Merton baseline formula; if it is, the neural models are simply receiving the parametric answer as a feature.","fun_headline_variants_meta":{"raw":{"variants":["Conv-KAN beats five rivals on CSI 300 options","Conv-KAN posts lowest error on 2020 CSI 300 options","KAN-based networks improve option pricing accuracy","Conv-KAN beats Black-Scholes on real option data","Noisy data helps Conv-KAN generalize on options"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00081,"raw_usage":{"total_tokens":3489,"prompt_tokens":817,"completion_tokens":2672,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":433,"completion_tokens_details":{"reasoning_tokens":2592}},"tokens_in":433,"tokens_out":2672,"duration_ms":15643,"temperature":1.0,"reasoning_tokens":2592,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:33:08.708398+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the Conv-KAN and B-S-M comparison on the same 2020 CSI 300 samples with the 'theoretical price' input removed or replaced by a neutral feature such as moneyness. If Conv-KAN's MSE no longer beats B-S-M's 0.01552, the reported advantage is an artifact of the setup rather than a property of the architecture.","supporting_citations":[{"cited_title":"The pricing of options and corporate liabilities[J]","cited_arxiv_id":null,"evidence_quote":"Supplies the original no-dividend Black-Scholes formula that serves as the first parametric baseline."},{"cited_title":"Theory of rational option pricing[J]","cited_arxiv_id":null,"evidence_quote":"Supplies the dividend-adjusted Black-Scholes-Merton formula, used as the improved baseline and as the source of the dividend-rate feature."},{"cited_title":"3D Tensor-based Deep Learning Models for Predicting Option Price[C]","cited_arxiv_id":null,"evidence_quote":"Provides the prior option-price prediction setup with LSTM and Conv-LSTM models that this paper compares against."},{"cited_title":"Option pricing using machine learning[J]","cited_arxiv_id":null,"evidence_quote":"Cited for the empirical claim that machine-learning models can outperform parametric option-pricing models, motivating the benchmark comparison."},{"cited_title":"Generalization and noise[J]","cited_arxiv_id":null,"evidence_quote":"Cited for the generalization-and-noise argument that justifies retaining raw noisy data instead of filtering on moneyness."},{"cited_title":"Convolutional LSTM network: A machine learning approach for precipitation nowcasting[J]","cited_arxiv_id":null,"evidence_quote":"Defines the Convolutional LSTM architecture used as the convolutional time-series baseline."}],"review_version":1}