{"id":"a1232450-2f11-4d34-ad60-096fbce0afce","arxiv_id":"2504.20088","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A residual network trained on eight years of Petrobras option data reports lower mean absolute error than Black-Scholes in the 3-19 BRL range, but the reported 64.3% reduction is contradicted by the paper's own table, which shows a 35.7% reduction.","lead":"A small deep-learning model trained on Brazilian Petrobras option data reports lower pricing error than Black-Scholes in a 3 to 19 real price range, but the abstract's headline 64.3% reduction is contradicted by the paper's own table, which implies a 35.7% reduction. The comparison is a useful emerging-market benchmark, not a breakthrough in option pricing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's 64.3% MAE reduction claim is contradicted by Table 3, which actually shows a 35.7% reduction.","rationale":"The reader's strongest claim is exactly the arithmetic mismatch I identified: the abstract reports a 64.3% reduction while Table 3 supports only a 35.7% reduction. The reader's stated weakest_assumption concerns data parsing and matching, which is a legitimate risk but not the most decisive issue. The most load-bearing concern is internal: the paper's own table falsifies the headline percentage. This does not require external data or assumptions about B3 parsing; it follows from the reported MAE values. The paper still shows the deep learning model with lower MAE in the selected band, so the corrected claim would be a modest outperformance, not a 64.3% error reduction. Because the central claim as written is unsupported, the REJECT verdict remains appropriate. I did not find a separate concern that would change the verdict; the numerical correction and the additional methodological weaknesses, such as the no-dividend Black-Scholes baseline, the historical-volatility input, the hybrid loss using Black-Scholes labels, and the lack of released code and data, all reinforce the rejection but are secondary to the false headline. A path to a conditional acceptance would require correcting the percentage, pre-specifying the range, providing uncertainty estimates, and making the data and code available.","tokens_in":9443,"tokens_out":3591,"duration_ms":36435,"concrete_test":"Recompute the percentage reduction from the values the paper itself reports in Table 3: (2.0295 - 1.3047)/2.0295 = 0.357, and 1.3047/2.0295 = 0.643. If Table 3 is accepted, the abstract and Section 4.3 must be corrected to state a 35.7% reduction, and the '64.3%' figure should be removed or relabeled as the ratio DL/BS. If the authors release the cleaned dataset and code, an additional check is to recompute test-set MAE on the 3-19 BRL band using the exact records and sample count, and to report a confidence interval; this would test whether the qualitative direction of the result survives the corrected percentage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantified claim, stated in the abstract and repeated in Section 4.3, is that the deep residual network achieved a 64.3% reduction in mean absolute error over Black-Scholes in the 3-19 BRL range. This claim is contradicted by the paper's own Table 3. With DL MAE = 1.3047 and BS MAE = 2.0295, the reduction is (2.0295 - 1.3047)/2.0295 = 0.357, or 35.7%. The value 64.3% is instead DL/BS = 1.3047/2.0295, i.e., the deep learning error is 64.3% of the Black-Scholes error, not a 64.3% reduction. If the claimed 64.3% reduction were true, the DL MAE would be about 0.724. The paper provides no confidence intervals, no sample count for the 3-19 BRL subset, and no pre-registered or pre-specified range definition; the range appears to be selected from test-set plots. Because the numerical headline is central to the claimed practical advantage, the paper as written does not support its main result. This is an internal arithmetic inconsistency, not a disagreement with market consensus, and it can be settled directly from the paper's own reported numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares a deep residual network trained on eight years of Brazilian B3 option data with the Black-Scholes formula for pricing Petrobras European call options. The authors scrape option data from B3, combine it with PETR4 stock prices, rolling volatility, and SELIC rates, and train a five-input network with a hybrid loss that uses market prices when the premium exceeds 0.1 BRL and Black-Scholes prices otherwise. They report a 64.3% reduction in mean absolute error over Black-Scholes in the 3-19 BRL price range on a test set from November 2024 to January 2025, and discuss performance by expiration category and ticker.","tokens_in":9643,"tokens_out":5664,"duration_ms":53593,"significance":"If substantiated, the result would be a useful case study showing that a modestly sized residual network can beat Black-Scholes in a specific price band for a liquid emerging-market underlying, with practical implications for pricing low-priced options. The paper's strengths are its detailed description of the data-scraping pipeline, the explicit reporting of MAE by expiration category, and the use of a consumer-grade GPU. However, the headline performance claim is arithmetically inconsistent with the paper's own Table 3, the 3-19 BRL range is selected post hoc, and there is no statistical uncertainty quantification, so the significance as written is not established.","major_comments":[{"comment":"The claimed '64.3% reduction in MAE' is contradicted by Table 3. Table 3 reports DL MAE 1.3047 and BS MAE 2.0295 in the 3-19 BRL range; the reduction is (2.0295 - 1.3047) / 2.0295 = 35.7%, not 64.3%. The 64.3% figure is the ratio DL/BS, i.e., the deep learning error as a percentage of the Black-Scholes error. Please correct the abstract, §4.3, and §5, and restate the magnitude of the improvement consistently with Table 3.","section":"Abstract; §4.3; §5"},{"comment":"The 3-19 BRL band appears to be selected after inspecting test-set performance, with no pre-specified rule or correction for multiple comparisons. The paper does not report how many test samples fall in this band, how sensitive the result is to the band boundaries, or confidence intervals for the MAE difference. Without this, the reported improvement cannot be distinguished from selection over price bins. Please report sample sizes, a pre-registered or at least a priori defined range, and uncertainty estimates.","section":"§4.2; Fig. 5"},{"comment":"The hybrid loss sets γ_i = 0 for samples with market premium at most 0.1 BRL, training the network to reproduce Black-Scholes prices for exactly the samples that are then used to compare the network against Black-Scholes in the overall test set. This makes the full-sample comparison partially circular. Although the 3-19 BRL band likely uses only market-labeled samples, the paper should quantify how many test samples in each reported comparison were trained with the Black-Scholes target and re-run the comparison excluding them.","section":"§3.2"},{"comment":"The data foundation is built entirely from bespoke web scraping and file parsing with no validation against a second source. Errors in parsing the .ex files, matching option records to PETR4 prices, or applying the 'closest preceding date' rule would corrupt both the training labels and the Black-Scholes inputs. Please validate a random sample of records against an independent data provider and report the error rate, and describe the de-duplication and cleaning rules in sufficient detail for replication.","section":"§2.2-2.3"},{"comment":"The text states that Fig. 2 'is not the original loss graph from that process' but 'serves as a representative example.' A result figure should display the actual recorded loss curves; presenting a representative or schematic version without a clear statement as such is misleading and prevents readers from assessing the hyperparameter search. Please provide the actual Optuna curves or remove the figure.","section":"§4.1; Fig. 2"}],"minor_comments":[{"comment":"The phrase 'Nobel prize for Economy' should be 'Nobel Memorial Prize in Economic Sciences.'","section":"§1"},{"comment":"The x-axis labels in Fig. 3 are garbled ('4 510152025') and the axis is not clearly labeled; please replot the figure.","section":"Fig. 3"},{"comment":"The column headers 'last_field' and 'return' are undefined; please define them in the caption or in the text.","section":"Table 2"},{"comment":"The abstract and §5 repeat '43.41% of all Petrobras option transactions' without defining the denominator; specify exactly how this percentage is computed, and use a round number or report the exact count.","section":"§2.1; §4.3"},{"comment":"The description of the rolling volatility calculation should state explicitly that the 21-day window ends on the trade date and that no future information is used; otherwise readers cannot assess look-ahead bias.","section":"§2.3"}],"recommendation":"major_revision","confidential_remarks":"The arithmetic error in the headline result is a serious internal inconsistency that must be corrected before further consideration. In addition, the post hoc range selection and the partial circularity of the hybrid loss mean the paper currently overstates the evidence. The manuscript would benefit from a professional statistical review; the current presentation is not yet at journal standard. If the authors can correct the magnitude, pre-specify the evaluation range, add uncertainty quantification, and validate the data pipeline, the corrected claim may be publishable as a case study. I recommend major revision rather than rejection because the underlying comparison is recoverable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Start with the thing that matters: the abstract's central claim is not what the paper's own Table 3 shows. The paper says the deep residual network achieved a 64.3% reduction in MAE for the 3–19 BRL band. Table 3 reports DL MAE 1.3047 vs BS MAE 2.0295. That is a 35.7% reduction, not 64.3%. The 64.3% figure is actually the ratio of DL error to BS error—the network error is 64.3% of the baseline, which is not the same as a 64.3% improvement. I checked: (2.0295−1.3047)/2.0295 = 0.357. So the headline is wrong by a factor of 1.8, in the direction that flatters the model.\n\nWhat is genuinely new here is the Brazilian Petrobras dataset. The authors scraped B3 directly, assembled eight years of option data, and trained a residual network on a laptop GPU. That is real work, and the paper is honest about its constraints and limitations. A cleaned version of that dataset, with code, would be useful to people working on emerging-market options.\n\nBut the soft spots are serious. First, the arithmetic error is not cosmetic—it is the paper's selling point. Second, the 3–19 BRL band appears to be chosen post hoc from the test-set plots; no pre-specified range, sample count, or confidence intervals are given. Third, the hybrid loss in Section 3.2 sets the training target to Black-Scholes whenever the market premium is at most 0.1 BRL, so part of the network is trained to reproduce the same baseline it is later compared against. The paper does not say how many test samples fall in that masked regime. Fourth, the Black-Scholes baseline uses 21-day historical volatility and omits dividends, even though the B3 data includes implied volatility per strike. That is a weak baseline, and the footnote dismissing dividends is unconvincing—omitting them systematically biases the baseline in a way that helps the comparison. Fifth, there is no code or data release, and the date-matching pipeline is validated against nothing.\n\nWho is this for? A reader looking for a cautionary example of post hoc range selection and contaminated training targets might get something out of it. As a contribution to option pricing, though, the load-bearing claim fails. I would not send this to referees in its current form. If the authors correct the percentage, pre-specify the evaluation range, replace or quantify the hybrid loss, and release code/data, a shorter empirical note could be credible. As written, the paper does not support its own abstract.","headline":"The paper's headline 64.3% MAE reduction is contradicted by its own Table 3 (actual reduction: 35.7%), and the hybrid loss trains part of the network on the Black-Scholes baseline it is later compared against.","tokens_in":10273,"tokens_out":3774,"would_cite":false,"duration_ms":35771,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91G20","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Residual network cuts option-price error by a third on Petrobras calls","keywords":["deep residual network","option pricing","Black-Scholes","Petrobras","B3 exchange","hybrid loss function","mean absolute error","Brazilian options"],"falsifier":"A reader can check the arithmetic directly: Table 3 gives $1.3047/2.0295 = 0.643$, so the reported 64.3% reduction should be about 35.7%; the stronger test is to re-run the comparison after re-parsing the B3 files with an independent parser and a second price source, since a date or price mismatch would change both the network's training targets and the Black-Scholes comparison.","tokens_in":9107,"feed_emoji":"📈","tokens_out":11891,"duration_ms":102095,"temperature":0.7,"pith_summary":"The paper tries to establish that a small deep residual network, trained on eight years of scraped Brazilian exchange data, can price Petrobras call options more accurately than the Black-Scholes formula in the 3–19 BRL band, where 43.41% of the option transactions sit. On its three-month test window, the network's mean absolute error is 1.3047 against Black-Scholes's 2.0295, and it wins in every expiration bucket. The abstract calls this a 64.3% reduction, but the table showing these figures implies the network error is 64.3% of the Black-Scholes error, a 35.7% reduction. The paper also reports that the network's error falls, rather than rises, as expiration lengthens, opposite to the usual Black-Scholes behavior. The authors conclude that deep learning is a viable alternative to traditional formulas for specific price ranges and contract structures.","feed_headline":"Residual network cuts option-price error by a third on Petrobras calls","feed_subtitle":"In the 3-19 BRL band, mean absolute error is 1.30 for the neural net vs 2.03 for Black-Scholes.","key_machinery":"The load-bearing mechanism is a residual fully connected network with a skip connection between the first and second hidden layers, trained under a hybrid loss $\\ell_i = \\gamma_i (p_i - m_i)^2 + (1 - \\gamma_i)(p_i - bs_i)^2$, where the binary mask $\\gamma_i$ is 1 when the market premium exceeds 0.1 BRL and 0 otherwise. The five inputs mirror the Black-Scholes variables: $\\ln(S)$, $\\ln(K)$, time to expiry, annualized 21-day rolling volatility, and the SELIC rate. The residual connection and the loss's fallback to Black-Scholes values are what let the model stay anchored to theory where market quotes are unreliable while still fitting the observed price surface.","core_discovery":"The central claim, stated on the authors' terms, is that a residual fully connected network with five inputs—$\\ln(S)$, $\\ln(K)$, time to expiration, annualized 21-day rolling volatility, and the SELIC rate—prices Petrobras call options in the 3–19 BRL range with lower mean absolute error than Black-Scholes on a three-month held-out test set from November 2024 to January 2025. The reported mean absolute errors are 1.3047 for the network and 2.0295 for Black-Scholes across that band, and the network also has lower error within each expiration bucket of one, two, and three months. The authors attribute the result to a hybrid loss that fits market prices when the quoted premium is above 0.1 BRL and falls back to Black-Scholes values otherwise, together with the residual shortcut between hidden layers. They also find that the network does not share Black-Scholes's tendency to lose accuracy as expiration lengthens. A caveat visible inside the paper: the abstract's 64.3% reduction number does not match Table 3, whose values imply the deep learning error is 64.3% of the Black-Scholes error—a 35.7% reduction.","pith_inferences":["The 64.3% figure in the abstract looks like a ratio inversion; correcting it to roughly 35.7% makes the headline more modest but does not erase the consistent per-expiration advantage reported in Table 3.","The result is tied to one bespoke scraping pipeline with no second-source validation; re-establishing the same comparison with an independent parser and a second price provider would show whether the edge survives.","The hybrid loss means the network is partly trained on Black-Scholes values, so part of its apparent advantage may come from inheriting the formula's structure rather than from learning market dynamics alone.","A natural extension, already suggested by the authors, is specialized models per price range; an equally natural check is to test the same architecture on other Brazilian underlyings or later time windows, which would separate the effect of the residual connection from the effect of the loss design."],"forward_implications":["For Petrobras options priced between 3 and 19 BRL, the network's mean absolute error is about one-third lower than Black-Scholes's on the 2024–2025 test window.","The network's error is lowest at three months to expiration, the reverse of the Black-Scholes pattern in this dataset, where error grows with time to expiration.","Options around 5 BRL that Black-Scholes systematically undervalues are priced closer by the network, suggesting the model captures a market feature the formula misses.","Outside the 3–19 BRL band the network overestimates prices, so the claimed advantage is range-specific rather than global.","Because the loss uses Black-Scholes prices whenever the market quote is below the reliability threshold, the network's output is partly a blend of empirical and theoretical pricing, not a purely data-driven price."],"supporting_citations":[{"why":"Supplies the Black-Scholes call-option formula that serves as the baseline and as the fallback target in the hybrid loss.","marker":"[2]"},{"why":"Establishes the residual-network approach to option pricing that this architecture is built on.","marker":"[13]"},{"why":"Provides the five input features and the base MLP architecture that this network adapts.","marker":"[19]"},{"why":"Motivates applying neural pricing to a high-volatility, lower-liquidity emerging market like Brazil.","marker":"[20]"},{"why":"Introduces the residual or skip connection that the model's hidden-layer shortcut comes from.","marker":"[21]"},{"why":"Justifies the batch normalization applied at each fully connected layer.","marker":"[22]"},{"why":"Supplies the optimizer used to train the network.","marker":"[23]"},{"why":"Provides the hyperparameter search procedure that selected learning rate, hidden size, dropout, and weight decay.","marker":"[24]"}],"fun_headline_variants":["Neural net trims Petrobras option error by 35.7%","Deep learning beats Black-Scholes on Petrobras options","Residual network holds accuracy on long-dated Petrobras options","Deep learning edges out Black-Scholes on Petrobras pricing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every scraped B3 record is parsed and matched to the correct trade date, stock price, volatility, and interest rate, with no check against a second source.","fun_headline_variants_meta":{"raw":{"variants":["Neural net trims Petrobras option error by 35.7%","Deep learning beats Black-Scholes on Petrobras options","Residual network holds accuracy on long-dated Petrobras options","Deep learning edges out Black-Scholes on Petrobras pricing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001131,"raw_usage":{"total_tokens":4736,"prompt_tokens":1018,"completion_tokens":3718,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":3647}},"tokens_in":634,"tokens_out":3718,"duration_ms":27608,"temperature":1.0,"reasoning_tokens":3647,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:14:35.145861+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader can check the arithmetic directly: Table 3 gives $1.3047/2.0295 = 0.643$, so the reported 64.3% reduction should be about 35.7%; the stronger test is to re-run the comparison after re-parsing the B3 files with an independent parser and a second price source, since a date or price mismatch would change both the network's training targets and the Black-Scholes comparison.","supporting_citations":[{"cited_title":"The pricing of options and corporate liabilities","cited_arxiv_id":null,"evidence_quote":"Supplies the Black-Scholes call-option formula that serves as the baseline and as the fallback target in the hybrid loss."},{"cited_title":"Option pricing based on the residual neural network","cited_arxiv_id":null,"evidence_quote":"Establishes the residual-network approach to option pricing that this architecture is built on."},{"cited_title":"Cs230: Deep learning - option pricing with deep learning, 2019","cited_arxiv_id":null,"evidence_quote":"Provides the five input features and the base MLP architecture that this network adapts."},{"cited_title":"A deep learning approach for option pricing in emerging markets","cited_arxiv_id":null,"evidence_quote":"Motivates applying neural pricing to a high-volatility, lower-liquidity emerging market like Brazil."},{"cited_title":"Batch normalization: Accelerating deep network training by reducing internal covariate shift","cited_arxiv_id":null,"evidence_quote":"Justifies the batch normalization applied at each fully connected layer."},{"cited_title":"Optuna: A next-generation hyperparameter optimization framework","cited_arxiv_id":null,"evidence_quote":"Provides the hyperparameter search procedure that selected learning rate, hidden size, dropout, and weight decay."}],"review_version":1}