{"id":"d726edb8-1a9a-40f7-a04a-200ae7cc78c1","arxiv_id":"2411.19317","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A neural network trained to invert rough Heston parameters from implied volatility surfaces relies most on short-maturity deep in-the-money prices, a pattern absent in the standard Heston model.","lead":"This paper trains a neural network to recover rough Heston model parameters from implied volatility surfaces, then applies four interpretability methods to see which surface points drive the network's predictions. The dominant drivers turn out to be short-maturity, deep in-the-money volatilities, a pattern tied to the rough model's steep left wing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline interpretability result may be an artifact of the uniform parameter sampling and non-identifiability, rather than a property of rough Heston; a retraining test would settle it.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the trained FNN is treated as a faithful surrogate for the true rough Heston parameter-to-surface map. My analysis agrees and sharpens it: the paper's own evidence (Section 3.5, Tables 4–5) shows both non-bijectivity and substantial out-of-sample errors for kappa and rho, so the network's attributions for these parameters—and possibly the overall pattern—may reflect training-distribution artifacts rather than model sensitivities. The consistency across LIME, DeepLIFT, LRP, and SHAP is supportive but not decisive, because all four methods are applied to the same network trained on the same uniform synthetic data; agreement among them does not establish faithfulness to the underlying model. The proposed concrete test—retraining under an alternative sampling distribution—directly addresses the weakest point, since the uniform parameter distribution is a modeling choice in Section 3.2 and not a consequence of the rough Heston model. If the attribution pattern is stable under this perturbation, the central claim is substantially strengthened; if not, the paper's interpretation must be revised to refer to the network's learned inverse rather than to the model. The reader's CONDITIONAL verdict already accounts for this uncertainty, so I do not change the verdict; however, the specific test should be a condition for acceptance.","tokens_in":15397,"tokens_out":6610,"duration_ms":60189,"concrete_test":"Retrain the FNN with the same architecture, data-generation scheme, and grid, but sample the parameter labels from a non-uniform distribution—e.g., log-uniform for κ, ν, θ, V0 and truncated normal for ρ and H—or, alternatively, from a mixture of two widely separated sub-boxes. Recompute the global SHAP and DeepLIFT attributions on the same test surface grid. If the top feature (K,T)=(0.6,0.6) and the moneyness/maturity pattern for each parameter persist across these sampling schemes, the claim is robust to the training prior; if the ranking changes substantially, the original attributions are distribution artifacts rather than model sensitivities.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the FNN inverse map for rough Heston is driven primarily by short-maturity deep in-the-money implied volatilities, and this is interpreted as a property of the model ('in line with research on the model itself' and the rough-Heston-specific left wing). The load-bearing step is the assumption that attribution scores on the trained FNN faithfully reveal the true rough Heston parameter-to-surface map. This assumption is insecure in two ways acknowledged in the paper itself. First, Section 3.5 reports that the map from parameters to surfaces is 'probably not bijective', so the inverse problem is ill-posed; for any surface, multiple parameter sets may be consistent, and the network's predictions depend on how the training objective resolved that ambiguity. Second, Tables 4 and 5 show that the FNN's out-of-sample errors for kappa and rho are large (0.1279 and 0.5593 for the narrow-range network; 0.1167 and 0.1650 for the broad-range network used in the interpretability study), so for these parameters the network is not a faithful surrogate away from the training distribution. The reported attributions for kappa and rho (and the overall pattern) may therefore reflect artifacts of uniform sampling over the parameter box and arbitrary tie-breaking among non-identifiable parameters, rather than an intrinsic property of the rough Heston model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains a feedforward neural network (FNN) to invert the rough Heston model's implied volatility surface into its six parameters (ρ, V0, κ, θ, ν, H) using synthetic data generated from two parameter ranges, and then applies four interpretability techniques (LIME, DeepLIFT, LRP, and DeepSHAP) to the trained network. The main finding is that the network attributes its parameter predictions predominantly to short-maturity deep in-the-money implied volatilities, with ρ read from both wings and θ from longer maturities; this pattern is reported as consistent across the four attribution methods. The authors interpret this as aligning with the rough-Heston-specific left-wing behavior of the volatility surface, and compare with a previous plain-Heston interpretability study. The paper also reports calibration accuracy and out-of-sample performance for the two networks, acknowledging non-identifiability of the parameter-to-surface map and large out-of-sample errors for κ and ρ.","tokens_in":15650,"tokens_out":6112,"duration_ms":54031,"significance":"If the central interpretability claim is robust, the paper offers a genuinely useful case study for deep calibration in rough volatility, showing where a neural-network calibrator concentrates its attention and providing a caution that liquid at-the-money quotes are not necessarily the dominant drivers of the learned inverse map. The convergence of four different attribution methods on the same qualitative pattern, and the consistency with known results on the rough Heston left wing (Keller-Ressel and Majid, Fukasawa), are notable strengths. However, the significance is currently tempered by the lack of any quantitative validation of the FNN attributions against the true forward map, the absence of error bars or seed-stability checks, and the acknowledged non-identifiability of the inverse problem; these issues leave open the possibility that the headline pattern is an artifact of the network training or the chosen grid.","major_comments":[{"comment":"The central claim, stated in Section 5 as 'short-maturity deep in-the-money volatilities actually provide the largest contribution overall', is presented as a property of the rough Heston model and its neural-network calibrator, but the attribution evidence comes from a single trained FNN whose inverse map is acknowledged in Section 3.5 to be 'not bijective'. The out-of-sample errors in Tables 4 and 5 are large for κ (0.1167) and ρ (0.1650) even for the broad-range network used in the interpretability study. These facts raise a load-bearing concern: the network's feature attributions may reflect how the training objective resolved non-identifiability and how the uniform parameter sampling shaped the loss landscape, rather than the intrinsic sensitivity structure of the rough Heston surface. To support the model-level interpretation in Section 4.3 and Section 5, the paper should either directly compare FNN attributions with the true local sensitivity of the forward map (e.g., finite-difference derivatives of σ_imp with respect to each parameter along the same grid), or add a retraining analysis with different random seeds, different sampling distributions, and perhaps different network architectures, to show that the attribution pattern is stable and not an artifact of a particular learned inverse.","section":"§3.5, Tables 4-5, §5"},{"comment":"The single most important feature identified by every attribution method is the grid corner (K,T)=(0.6,0.6), which is the extreme lower-left point of the chosen strike-maturity grid (K ∈ {0.6,...,1.4}, T ∈ {0.6,...,2}). No robustness check is provided against the choice of grid boundaries. If the strike grid extended to lower strikes, the 'deep in-the-money' attribution might shift toward newly included points, and the specific claim about the left wing at short maturities could be affected. The authors should either extend the grid (e.g., include strikes 0.2-0.5) or perform a test on a sub-grid that removes the corner from the boundary, to demonstrate that the headline pattern is not a boundary artifact.","section":"§3.3, §4.1, Figures 6-12"},{"comment":"All attribution results are reported for a single trained network (the broad-range FNN), with no error bars, confidence intervals, or stability analysis across network initializations. The paper states in Section 3.5 that the FNN calibration was repeated under different configurations, but the interpretability study is based on a single final network, and the four attribution methods are evaluated on that same network. Consequently, the observed agreement among LIME, DeepLIFT, LRP, and SHAP does not by itself rule out network-specific artifacts or chance patterns. The authors should compute attribution heatmaps for at least several random seeds (and ideally for both scaling approaches) and report the dispersion of the average absolute attribution for the top features, to establish that the qualitative pattern is statistically stable.","section":"§4.1-4.2, Figures 5-12, §3.5"}],"minor_comments":[{"comment":"The 'accuracy' curves in Figures 3 and 4 are not defined for a regression problem; please clarify the metric or replace it with a standard regression accuracy measure, e.g., 1 - normalized mean squared error.","section":"§3.5, Figures 3-4"},{"comment":"The attribution methods are applied to scaled and ZCA-whitened data, but the paper does not explain how attributions computed on whitened features are mapped back to the original (K,T) grid shown in the heatmaps. Since ZCA is an invertible linear transformation, the back-transformation should be stated explicitly to avoid ambiguity.","section":"§3.4, §4.1"},{"comment":"The sentence 'it should be noted that the feature influencing the prediction most is far from the money' is confusing because the feature (K,T)=(0.6,0.6) is deep in the money for a call option with spot 1; the authors likely mean 'far from at-the-money' rather than 'far from the money', and should rephrase.","section":"§4.2"},{"comment":"The error metric in equation (3.3) is a mean-squared logarithmic ratio, so the reported numbers are dimensionless; stating this explicitly next to the tables would help readers interpret the magnitudes.","section":"§3.5, Tables 4-5"},{"comment":"The description of LIME omits practical implementation details such as the number of perturbed samples, the kernel width, and how the neighborhood is defined; adding these would improve reproducibility.","section":"§4.1, LIME description"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of q-fin.CP and builds transparently on the authors' earlier Heston interpretability work [17]; there is no circularity concern. The main risk is that the central claim—that the FNN is driven by the rough-Heston left wing—is presented both as a property of the network and as a property of the model, without sufficient evidence to separate the two. The proposed additional experiments (seed stability, grid extension, and direct comparison with forward-map sensitivity) are feasible within the manuscript's scope and would materially strengthen the paper. I do not see a need for rejection, but the current qualitative evidence is not yet sufficient for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Know this paper as the first interpretability study for a rough Heston calibration network. Its main empirical claim—that short-maturity deep in-the-money implied volatilities dominate the FNN's parameter predictions, consistently across LIME, DeepLIFT, LRP, and SHAP—is new relative to the same authors' Heston paper and is consistent with the known steep left wing of rough Heston at short maturities. That is a real finding, and the paper connects it to the right theory (Fukasawa's short-time skew, Keller-Ressel and Majid's comparison principle). The comparison table against plain Heston is a useful touch.\n\nThe paper is honest about its own limitations: Section 3.5 reports non-bijectivity of the parameter-to-surface map and out-of-sample failures for kappa and rho on the narrow range, which is why they switch to the broad-range network. The interpretability analysis is qualitatively convincing because four methods agree. But the evidence is all qualitative: no error bars, no seeds, no code or data release. The final network was selected after an out-of-sample failure, which introduces selection bias in the story. The stress-test concern—that uniform sampling over the parameter box plus non-identifiability could shape what the network learns—is plausible and deserves a retraining test: train the same architecture on a different sampling distribution and see if the attribution pattern persists. That would settle whether the left-wing dominance is a property of rough Heston or of the training design. Also, the LRP text in Section 4.1 appears to contradict the figures and tables; it says long maturity for kappa and nu when the heatmaps show short maturity. That needs fixing.\n\nNone of this invalidates the central claim as a statement about the trained FNN. The paper should be careful not to overstate it as a direct statement about the model itself; attribution scores describe the network, not the model. But given the theory link, the finding is likely to transfer.\n\nBottom line: this is a legitimate empirical study with a novel, theory-consistent result. It deserves a serious referee and would benefit from a major revision that adds code and data, quantitative attribution stability (e.g., standard deviations across retrained networks), and corrects the LRP text. I'd send it to review with the expectation of heavy revision.","headline":"First interpretability study for rough Heston calibration nets: consistent finding that short-maturity deep ITM vols dominate, but the evidence is qualitative and the training design may shape the result.","tokens_in":16183,"tokens_out":2223,"would_cite":false,"duration_ms":18937,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","91G20","91G60"],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural network trained to read rough Heston parameters from volatility surfaces relies most on short-maturity, deep in-the-money options.","keywords":["option pricing","rough volatility","rough Heston model","deep learning","neural network interpretability","Shapley values","surrogate models","implied volatility"],"falsifier":"Train the same FNN architecture and data pipeline on rough Heston surfaces with H fixed at 0.5 (the non-rough Brownian limit) and compare the LIME, DeepLIFT, LRP, and SHAP attribution maps. If the short-maturity deep in-the-money region remains the top contributor, the paper's explanation that the network keys on rough Heston's distinctive left wing is wrong; if the attribution shifts toward at-the-money or spreads evenly, the claim is supported.","tokens_in":15187,"feed_emoji":"📈","tokens_out":11376,"duration_ms":86386,"temperature":0.7,"pith_summary":"The paper asks: when a neural network is trained to recover rough Heston model parameters from an implied volatility surface, which points on that surface actually drive the predictions? The authors train a feedforward network on synthetic rough Heston surfaces and interrogate it with four attribution methods—LIME, DeepLIFT, LRP, and SHAP. Their central result is that short-maturity, deep in-the-money implied volatilities contribute the largest share of the parameter predictions overall, while correlation ρ is read from both wings and the long-term variance θ from longer maturities. This runs against the common expectation that liquid at-the-money quotes dominate calibration, and it matters because interpretability tools give a safety check for deep-learning calibrations: knowing which surface region drives the network tells practitioners where calibration errors and data noise will have the most impact.","feed_headline":"Neural calibration reads rough Heston from its left wing","feed_subtitle":"Short-maturity, deep in-the-money volatilities drive parameter estimates, not at-the-money quotes.","key_machinery":"The load-bearing object is the trained feedforward neural network (one hidden layer, ELU activations, trained on 10,000 synthetic rough Heston implied-volatility surfaces with ZCA-whitened inputs) viewed as an inverse map from surface points to the six parameters {ρ, V0, κ, θ, ν, H}. The argument is carried by four attribution methods that assign each surface point a relevance score for each predicted parameter: LIME (local linear surrogate), DeepLIFT (backpropagated activation differences), LRP (layer-wise relevance propagation), and SHAP (Shapley-value-based global attributions). The convergence of all four methods on the same surface regions is what turns the attributions into a claim about the model's perceived sensitivity structure.","core_discovery":"The central discovery is that the FNN's learned inverse map from implied volatility surfaces to rough Heston parameters is dominated, across LIME, DeepLIFT, LRP, and SHAP, by the deep in-the-money, short-maturity region of the smile—most consistently the point (K, T) = (0.6, 0.6) on the authors' grid. The mean-reversion speed κ, the volatility-of-volatility ν, and the roughness parameter H are read almost exclusively from this left wing at short maturities; ρ is read from both wings at extreme expiries; V0 from short-maturity wings; and θ from long maturities according to LRP and DeepLIFT, though the SHAP global analysis instead highlights short-expiry deep-ITM points. The authors explain the left-wing dominance by the known result that rough Heston's short-maturity left wing is dramatically steeper than plain Heston's, so the network weights the surface region most distinctive of the rough model. They contrast this with their prior interpretability study of the plain Heston model, where short maturity mattered but no clear moneyness preference appeared.","pith_inferences":["A natural extension would be to train the same network on market-calibrated parameter distributions instead of uniform draws; the uniform prior may itself inflate the influence of the left wing.","Applying the same attribution pipeline to other rough models (e.g., rough Bergomi) would test whether deep-ITM short-maturity dominance is a universal signature of rough volatility.","If the attribution pattern persists on real SPX data, calibration algorithms should either up-weight trustworthy deep ITM quotes or use robust losses to limit the influence of illiquid noise in that region.","The paper's own observation that SHAP's global reading for θ contradicts the local methods' long-maturity picture suggests that local-versus-global disagreements may themselves carry information about parameter identifiability."],"forward_implications":["Calibration error on the short-maturity, deep in-the-money region will propagate most strongly into parameter estimates, so FNN-based calibrations of rough Heston should validate accuracy there first.","The out-of-sample failures for κ and ρ on the narrow-range network confirm that the parameter-to-surface map is not bijective; attributions for weakly identified parameters should be interpreted with caution.","The agreement among LIME, DeepLIFT, LRP, and SHAP indicates that the left-wing dominance is a stable property of the learned map rather than an artifact of one attribution algorithm.","The comparison with plain Heston implies that interpretability patterns can serve as a diagnostic: a network calibrated on data that does not exhibit rough Heston's steep left wing would be expected to show a different attribution profile."],"supporting_citations":[{"why":"Baseline interpretability study for the plain Heston model; supplies the methodology, the comparison tables, and the finding that short maturity matters but moneyness does not.","marker":"[17]"},{"why":"Provides the characteristic function and pricing scheme used to generate the synthetic rough Heston implied volatility surfaces.","marker":"[21]"},{"why":"Theoretical comparison principle showing rough Heston's short-maturity left wing is steeper than plain Heston's; used to explain the deep-ITM attribution pattern.","marker":"[32]"},{"why":"LIME, one of the four attribution methods applied to the trained network.","marker":"[43]"},{"why":"DeepLIFT, one of the four attribution methods applied to the trained network.","marker":"[47]"},{"why":"Layer-wise Relevance Propagation, one of the four attribution methods.","marker":"[5]"},{"why":"SHAP, the global Shapley-value attribution method used for the global interpretability analysis.","marker":"[37]"},{"why":"Direct neural-network calibration approach that the FNN training setup follows.","marker":"[44]"}],"fun_headline_variants":["Left wing drives neural rough Heston calibration","Rough Heston left wing steers neural estimates","DL interpretability: rough volatility read from left wing","Short-maturity left wing dominates rough Heston neural fits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The analysis treats the trained feedforward network as a faithful surrogate for the rough Heston model's true parameter-to-surface map, even though that map is probably not one-to-one and the network itself fails to predict κ and ρ out-of-sample on the narrower parameter range.","fun_headline_variants_meta":{"raw":{"variants":["Left wing drives neural rough Heston calibration","Rough Heston left wing steers neural estimates","DL interpretability: rough volatility read from left wing","Short-maturity left wing dominates rough Heston neural fits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00051,"raw_usage":{"total_tokens":2448,"prompt_tokens":880,"completion_tokens":1568,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":1506}},"tokens_in":496,"tokens_out":1568,"duration_ms":10627,"temperature":1.0,"reasoning_tokens":1506,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:18:17.580563+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same FNN architecture and data pipeline on rough Heston surfaces with H fixed at 0.5 (the non-rough Brownian limit) and compare the LIME, DeepLIFT, LRP, and SHAP attribution maps. If the short-maturity deep in-the-money region remains the top contributor, the paper's explanation that the network keys on rough Heston's distinctive left wing is wrong; if the attribution shifts toward at-the-money or spreads evenly, the claim is supported.","supporting_citations":[{"cited_title":"E l Euch and M","cited_arxiv_id":null,"evidence_quote":"Provides the characteristic function and pricing scheme used to generate the synthetic rough Heston implied volatility surfaces."},{"cited_title":"K eller-Ressel and A","cited_arxiv_id":null,"evidence_quote":"Theoretical comparison principle showing rough Heston's short-maturity left wing is steeper than plain Heston's; used to explain the deep-ITM attribution pattern."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LIME, one of the four attribution methods applied to the trained network."},{"cited_title":"S hrikumar, P","cited_arxiv_id":null,"evidence_quote":"DeepLIFT, one of the four attribution methods applied to the trained network."},{"cited_title":"B ach, A","cited_arxiv_id":null,"evidence_quote":"Layer-wise Relevance Propagation, one of the four attribution methods."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SHAP, the global Shapley-value attribution method used for the global interpretability analysis."},{"cited_title":"Volatility model calibration with neural networks a comparison between direct and indirect methods","cited_arxiv_id":"2007.03494","evidence_quote":"Direct neural-network calibration approach that the FNN training setup follows."}],"review_version":1}