Hedging with memory: shallow and deep learning with signatures

Eduardo Abi Jaber; Louis-Amand G\'erard

REVIEW 3 major objections 3 minor 1 cited by

Hedging with memory: shallow and deep learning with signatures

T0 review · 3 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Path signatures let small neural networks hedge exotic derivatives better than LSTM baselines, using far less compute.

desk verdict Solid abstract with a plausible method; the real test is whether the SVM calibration carries over to the true non-Markovian dynamics. read the letter →

arxiv 2508.02759 v1 pith:NS6ZFFJ7 submitted 2025-08-03 stat.ML cs.LGq-fin.CPq-fin.MF

Eduardo Abi Jaber , Louis-Amand Gérard This is my paper · ORCID

classification stat.MLcs.LGq-fin.CPq-fin.MF MSC 91G2068T07

keywords pathsignaturesexoticderivativeshedgingnon-MarkovianstochasticvolatilitydeeplearningsignaturemodelexpectedLSTMbaseline

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that path signatures—a structured mathematical summary of an entire price path—are an effective and cheap feature set for learning hedging strategies for exotic derivatives under non-Markovian stochastic volatility models. In the deep setting, feedforward networks fed with signature features beat LSTM networks in most tested cases while needing orders of magnitude less training compute. In the shallow setting, one regression approach learns directly from the expected signature of the price process, and another calibrates a signature volatility model to the expected signature of volatility and then solves the hedging problem inside that calibrated model. The second approach is claimed to give more accurate and stable hedges across different payoffs and volatility dynamics. If these claims hold, hedging under path-dependent volatility becomes tractable without recurrent networks or heavy sequential training.

What carries the argument

The key machinery is the path signature, a graded sequence of iterated integrals that captures the order and nonlinear memory of a path in a coordinate-free way; the expected signature is this signature averaged over a model's probability distribution and serves as a calibration target. The signature volatility model is a parametric model of the volatility process whose expected signature can be matched to data. The argument works by converting non-Markovian path dependence into a finite truncated signature feature vector, allowing feedforward networks to approximate hedging strategies that LSTM networks would otherwise learn slowly.

What would settle it

Construct two non-Markovian volatility dynamics that have the same expected signature up to the truncation order used in calibration but different path-dependent option deltas for a given exotic payoff; if hedging strategies derived from the calibrated signature volatility model perform materially differently on the two dynamics, the sufficiency of expected-signature calibration is refuted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that path signatures serve as a superior input representation for learning hedging strategies in non-Markovian models. Feedforward neural networks using signature features outperform LSTM baselines in most test cases, with orders of magnitude less training compute. Additionally, in the shallow setting, calibrating a signature volatility model to the expected signature of volatility and then solving the hedging problem within that model yields more accurate and stable hedging results across different payoffs and volatility dynamics than directly regressing the hedging strategy on the expected signature of the price process.

Load-bearing premise

The load-bearing premise is that the expected signature of the volatility process captures the memory structure needed for hedging: if two volatility dynamics share the same expected signature but require different hedges, the calibrated model's strategies will not generalize.

Editorial extensions

If this is right

Feedforward networks with signature features can replace recurrent architectures in derivative hedging, cutting training compute by orders of magnitude.
The expected signature of the price process suffices as a direct regression input for hedging strategies in non-Markovian models.
A signature volatility model calibrated on the expected signature of volatility provides a tractable proxy model in which hedging problems can be solved more accurately and stably.
The approach extends across different payoffs and volatility dynamics without re-engineering the feature representation.
Hedging under non-Markovian dynamics becomes a static feature-engineering problem rather than a sequential modeling problem.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

We infer that expected-signature calibration could serve as a universal statistic for other path-dependent pricing and risk-management tasks, reducing model risk to moment matching in the signature sense.
The compute advantage suggests signature features may also help in high-frequency trading or portfolio optimization where LSTM training is prohibitive; this extrapolation goes beyond the paper.
A testable extension is to compare signature-FFN hedges against LSTM on out-of-sample paths with longer memory or heavier tails to map the boundary of the 'most cases' claim.
Because truncated signatures are finite-dimensional, the approach might admit analytic error bounds for hedging strategies in terms of truncation order, though the paper does not state this.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

What you should know: this is a credible ML-finance paper, but the public text is only the abstract, so every strong claim is unverifiable from the record. The core idea is sensible: use path signatures as features in feedforward nets, and compare against LSTMs for hedging exotics under non-Markovian volatility. The reported win of 'orders of magnitude less training compute' is the kind of concrete advantage that makes the paper worth reading, if it holds.

What it does well: it connects two active strands — signature methods in finance and deep hedgers — and does so with a clear shallow-versus-deep framing. The second shallow approach, fitting a signature volatility model to the expected signature of volatility and then solving the hedging problem inside that model, is methodologically interesting. If the calibrated model actually captures the relevant conditional memory, that is a genuine advance over brute-force LSTM baselines.

The soft spots, in proportion: first, the abstract gives no quantitative evidence — no numbers, no error bars, no payoff families, no dataset descriptions, no in-sample/out-of-sample distinction. That is normal for an abstract, but it means the empirical claims are unaudited. Second, the stress-test concern is on target: matching unconditional expected signatures does not automatically give you the conditional law of future volatility given the path, which is what hedging depends on. The paper might well contain a transfer argument, a continuity estimate, or a misspecification analysis — or it might not. The abstract does not say. That is a load-bearing point and should be the first question a referee asks. Third, the LSTM comparison is only as good as the baseline tuning; an apples-to-apples compute budget matters, and again we cannot check it here.

All of this is abstract-only friction, not evidence of a flaw. The paper deserves a serious referee. The recommendation: send it to peer review, and make sure the reviewers push on the calibration-to-hedging transfer condition and on the empirical setup.

For your own use: I'd want to read the full paper before citing it, but the idea is worth tracking.

Referee Report

3 major / 3 minor

Summary. The abstract of arXiv:2508.02759 proposes two machine-learning approaches for hedging exotic derivatives under non-Markovian stochastic volatility models. In the deep-learning setting, path signatures are used as features in feedforward neural networks and are claimed to outperform LSTMs in most tested cases while using orders of magnitude less training compute. In the shallow-learning setting, the authors compare a direct regression of the hedging strategy on the expected signature of the price process with an approach that first calibrates a signature volatility model (SVM) to the expected signature of the volatility process and then solves the hedging problem inside that calibrated SVM. The abstract claims that the SVM-calibration approach yields more accurate and stable hedging results across different payoffs and volatility dynamics.

Significance. If the claims in the abstract hold, the paper would provide a practically valuable and computationally efficient alternative to LSTM-based hedging for non-Markovian volatility models, with a clear methodological contribution in using expected signatures to calibrate a tractable hedging model. The deep-learning result, if reproducible, would be a significant efficiency gain. The shallow-learning contribution has conceptual appeal because it avoids simulation-intensive training. However, the strength of these contributions cannot be assessed from the abstract alone, which contains no quantitative results, no specification of the volatility models or payoff families, no error bars, and no in-sample/out-of-sample distinction. The formal significance evaluation therefore requires the full manuscript.

major comments (3)

[Abstract (second shallow-learning approach)] Please note: this comment is based on the abstract alone; the full text may already contain such a proof or misspecification analysis, but it is not visible from the submitted material.
[Abstract (deep-learning comparison)] The empirical claim that signature-based feedforward networks "outperform LSTMs in most cases, with orders of magnitude less training compute" is asserted without any quantitative detail. The abstract does not specify the datasets, the payoff families, the stochastic volatility models, the network architectures, the hyperparameter choices, the error metric, or whether the evaluation is in-sample or out-of-sample. As written, this claim is not falsifiable from the abstract. Given that the abstract presents this as a headline result, the authors should provide a quantitative summary in the abstract (e.g., median error reduction, number of cases, compute ratio) or in the full text at least; in a revised version, a few numbers would make the claim assessable.
[Abstract (general reproducibility)] The abstract reports no error bars or statistical significance for the comparative claims. Since the LSTM comparison may be sensitive to baseline tuning and the SVM calibration may be sensitive to the choice of truncation level or optimization procedure, the lack of any uncertainty quantification makes it impossible to distinguish a robust improvement from a single favorable run. The full paper should include confidence intervals or repeated-seed experiments for the key comparisons; without this, the strength of the stated claims is disproportionate to the evidence presented in the abstract.

minor comments (3)

[Abstract] The phrase "orders of magnitude less training compute" is vague; please specify whether this refers to wall-clock time, FLOPs, or GPU-hours, and on what hardware the comparison was made.
[Abstract] The term "expected signature of the volatility" is used without defining which norm or topology is intended; a brief clarification or a reference would help readers who are not specialists in signature methods.
[Abstract] The phrase "more accurate and stable" is ambiguous: "stable" could mean low variance across random seeds, robustness to model misspecification, or small hedging error variation across paths. The full text should define the stability metric explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity demonstrable from the abstract; downstream hedging evaluation does not reduce to calibration moments on the available evidence.

full rationale

This is an abstract-only review, so no equations, fitted-parameter definitions, or self-citation chains are available to inspect. The abstract's second shallow approach calibrates a signature volatility model on the expected signature of volatility and then solves the hedging problem in that calibrated model. For circularity to be established, one would need to show that the reported hedging accuracy is the same quantity used for calibration, or that the 'prediction' is forced by construction. The abstract does not state the evaluation metric, so that reduction is not exhibited. The phrase 'more accurate and stable results across different payoffs and volatility dynamics' refers to a downstream hedging task, which is not definitionally identical to matching expected signature moments. Similarly, the LSTM comparison is an empirical benchmark, not a circular step. Under the hard rules requiring quoted evidence and a specific reduction, no circular step can be identified. The appropriate finding is therefore no significant circularity, score 0.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

Because only the abstract is available, the ledger is minimal. The main implied assumption is that signatures and expected signatures contain enough information for hedging under non-Markovian volatility. The signature volatility model must have fitted parameters, whose values are not reported. No invented entities are declared in the abstract.

free parameters (1)

Signature volatility model parameters = [not reported in abstract]
The abstract states that the volatility dynamics are modeled with a signature volatility model calibrated on the expected signature of volatility, so it contains parameters fitted to data. The abstract does not report their values.

assumptions (3)

domain assumption Path signatures are well-defined and capture the pathwise memory of price and volatility processes.
The method relies on signatures as meaningful features for non-Markovian dynamics; the abstract presupposes this.
domain assumption The expected signature of the volatility process is a sufficient calibration target for producing accurate hedging strategies.
The second shallow approach calibrates a signature volatility model to the expected signature and then solves the hedging problem inside that model; this requires the expected signature to carry the information relevant to hedging.
domain assumption The tested volatility models and payoffs are representative enough to support the claim of more accurate and stable results 'across different payoffs and volatility dynamics'.
The abstract makes a general claim from presumably a finite set of experiments; the full text is needed to verify that the test set is not cherry-picked.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hedging with memory: shallow and deep learning with signatures." pith.science (2026). https://pith.science/paper/NS6ZFFJ7

@misc{pith2026250802759,
  author       = {Pith},
  title        = {Pith review of: Hedging with memory: shallow and deep learning with signatures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NS6ZFFJ7}},
  note         = {Machine review of arXiv:2508.02759}
}

read the original abstract

We investigate the use of path signatures in a machine learning context for hedging exotic derivatives under non-Markovian stochastic volatility models. In a deep learning setting, we use signatures as features in feedforward neural networks and show that they outperform LSTMs in most cases, with orders of magnitude less training compute. In a shallow learning setting, we compare two regression approaches: the first directly learns the hedging strategy from the expected signature of the price process; the second models the dynamics of volatility using a signature volatility model, calibrated on the expected signature of the volatility. Solving the hedging problem in the calibrated signature volatility model yields more accurate and stable results across different payoffs and volatility dynamics.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Stochastic control with signatures via Riccati equations on the tensor algebra
math.OC 2026-07 accept novelty 7.0 of 10

Non-Markovian path-dependent control problems with signature rewards admit optimal feedback controls and value processes as local linear signature expansions whose coefficients solve infinite-dimensional Riccati equat...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.