SigGate: Enhancing Recurrent Neural Networks with Signature-Based Gating Mechanisms

Hugo Inzirillo; R\'emi Genet

REVIEW 4 major objections 4 minor 25 references

SigGate: Enhancing Recurrent Neural Networks with Signature-Based Gating Mechanisms

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Path signatures in the forget and reset gates improve RNN forecasting on crypto data.

desk verdict A plausible but thinly supported RNN gating variant whose own tables contradict the abstract's claim of consistent improvement. read the letter →

arxiv 2502.09318 v1 pith:IVL52EPX submitted 2025-02-13 cs.LG

Rémi Genet , Hugo Inzirillo This is my paper · ORCID

classification cs.LG MSC 68T0760L1062M10

keywords pathsignaturesrecurrentneuralnetworksLSTMGRUgatingmechanismstimeseriesforecastingcryptocurrencylong-termdependencies

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

This paper proposes a small, targeted change to two standard recurrent architectures: replace the forget gate in an LSTM and the reset gate in a GRU with a function of the path signature of the input history, rather than a function of the current input and previous hidden state. The signature is computed on a learnable five-dimensional projection of the input, truncated to depth two to four, and time-normalized. On hourly cryptocurrency data, the authors report that the best signature-based LSTM and GRU variants reach higher $R^2$ than their standard counterparts on absolute-return and volume forecasting, with the clearest gains at longer prediction horizons. The point of the paper is that memory-retention decisions benefit from a compact geometric summary of the whole input trajectory.

What carries the argument

The load-bearing object is the truncated, time-normalized path signature $S(\tilde{X})_{0,t}$, the collection of iterated integrals of a learnably projected input path up to time $t$. For a path in $\mathbb{R}^5$, depth $M=3$ gives terms for all coordinate increments, their pairwise products, and triple products, encoding not just where the input is but how it has moved. Feeding this vector through a learned matrix and a sigmoid produces the forget or reset gate, so the memory decision is a function of the whole trajectory rather than the latest point. The projection $W_{\mathrm{sig}}$ keeps the signature dimension manageable, and time normalization by $t$ keeps gate values comparable across sequence lengths.

What would settle it

An ablation that swaps the signature gate for a plain neural-network gate with the same input projection and parameter count, trained on the same five runs, would settle the claim: if the non-signature gate matches the SigLSTM and SigGRU $R^2$ values in Tables I and III, the gains are not due to the signature structure.

Watch

Extended reading notes

Core claim

On the paper's own terms: conditioning an RNN's memory-retention decision on the geometric signature of the entire input path improves forecasting. Concretely, SigLSTM replaces $f_t = \sigma(W_f x_t + U_f h_{t-1} + b_f)$ with $f_t = \sigma(W_f S(\tilde{X})_{0,t} + b_f)$, and SigGRU does the same for the reset gate; all other gates stay standard. In the authors' experiments on Binance USDT markets, SignatureLSTM-3-2 reaches the best one-step $R^2$ of 0.1642 for absolute returns (baseline LSTM: 0.1620) and SignatureLSTM-3-3-3 the best six-step $R^2$ of 0.1324 (baseline: 0.1257), while SignatureGRU-3-3 leads volume prediction at every horizon (0.4051, 0.1210, and 0.0963 vs GRU's 0.3972, 0.0949, and 0.0807). The authors read the pattern as signatures amplifying the strengths of whichever base architecture suits the task, not overturning those strengths.

Load-bearing premise

The method assumes that removing the previous hidden state from the forget and reset gates and replacing it with the input path signature is harmless or helpful, and this removal is never tested on its own, so the reported gains could in principle come from the added parameters or the projection rather than from the signature.

Editorial extensions

If this is right

Signature gating can be added to existing LSTM and GRU code by changing only the forget or reset gate equation, leaving all other gates and training procedures intact.
Longer-horizon forecasts benefit most, because standard LSTM volume forecasts turn negative at 9 and 15 steps while SignatureGRU retains positive $R^2$.
Architecture choice remains task-dependent: LSTM families win on absolute returns, GRU families win on volume, and signatures improve the leading family in each case.
Training time grows by 2 to 6 times depending on architecture and signature depth, so the accuracy gain must be weighed against compute in deployment.
Signature depth around three gives the best accuracy-cost balance; depth four adds compute with diminishing returns.

Reading between the lines

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

The paper does not run a matched-parameter control where the signature is replaced by a plain nonlinear readout of the same projected history; such a control would isolate the signature structure as the source of gains.
The five-dimensional projection and truncation depth are hyperparameters; nothing in the method fixes them, so larger projections or learned depth selection are natural follow-ups the paper leaves open.
The same gating substitution could be applied to the update gate in a GRU or the input gate in an LSTM; the paper modifies only forget and reset gates.
Because each configuration was run five times, paired-difference tests across those runs would tell whether the reported $R^2$ gaps are statistical noise; the paper reports means and standard deviations only.

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

Colleague,

The one thing to know: this is a small but real architectural variant—swap the LSTM forget gate and GRU reset gate for a time-normalized path signature of a learned low-dim projection—and the authors have shipped code and a PyPI package. The exposition of the signature machinery and the Keras implementation is clear. They run two forecasting tasks (BTC volume and absolute returns), several horizons, multiple depths, and report means and stds over five runs. That is more than a toy.

The catch is that the headline claim does not survive their own tables. The abstract says SigLSTM and SigGRU “outperform their traditional counterparts,” but in Table I every two-layer SigGRU is below plain GRU on 1-step volatility, and several SigLSTM variants are worse; in Table III the three-layer SigLSTM collapses on 9-step volume (-0.287 vs 0.061). With stds in the 0.003–0.02 range and no significance tests, most of the “wins” are noise-level. The paper also changes two things at once: it drops the h_{t-1} dependence from the gate and adds a learned projection plus a gate matrix from signature space to hidden space. For the volume task with hidden size 100 and depth 4, the gate matrix alone is around 78k parameters, several times the entire reset gate it replaces. No ablation, no parameter-matched baseline, no attempt to restore h_{t-1}. So the favorable numbers cannot be attributed to the signature.

What is genuinely useful: the idea is plausible, the related work is honest (they cite the LSTM-with-signatures PDE paper and Rough Transformers), and the code release means a motivated reader can check. The paper reads like a workshop-level empirical note that got dressed up as a discovery. The soft spots are the load-bearing ones: the central claim is overstated and the experiments don't isolate the mechanism.

Who is this for? Someone working on signature-based sequence models or RNN gating variants might mine it for the architecture idea, but they should not cite it as evidence of consistent gains. I'd send it to a referee before desk rejecting, because the variant is new and the empirical protocol is fixable—but the referee would need to require parameter-matched baselines, ablations that keep h_{t-1}, and significance testing. As it stands, I would not accept it.

Referee Report

4 major / 4 minor

Summary. The manuscript proposes SigLSTM and SigGRU, variants of LSTM/GRU in which the forget gate (LSTM) and reset gate (GRU) are replaced by a sigmoidal function of a time-normalized path signature of a learned low-dimensional projection of the input sequence, while the other gates remain unchanged. The models are evaluated on two cryptocurrency forecasting tasks (hourly BTC absolute returns and hourly trading volumes) at several horizons, with R2 means and standard deviations over five runs, together with training times. The central claim is that signature-based gating consistently improves accuracy over standard LSTM/GRU.

Significance. The idea of using path signatures for gate control is interesting and worth exploring, and the open-source release of the implementation is a practical strength. However, the results as presented do not support the paper's central claim: the comparisons are not parameter-matched, no ablation isolates the signature mechanism, several reported configurations are worse than baselines, and no statistical test is provided. The significance of the contribution therefore cannot be assessed from the current evidence.

major comments (4)

[V, Tables I and III] The abstract and conclusion claim that the models 'consistently improve' and 'outperform their traditional counterparts,' but the paper's own tables contain multiple counterexamples. In Table I, every two-layer SigGRU variant is worse than plain GRU for 1-step volatility (0.1531-0.1589 vs. 0.1604), and SigLSTM-2-2 (0.1597) and SigLSTM-3-3 (0.1544) are worse than LSTM (0.1620). In Table III, the three-layer SigLSTM-3-3-3 is much worse than LSTM-3 at 9-step volume (-0.2870 vs. -0.0662). The 'consistently outperform' claim is therefore not supported by the reported data.
[III-B, Signature Forget Gate and Signature Reset Gate equations] The design change is confounded with a large increase in parameter count and with the removal of the recurrent term from the gate. For a projected signature dimension of 780 (depth 4, projection dimension 5) and hidden size 100, the signature gate matrix contains 78,000 parameters, whereas the original GRU reset gate has W (19x100) plus U (100x100), about 11,900 parameters for the 19-dimensional volume task. Because no parameter-matched baseline and no ablation that restores h_{t-1} is reported, the observed differences cannot be attributed to the path-signature mechanism rather than to added capacity or to dropping the recurrent dependency.
[IV.D] The training protocol states that models are evaluated five times 'to assess stability and statistical significance,' but the paper reports no significance test, confidence interval, or paired comparison. Several key differences are within one standard deviation of the baseline; for example, Table I 1-step LSTM (0.1620 +/- 0.0017) versus SigLSTM-3-2 (0.1642 +/- 0.0043) is about one standard error apart. With five runs per configuration, the claimed improvements cannot be distinguished from noise.
[V] The narrative selects favorable configurations after the fact: the text highlights SigLSTM-3-2 for 1-step volatility and SigLSTM-3-3-3 for 6-step volatility while giving less prominence to the many worse-than-baseline entries in the same tables. A systematic, multiple-comparison-aware evaluation is needed before any claim of consistent improvement can be made.

minor comments (4)

[IV.C] Configuration labels such as SignatureLSTM-3-2 and SignatureGRU-3-3-3 are never formally defined; state which number is signature depth, which is projection dimension, and which is layer count.
[IV.C and Tables I and III] The rows labeled 'Extended Signature Input Models' are not described in the model variants; clarify whether these use the expanded projection dimension and why their settings differ.
[III.A] Time normalization S(X)_{0,t}/t is undefined at t=0; state how the first time step is initialized in the implementation.
[Abstract and Tables II and IV] The claim that the approach 'maintains the computational efficiency of traditional counterparts' is contradicted by the reported training times, which are 5-6 times longer for SigLSTM and 2-3 times longer for SigGRU; please rephrase or provide a cost-benefit analysis.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity identified: the paper's claims are empirical, benchmarked against non-signature LSTM/GRU baselines, and not derived from its own definitions.

full rationale

No circularity found. The paper offers no analytical derivation; its central claim is empirical: SigLSTM and SigGRU outperform standard LSTM and GRU baselines. The signature gate equations are explicitly defined in Section III-B as f_t = sigma(W_f S(X~)_{0,t}+b_f) and r_t = sigma(W_r S(X~)_{0,t}+b_r), and the reported R2 values come from training these architectures and the baselines under a shared protocol. The self-citations are [17] (SigKAN) as prior motivation and [25] (keras-sig) as implementation tooling; neither is invoked as proof of the SigLSTM/SigGRU performance claim, so they are not load-bearing. No parameter is fitted to a subset of a target dataset and then renamed a prediction; no uniqueness theorem from prior work is imported to force the architecture; no known result is merely relabeled. Some statements in the abstract are contradicted by the authors' own Table I, and the signature gates simultaneously change both the input dependence and the parameter count, but those are correctness and attribution risks, not circularity. Since no step reduces to its own inputs by construction, the honest finding is no circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical or mathematical entities are introduced; path signatures are established prior work. The central claim rests on standard signature theory, on the adequacy of depth 2 to 4 truncation of a 5-dimensional learned projection, and on the safety of dropping the hidden state from the gate. The signature depth and projection dimension are the main hand-chosen free parameters.

free parameters (3)

signature depth M = 2, 3, 4
Chosen by hand; the paper tests several and highlights the best per task. M sets the signature dimension entering the gate, so it is a free knob of the claimed improvement.
input projection dimension = 5 and 10
The input is projected to 5 dimensions (or 10 in extended variants) before signature computation, chosen for computational tractability; W_sig is learned.
hidden size = 100
Fixed for all models for fair comparison, but it caps the capacity of the gate matrices W_f and W_r.

assumptions (4)

standard math Chen's theorem and path signature properties (iterated integrals, time normalization) are valid for the discretized input path.
Section III-A invokes signature construction after defining the input path; relies on standard rough path theory.
domain assumption The truncated, time-normalized signature of the learned projection S(X~)_{0,t} retains enough information about the input history to make gate decisions.
Section III-B replaces the forget/reset gate with this signature; no information-theoretic or empirical ablation supports the truncation depth.
domain assumption The forget/reset gate can safely discard dependence on the previous hidden state and cell state.
Section III-B equations for f_t and r_t contain no h_{t-1} or c_t term; the paper does not test this removal.
domain assumption Binance USDT data from January 2020 to December 2022 is representative for evaluating the models.
Section IV-A defines the dataset; no experiments on other domains support the abstract's claim of 'various sequential learning tasks'.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SigGate: Enhancing Recurrent Neural Networks with Signature-Based Gating Mechanisms." pith.science (2026). https://pith.science/paper/IVL52EPX

@misc{pith2026250209318,
  author       = {Pith},
  title        = {Pith review of: SigGate: Enhancing Recurrent Neural Networks with Signature-Based Gating Mechanisms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IVL52EPX}},
  note         = {Machine review of arXiv:2502.09318}
}

read the original abstract

In this paper, we propose a novel approach that enhances recurrent neural networks (RNNs) by incorporating path signatures into their gating mechanisms. Our method modifies both Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures by replacing their forget and reset gates, respectively, with learnable path signatures. These signatures, which capture the geometric features of the entire path history, provide a richer context for controlling information flow through the network's memory. This modification allows the networks to make memory decisions based on the full historical context rather than just the current input and state. Through experimental studies, we demonstrate that our Signature-LSTM (SigLSTM) and Signature-GRU (SigGRU) models outperform their traditional counterparts across various sequential learning tasks. By leveraging path signatures in recurrent architectures, this method offers new opportunities to enhance performance in time series analysis and forecasting applications.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 12 canonical work pages

[1]

R. H. Shumway, D. S. Stoffer, and D. S. Stoffer, Time series analysis and its applications . Springer, 2000, vol. 3

work page 2000
[2]

G. E. Box, G. M. Jenkins, G. C. Reinsel, and G. M. Ljung, Time series analysis: forecasting and control . John Wiley & Sons, 2015

2015
[3]

Essai d’une recherche statistique sur le t exte du roman “Eugene Onegin

A. A. Markov, “Essai d’une recherche statistique sur le t exte du roman “Eugene Onegin” illustrant la liaison des epreuve en chain ( ‘Example of a statistical investigation of the text of “Eugene Onegin" i llustrating the dependence between samples in chain’),” Izvistia Imperatorskoi Akademii Nauk (Bulletin de l’Académie Impériale des Scienc es de St.- Pé...

work page 1913
[4]

Deep learning,

Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,” nature, vol. 521, no. 7553, pp. 436–444, 2015

2015
[5]

Medsker and L

L. Medsker and L. C. Jain, Recurrent neural networks: design and applications. CRC press, 1999

1999
[6]

Time-series forecasting with deep learning: a survey,

B. Lim and S. Zohren, “Time-series forecasting with deep learning: a survey,” Philosophical Transactions of the Royal Society A , vol. 379, no. 2194, p. 20200209, 2021

2021
[7]

Tkan: Temporal kolmogorov-a rnold net- works,

R. Genet and H. Inzirillo, “Tkan: Temporal kolmogorov-a rnold net- works,” arXiv preprint arXiv:2405.07344 , 2024

arXiv 2024
[8]

Temporal fusi on transform- ers for interpretable multi-horizon time series forecasti ng,

B. Lim, S. Ö. Arık, N. Loeff, and T. Pﬁster, “Temporal fusi on transform- ers for interpretable multi-horizon time series forecasti ng,” International Journal of F orecasting, vol. 37, no. 4, pp. 1748–1764, 2021

work page 2021

Show all 25 references

[9]

A temporal kolmogorov-arnol d transformer for time series forecasting,

R. Genet and H. Inzirillo, “A temporal kolmogorov-arnol d transformer for time series forecasting,” arXiv preprint arXiv:2406.02486 , 2024

2024 arXiv
[10]

Long short-term mem ory,

S. Hochreiter and J. Schmidhuber, “Long short-term mem ory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

1997
[11]

Learning phrase representation s using rnn encoder-decoder for statistical machine translation,

K. Cho, B. V an Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning phrase representation s using rnn encoder-decoder for statistical machine translation, ” arXiv preprint arXiv:1406.1078, 2014

2014 arXiv
[12]

Integration of paths–a faithful represen tation of paths by noncommutative formal power series,

K.-T. Chen, “Integration of paths–a faithful represen tation of paths by noncommutative formal power series,” Transactions of the American Mathematical Society , vol. 89, no. 2, pp. 395–407, 1958

1958
[13]

A primer on the signatu re method in machine learning,

I. Chevyrev and A. Kormilitzin, “A primer on the signatu re method in machine learning,” arXiv preprint arXiv:1603.03788 , 2016

2016
[14]

Embedding and learning with signatures ,

A. Fermanian, “Embedding and learning with signatures ,” Computa- tional Statistics & Data Analysis , vol. 157, p. 107148, 2021

2021
[15]

Ext racting information from the signature of a ﬁnancial data stream,

L. G. Gyurkó, T. Lyons, M. Kontkowski, and J. Field, “Ext racting information from the signature of a ﬁnancial data stream,” arXiv preprint arXiv:1307.7244, 2013

2013 arXiv
[16]

Deep signature s tatistics for likelihood-free time-series models,

J. Dyer, P . W. Cannon, and S. M. Schmon, “Deep signature s tatistics for likelihood-free time-series models,” in ICML W orkshop on Invertible Neural Networks, Normalizing Flows, and Explicit Likeliho od Models , 2021

2021
[17]

Sigkan: Signature-weighte d kolmogorov- arnold networks for time series,

H. Inzirillo and R. Genet, “Sigkan: Signature-weighte d kolmogorov- arnold networks for time series,” arXiv preprint arXiv:2406.17890, 2024

2024 arXiv
[18]

Improving the gating mechanism of recurrent neural networks,

A. Gu, C. Gulcehre, T. Paine, M. Hoffman, and R. Pascanu, “Improving the gating mechanism of recurrent neural networks,” in International conference on machine learning . PMLR, 2020, pp. 3800–3809

2020
[19]

Simpliﬁed gating in long short-te rm memory (lstm) recurrent neural networks,

Y . Lu and F. M. Salem, “Simpliﬁed gating in long short-te rm memory (lstm) recurrent neural networks,” in 2017 IEEE 60th international midwest symposium on circuits and systems (MWSCAS) . IEEE, 2017, pp. 1601–1604

2017
[20]

Reﬁned gate: A simple and effective gating mechanism for recurrent units,

Z. Cheng, Y . Xu, M. Cheng, Y . Qiao, S. Pu, Y . Niu, and F. Wu, “Reﬁned gate: A simple and effective gating mechanism for recurrent units,” arXiv preprint arXiv:2002.11338, 2020

2002 arXiv
[21]

Differential equations driven by rough si gnals,

T. J. Lyons, “Differential equations driven by rough si gnals,” Revista Matemática Iberoamericana, vol. 14, no. 2, pp. 215–310, 1998

1998
[22]

Solving p ath depen- dent pdes with lstm networks and path signatures,

M. Sabate-Vidales, D. Šiška, and L. Szpruch, “Solving p ath depen- dent pdes with lstm networks and path signatures,” arXiv preprint arXiv:2011.10630, 2020

2011 arXiv
[23]

Rough transformers for continuous and efﬁcient time-series mode lling,

F. Moreno-Pino, A. Arroyo, H. Waldon, X. Dong, and Á. Car tea, “Rough transformers for continuous and efﬁcient time-series mode lling,” arXiv preprint arXiv:2403.10288, 2024

2024 arXiv
[24]

Clustering digital assets using path si gnatures: Application to portfolio construction,

H. Inzirillo, “Clustering digital assets using path si gnatures: Application to portfolio construction,” arXiv preprint arXiv:2410.23297 , 2024

2024 arXiv
[25]

Keras sig: Efﬁcient path sig nature computa- tion on gpu in keras 3,

R. Genet and H. Inzirillo, “Keras sig: Efﬁcient path sig nature computa- tion on gpu in keras 3,” arXiv preprint arXiv:2501.08455 , 2025. 6

2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

R. H. Shumway, D. S. Stoffer, and D. S. Stoffer, Time series analysis and its applications . Springer, 2000, vol. 3

work page 2000

[2] [2]

G. E. Box, G. M. Jenkins, G. C. Reinsel, and G. M. Ljung, Time series analysis: forecasting and control . John Wiley & Sons, 2015

2015

[3] [3]

Essai d’une recherche statistique sur le t exte du roman “Eugene Onegin

A. A. Markov, “Essai d’une recherche statistique sur le t exte du roman “Eugene Onegin” illustrant la liaison des epreuve en chain ( ‘Example of a statistical investigation of the text of “Eugene Onegin" i llustrating the dependence between samples in chain’),” Izvistia Imperatorskoi Akademii Nauk (Bulletin de l’Académie Impériale des Scienc es de St.- Pé...

work page 1913

[4] [4]

Deep learning,

Y . LeCun, Y . Bengio, and G. Hinton, “Deep learning,” nature, vol. 521, no. 7553, pp. 436–444, 2015

2015

[5] [5]

Medsker and L

L. Medsker and L. C. Jain, Recurrent neural networks: design and applications. CRC press, 1999

1999

[6] [6]

Time-series forecasting with deep learning: a survey,

B. Lim and S. Zohren, “Time-series forecasting with deep learning: a survey,” Philosophical Transactions of the Royal Society A , vol. 379, no. 2194, p. 20200209, 2021

2021

[7] [7]

Tkan: Temporal kolmogorov-a rnold net- works,

R. Genet and H. Inzirillo, “Tkan: Temporal kolmogorov-a rnold net- works,” arXiv preprint arXiv:2405.07344 , 2024

arXiv 2024

[8] [8]

Temporal fusi on transform- ers for interpretable multi-horizon time series forecasti ng,

B. Lim, S. Ö. Arık, N. Loeff, and T. Pﬁster, “Temporal fusi on transform- ers for interpretable multi-horizon time series forecasti ng,” International Journal of F orecasting, vol. 37, no. 4, pp. 1748–1764, 2021

work page 2021