Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

One matrix inversion solves path-dependent statistical-arbitrage execution.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 09:23 UTC pith:4ATXSTHX

load-bearing objection The quadratic reduction is real and useful; the backtest's endpoint feature evaluation is the main thing to fix before trusting the empirical numbers. the 3 major comments →

arxiv 2606.31387 v2 pith:4ATXSTHX submitted 2026-06-30 q-fin.TR

Signature-Based Optimal Execution for Statistical Arbitrage with Path-Dependent Trading Signals

Gianmarco Morbelli , Sven Karbach , Mike Derksen This is my paper
classification q-fin.TR MSC 91G80
keywords path signaturesoptimal executionstatistical arbitragepairs tradingquadratic programmingtemporary market impactdollar neutralitysignature trading
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to show that a realistically expensive execution problem — trading a statistical-arbitrage signal while paying temporary market impact, controlling inventory, staying dollar-neutral, and liquidating by the horizon — becomes a plain finite-dimensional quadratic optimisation when both the predicted return and the trading speed are written as linear functionals of the same truncated path-signature features. The payoff is operational: all path dependence is pushed into moment tensors estimated offline, the optimal policy is one matrix inversion, and a live trade is one matrix-vector product. The authors also claim that their fitted signature policy earns a higher return on turnover than a standard z-score threshold rule on synthetic mean-reverting pairs (about 9 vs 6 basis points) and on a single two-stock historical window (about 9 vs 2 basis points). A sympathetic reader would care because it suggests path-dependent execution can be added to statistical arbitrage without a real-time dynamic-programming layer.

Core claim

The central claim is Theorem 2.9: under temporary impact with an exogenous information path, with signal α_t = Kx_t and trading speed v_t = Bx_t both linear in the truncated signature x_t, the objective combining signal reward, temporary impact, inventory risk, dollar-neutrality, and terminal liquidation collapses exactly to J(θ) = θᵀAθ + bᵀθ + c with θ = vec(B). The matrices A and b are explicit expectations of Kronecker products of signature Gram blocks, prices, and penalty matrices; A is negative semidefinite, so the problem is concave, and the unconstrained optimum is θ* = −½ A⁻¹b. The reduction is exact within the policy class: approximation enters only through truncation, moment estima

What carries the argument

The workhorse is the truncated signature x_t of a time-augmented market path Z_t — a coordinate vector containing all iterated integrals up to level N. Both alpha and the control are linear maps on this same vector, α_t = Kx_t and v_t = Bx_t, so inventory is affine in θ = vec(B) through y_t = ∫₀ᵗ x_u du. Products like vᵀΛv and (QᵀP)² then become Kronecker products of signature Gram blocks with impact and price matrices; taking expectations and integrating in time yields the deterministic quadratic-form data A, b, c. Assumption 2.3 — that the law of Z does not depend on B — keeps P_t and x_t exogenous, which is what makes every term at most quadratic in θ.

Load-bearing premise

The load-bearing premise is that the information path, including the mid-price in the dollar-neutrality term, is exogenous: trading does not feed back into prices or into the signal; if permanent or transient impact makes P_t and x_t depend on the chosen B, the objective stops being quadratic in the policy coefficients.

What would settle it

Switch on permanent price impact so the mid-price used in the dollar-neutrality penalty depends on cumulative inventory, then repeat the paper's identity check — direct path simulation of the objective versus θᵀAθ + bᵀθ + c on random θ. The two will diverge, showing the reduction is specific to temporary impact. Separately, run the same workflow on dozens of pairs and out-of-sample windows; if the signature policy's return on turnover does not systematically beat the z-score benchmark after costs, the empirical outperformance claim is not general.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Offline calibration, instant execution: once A and b are estimated from historical paths or simulations, the optimal trading matrix is one regularized linear solve, and the live trading speed is one matrix-vector product with no dynamic-programming step.
  • Classical schedules are a special case: when the features are restricted to time only, the quadratic program reduces to deterministic liquidation schedules of the classical Almgren–Chriss type, now approximated by polynomial trading speeds.
  • Constraints stay cheap: trading-speed bounds, inventory caps, and hard dollar-neutrality requirements become finite-dimensional convex constraints on θ, preserving a tractable quadratic programme.
  • Richer path features become usable: second-order coordinates such as the Lévy area encode lead-lag geometry that z-score thresholds ignore, and the same quadratic reduction still applies.
  • Reported performance: in the paper's experiments the signature policy achieves roughly 9 basis points return on turnover versus 6 for the z-score benchmark on synthetic data, and roughly 9 versus 2 basis points on a single two-stock historical window.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the exogeneity assumption survives in practice, the same reduction should extend to any alpha expressible as a linear functional of an exogenous path — for example order-flow imbalance or residual-factor signals — without re-deriving the control theory; only the moment-estimation block changes.
  • The near-null eigenvalue tail reported for the empirical curvature matrix suggests that the 'one matrix solve' story is conditional on regularization; an out-of-sample sweep across ridge values would quantify how much of the benchmark outperformance is real signal versus shrinkage luck.
  • A natural testable extension is to replace the soft dollar-neutrality and terminal-liquidation penalties with hard constraints and check whether the constrained quadratic programme retains its edge; that is the practical desk version of the same problem.
  • Because the identity J(θ) = θᵀAθ + bᵀθ + c is pathwise algebraic, it should hold for any stochastic model for the information path satisfying mild integrability; re-running the paper's identity check under a different data-generating process would be a cheap robustness probe.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper proposes a signature-based optimal execution framework for statistical arbitrage. Both the alpha signal, α_t = K x_t, and the trading speed, v_t = B x_t, are modeled as linear functionals of the truncated signature x_t of an exogenous market/information path Z_t. The central theoretical result, Theorem 2.9, asserts that the path-dependent execution objective (2.12) — signal reward, temporary impact, inventory risk, dollar neutrality, and terminal liquidation — reduces exactly to a finite-dimensional concave quadratic program J(θ) = θ^T A θ + b^T θ + c, with A, b, c given explicitly in (2.15)–(2.17), so that the optimal coefficient vector is obtained by one matrix inversion. The paper includes a numerical check of the reduction (Figure 5), synthetic experiments under a common-trend/OU-spread model, and a single Shell–BP historical window, reporting higher return on turnover than a z-score threshold benchmark (≈9 vs 6 bps synthetic; ≈9 vs 2 bps historical).

Significance. If the main theorem holds as stated, the paper makes a useful contribution: it shows that a non-Markovian execution problem, restricted to signature-linear controls, can be solved as a static quadratic program without dynamic programming or HJB equations. The strengths are the explicit tensor assembly, the closed-form OU moment blocks in Appendix C, the shuffle-algebra reformulation in Appendix B, the numerical validation of the reduction to floating-point precision, and the paper's candid statement of limitations. However, the empirical contribution is currently undermined by the endpoint feature-evaluation convention in Section C.4, which introduces look-ahead into the reported ROT comparisons. The theoretical result is sound within its stated scope, but the empirical claims need to be re-run with a causal protocol before they can be taken as evidence of out-of-sample performance.

major comments (3)
  1. [Section C.4 / Sections 3.3–3.4] The paper states in Section C.4 that 'features are evaluated at bucket endpoints in both the matrix assembly and the backtest' and that a strictly predictable live implementation would use left-endpoint features. This is not merely an implementation robustness issue: using the endpoint signature means the trading speed assigned to an execution bucket is computed from information spanning the entire bucket, including the concurrent price move. Since calibration and evaluation share this convention, the fitted policy can exploit within-bucket price information, inflating the reported ROT comparisons (≈9 vs 6 bps synthetic; ≈9 vs 2 bps Shell–BP). Please rerun the empirical workflow with strictly causal left-endpoint features and report both metrics; if the difference is material, the empirical conclusions must be revised.
  2. [Assumption 2.3 / Section 2.2] The quadratic reduction relies fundamentally on Assumption 2.3: the law of the information path Z does not depend on B, i.e. a temporary-impact setting. If trading has permanent or transient impact, or if the spread signal itself responds to the strategy's trades, then P_t and x_t become functions of B, and the terms (Q_t^T P_t)^2 and Q_t^T K x_t are no longer quadratic in θ. This is a genuine scope restriction. The paper acknowledges the temporary-impact assumption, but the title and framing should state clearly that permanent impact and any feedback of trading into the signal are outside the model, and a brief discussion of where the algebra breaks would help readers avoid over-generalizing Theorem 2.9.
  3. [Section 2.6 / Figures 6 and 8] The practical claim of a 'one matrix solve' depends on A being negative definite. Figure 6 shows that only 29 of 42 directions lie above the ridge level, and Figure 8 shows that return on turnover is materially sensitive to λ_ridge. Thus the reported solution is not a clean closed form but a regularized empirical solve whose output depends on the chosen ridge. Please report a systematic ridge-selection procedure (e.g. cross-validation) and show out-of-sample ROT across a range of λ_ridge, rather than presenting a single λ_ridge = 1e−8 solve as the canonical result.
minor comments (5)
  1. [Abstract] 'We shows' should be 'We show'.
  2. [Eq. (2.15)] The symbol for the temporary-impact matrix appears as 'eΛ' in the displayed formula; it should be '\tildeΛ' consistently with the text.
  3. [Section 3.3 vs Section C.4] Section 3.3 reports 10,000 training paths and 5,000 test paths, while Section C.4 and Figures 6–8 use 1,000 training and 1,000 test paths. Please clarify which configuration underlies Figure 1 and ensure the headline result is tied to a single, clearly described experimental setup.
  4. [Figure 5 / Section D] The text reports a maximum relative error of 2.2 × 10^−7, while the Figure 5 caption reports 1.5 × 10^−7. Reconcile the numbers.
  5. [Section 3.2] The z-score benchmark is a fixed-threshold, position-capped rule that does not solve an execution problem with impact or inventory penalties. The ROT comparison should be interpreted with this limitation stated explicitly; the benchmark is a reference rule, not a strong baselining of the execution component.

Circularity Check

0 steps flagged

No significant circularity: Theorem 2.9 is a direct algebraic reduction, and the empirical comparison uses a legitimate train/test protocol.

full rationale

The paper's central claim, Theorem 2.9, is a self-contained algebraic derivation rather than a circular one. Substituting the signature-linear control v_t = B x_t and the induced affine inventory Q_t = Q_0 + (∫_0^t x_u du)^⊤ B into the objective (2.12) mechanically produces a quadratic form in θ = vec(B); the matrices A, b, c in (2.15)-(2.17) are defined as expectations of moments of the exogenous path and the fixed signal matrix K, not in terms of the optimal policy. The proof exhibits each objective term's contribution to A, b, and c explicitly, and Appendix D checks the identity against direct pathwise simulation to floating-point precision. There is no fitted parameter renamed as a prediction: the empirical ROT comparison evaluates the fitted policy on held-out paths (5000 synthetic test paths; Nov-Dec 2025 for Shell-BP) using the same z-score alpha for both arms, so the outperformance is a simulation/backtest result, not a number forced by construction. The paper cites prior signature-execution work, but none of the cited results are from the present authors and none is load-bearing: the quadratic reduction depends only on the stated linear-feature ansatz and the exogeneity assumption. The endpoint feature convention noted in Section C.4 is a potential look-ahead/implementation concern about the empirical backtest, not a circularity of the derivation; the paper itself flags that a live implementation would use left-endpoint features, and this does not make the theorem or the benchmark comparison circular. Overall, the derivation chain is independent, explicit, and numerically verified, so no circular step is present.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The theorem itself is clean: its only inputs are the objective weights (Lambda, eta, phi, gamma), the truncation N, and the moment tensors - no hidden fitted constants enter the reduction. All hand-set quantities live in the empirical demonstration: c_alpha, Lambda, lambda_ridge, benchmark caps, and the DGP parameters. The framework introduces no new entities; the projected basis psi_t of Appendix C is a deliberate sub-model used for validation, not an addition to the ontology. The practical reach of the theorem is gated by the moment-estimation step: Figure 6 shows 13 of 42 empirical eigendirections sit below the ridge, so a large slice of the policy class is, in practice, regularization-dependent rather than data-driven.

free parameters (8)
  • c_alpha (signal scale) = 1.5 (Section 3.3); 5.02e-2 = 0.5*kappa*bsd(S) (C.5)
    Converts the dimensionless z-score into instantaneous-return units; the headline ROT comparison depends on it and its value differs across sections.
  • Lambda temporary-impact matrix = diag(1e-4, 1e-5) Section 3.3; diag(1e-1, 1e-2) Section 3.4; diag(3e-3, 3e-4) C.5
    Curvature input to A; hand-set and inconsistent across experiments; ROT results scale with it.
  • lambda_ridge = 1e-8 (Section 3.3); 0 (Section 3.4); sensitivity sweep in Figure 8
    Required because only 29 of 42 eigendirections are above the ridge (Figure 6); Figure 8 shows ROT is especially sensitive to it; the historical run uses an unregularised solve despite the paper's own recommendation.
  • eta (dollar-neutrality weight) = 0.1 (Section 3.3); 0.01 (Section 3.4 and C.5)
    Hand-set soft penalty weight in A; different across experiments.
  • gamma (terminal-liquidation weight) = 0.1 (synthetic); 1 (historical)
    Hand-set terminal inventory penalty; changes inventory trajectory and ROT.
  • Benchmark policy parameters = entry/exit z-scores 2/0; capital caps 1000 (synthetic) and GBP 500,000 (historical)
    Define the comparator; the signature policy has no equivalent cap, so the comparison is asymmetric.
  • Synthetic DGP parameters = kappa=50, sigma_M=sigma_X=0.02, rho=0.3, mu=0
    Define the test environment; the ROT advantage is demonstrated only in this environment plus one chosen historical pair.
  • Truncation level N and z-score window tau = N=2; tau=600 grid steps (Section 3.3 warm-up), 8h rolling (Section 3.4)
    Feature-set size and signal lookback are hand-set; Figure 7 shows N=1 vs N=2 materially changes objective and turnover.
axioms (6)
  • domain assumption Assumption 2.1: the information path Z admits a geometric rough-path lift; the truncated signature x_t is well defined.
    Invoked at the start of Section 2.1; standard in the signature literature and inherited from [9,17], not reproven here.
  • domain assumption Assumption 2.2: integrability of feature moments, E[int ||x_t||^2 dt] < inf and E[int ||P_t||^2 ||y_t||^2 dt] < inf.
    Used in Theorem 2.9 to exchange expectation and time integrals; a technical regularity condition not verified in the empirical pipeline, where moments are estimated from samples.
  • domain assumption Assumption 2.3: exogeneity of the information path Z with respect to the control B (temporary-impact-only setting).
    Load-bearing: with permanent or transient impact the quadratic reduction fails; see weakest_assumption.
  • domain assumption Density/universality of linear functionals on signatures in spaces of continuous path functionals.
    Justifies restricting the policy class to v_t = B x_t; cited from [17,19,25], not proven in the paper; finite-N adequacy remains an empirical matter.
  • ad hoc to paper Synthetic DGP (3.1)-(3.4): common-trend log-prices with an OU spread and correlated Brownian drivers.
    Defines the controlled benchmark environment; the ROT claim is only demonstrated under this model.
  • domain assumption For the OU-projected closed-form blocks only: the spread follows (C.4) with known kappa, sigma.
    Used to derive G_psi and G_r in Appendix C; the headline optimiser uses empirical moments instead of these closed forms.

pith-pipeline@v1.3.0-alltime-deepseek · 25394 in / 30930 out tokens · 287803 ms · 2026-08-02T09:23:37.677119+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Signature-Based Optimal Execution for Statistical Arbitrage with Path-Dependent Trading Signals." pith.science (2026). https://pith.science/paper/4ATXSTHX

@misc{pith2026260631387,
  author       = {Pith},
  title        = {Pith review of: Signature-Based Optimal Execution for Statistical Arbitrage with Path-Dependent Trading Signals},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4ATXSTHX}},
  note         = {Machine review of arXiv:2606.31387}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We develop a signature-based framework for optimal execution in statistical arbitrage strategies with path-dependent predictive signals. Both the alpha process and the trading speed are modelled as linear functionals of the truncated signature of a time-augmented market path, placing signal generation and execution on the same truncated signature basis. This allows the trading rule to react to the realised history of the signal while accounting for temporary impact, inventory exposure, terminal liquidation, and approximate dollar neutrality. The main contribution is a quadratic reduction theorem: within the class of signature-linear trading speeds, the restricted path-dependent execution problem becomes a finite-dimensional concave quadratic programme in the policy coefficients. After running synthetic experiments under a mean-reverting log-spread model, we find that the fitted policy achieves a higher return on turnover than a classical $z$-score threshold benchmark. We show how the same workflow can be deployed on a historical equity pairs-trading backtest, where the fitted signature policy again outperforms the benchmark in accounting terms.

Figures

Figures reproduced from arXiv: 2606.31387 by Gianmarco Morbelli, Mike Derksen, Sven Karbach.

Figure 1
Figure 1. Figure 1: Synthetic benchmark against a classical z-score pairs-trading rule. The two assets follow the common-trend log-spread model of Section 3.1. We display the log spread, the z-score signal, price-scaled trading speed and inventory paths, and cumulative PnL for the signature strategy and the benchmark of Section 3.2. The plotted values are averages over 5000 test paths after a 600-step warm-up window used to e… view at source ↗
Figure 2
Figure 2. Figure 2: The historical Shell–BP deployment backtest on a single out-of-sample test window. The panels show the log spread, the z-score signal, price-scaled trading and inventory exposure in the two legs, and cumulative PnL of the signature strategy against the z-score benchmark of Section 3.2. The signature policy is calibrated on past fixed-length trading windows and then applied on the test window. The figure il… view at source ↗
Figure 3
Figure 3. Figure 3: Convergence of the empirically estimated Gram matrices Gψ and Gr to their closed-form OU moment targets, evaluated on the simulator time grid so that the comparison isolates Monte Carlo sampling error. The error decays at the O(M−1/2 ) Monte Carlo reference rate (grey), consistent with consistent estimation of the OU-projected blocks; imposing the spread model removes this sampling noise for those blocks. … view at source ↗
Figure 7
Figure 7. Figure 7: Effect of signature truncation order N. Out-of￾sample per-path distributions (boxes: 25–75%, whiskers: 5– 95%, diamonds: means) of the reduced objective, terminal wealth, terminal inventory norm, and turnover at N = 1 versus N = 2. Adding the second-order block (time ordering, L´evy area) shifts the reduced-objective and terminal-wealth accounting distributions upward, while the turnover and inventory dist… view at source ↗
Figure 5
Figure 5. Figure 5: Quadratic-reduction check. The matrix-form value θ ⊤Aθ + b ⊤θ + c versus the directly path-simulated objective Jsim(θ), for the optimiser θ ∗ (star) and 48 random coefficient vectors. All points lie on the identity line (maximum relative error 1.5 × 10−7 ), checking the matrix implementation of Theorem 2.9 numerically [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Left: sorted signed eigenvalues of −Aˆ (positive semidefinite under concavity) on a signed-log scale, with zero and the ridge level λridge marked. Plotting signed eigenvalues keeps any concavity violation visible. Right: heatmap of Aˆ showing its Kronecker block structure from signature Gram matrices and execution penalties. therefore makes the linear solve well posed precisely in those near-null signature… view at source ↗
Figure 8
Figure 8. Figure 8: reports the resulting mean accounting re￾turn on turnover under the canonical synthetic workflow. The purpose of this diagnostic is not to optimise these parameters, but to identify which modelling parameters materially affect out-of-sample accounting metrics. The model is especially sensitive to the ridge parameter, which controls the use of near-null directions in the empirical curvature matrix: too litt… view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Expected signatures via partial integration, coordinate change and symmetrization

    math.PR 2026-07 conditional novelty 8.0

    A coordinate transform on the tensor algebra lets expected signatures of augmented rough processes be computed by deterministic integration of Gaussian/polynomial correlators, dramatically reducing numerical cost.

Reference graph

Works this paper leans on

28 extracted references · 4 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Gatev, W

    E. Gatev, W. N. Goetzmann, K. G. Rouwenhorst, Pairs trading: Performance of a relative-value arbitrage rule, The Review of Financial Studies 19 (3) (2006) 797–827. doi:10.1093/rfs/hhj020

    doi:10.1093/rfs/hhj020 2006

  2. [2]

    Avellaneda, J.-H

    M. Avellaneda, J.-H. Lee, Statistical arbitrage in the US equities market, Quantitative Finance 10 (7) (2010) 761–782. doi:10.1080/14697680903124632

    doi:10.1080/14697680903124632 2010

  3. [3]

    Almgren, N

    R. Almgren, N. Chriss, Optimal execution of portfolio transactions, Journal of Risk 3 (2) (2000) 5–39. doi: 10.21314/JOR.2001.041

    doi:10.21314/jor.2001.041 2000

  4. [4]

    Cartea, S

    ´A. Cartea, S. Jaimungal, J. Penalva, Algorithmic and High- Frequency Trading, Cambridge University Press, 2015

    2015

  5. [5]

    Lorenz, A

    C. Lorenz, A. Schied, Drift dependence of optimal trade execution strategies under transient price impact, Finance and Stochastics 17 (4) (2013) 743–770. doi:10.1007/s00780- 013-0211-x

    doi:10.1007/s00780- 2013

  6. [6]

    Curato, J

    G. Curato, J. Gatheral, F. Lillo, Optimal execution with non-linear transient market impact, Quantitative Finance 17 (1) (2017) 41–54.doi:10.1080/14697688.2016.1181274

    arXiv 2017

  7. [7]

    P. N. Kolm, J. Turiel, N. Westray, Deep order flow imbalance: Extracting alpha at multiple horizons from the limit order book, Mathematical Finance 33 (4) (2023) 1044–1081. doi: 10.1111/mafi.12413

    doi:10.1111/mafi.12413 2023

  8. [8]

    L. G. Gyurk´ o, T. Lyons, M. Kontkowski, J. Field, Extracting information from the signature of a financial data stream, arXiv preprint arXiv:1307.7244arXiv:1307.7244

    Pith/arXiv arXiv

  9. [9]

    Kalsi, T

    J. Kalsi, T. Lyons, I. Perez Arribas, Optimal execution with rough path signatures, SIAM Journal on Financial Mathe- matics 11 (2) (2020) 470–493.doi:10.1137/19M1259778

    doi:10.1137/19m1259778 2020

  10. [10]

    Cartea, I

    ´A. Cartea, I. Perez Arribas, L. S´ anchez-Betancourt, Double- execution strategies using path signatures, SIAM Journal on Financial Mathematics 13 (4) (2022) 1379–1417. doi: 10.1137/21M1456467

    doi:10.1137/21m1456467 2022

  11. [11]

    Futter, B

    O. Futter, B. Horvath, M. Wiese, Signature trading: A path- dependent extension of the mean-variance framework with exogenous signals (2023).arXiv:2308.15135. URLhttps://arxiv.org/abs/2308.15135

    Pith/arXiv arXiv 2023

  12. [12]

    Buehler, B

    H. Buehler, B. Horvath, T. Lyons, I. Perez Arribas, B. Wood, Generating financial markets with signatures, SSRN Elec- tronic Journaldoi:10.2139/ssrn.3657366

    doi:10.2139/ssrn.3657366

  13. [13]

    Kidger, J

    P. Kidger, J. Morrill, J. Foster, T. Lyons, Neural controlled differential equations for irregular time series, in: Advances in Neural Information Processing Systems, Vol. 33, 2020, pp. 6696–6707. URL https://proceedings.neurips.cc/paper_files/ paper/2020/file/4a5876b450b45371f6cfe5047ac8cd45- Paper.pdf

    2020

  14. [14]

    Bonnier, P

    P. Bonnier, P. Kidger, I. P. Arribas, C. Salvi, T. Lyons, Deep signature transforms, Curran Associates Inc., Red Hook, NY, USA, 2019

    2019

  15. [15]

    C.-I. Lu, J. Sester, Generative model for financial time series trained with mmd using a signature kernel, arXiv preprint arXiv:2407.19848arXiv:2407.19848

    arXiv

  16. [16]

    Manten, C

    G. Manten, C. Casolo, E. Ferrucci, S. W. Mogensen, C. Salvi, N. Kilbertus, Signature kernel conditional independence tests in causal discovery for stochastic processes (2025). arXiv:2402.18477. URLhttps://arxiv.org/abs/2402.18477

    Pith/arXiv arXiv 2025

  17. [17]

    T. J. Lyons, M. Caruana, T. L´ evy, Differential Equations Driven by Rough Paths, Vol. 1908 of Lecture Notes in Mathematics, Springer, 2007. doi:10.1007/978-3-540- 71285-5

    doi:10.1007/978-3-540- 1908

  18. [18]

    Lyons, Differential equations driven by rough signals, Revista Matem´ atica Iberoamericana 14 (2) (1998) 215–310

    T. Lyons, Differential equations driven by rough signals, Revista Matem´ atica Iberoamericana 14 (2) (1998) 215–310

    1998

  19. [19]

    Chevyrev, A

    I. Chevyrev, A. Kormilitzin, A Primer on the Signature Method in Machine Learning, Springer Nature Switzerland, 2025, p. 3–64.doi:10.1007/978-3-031-97239-3_1. URLhttp://dx.doi.org/10.1007/978-3-031-97239-3_1

    doi:10.1007/978-3-031-97239-3_1 2025

  20. [20]

    Levin, T

    D. Levin, T. Lyons, H. Ni, Learning from the past, predicting the statistics for the future, learning an evolving system, arXiv preprint arXiv:1309.0260arXiv:1309.0260

    Pith/arXiv arXiv

  21. [21]

    Fermanian, Embedding and learning with signatures, Computational Statistics & Data Analysis 157 (2021) 107148

    A. Fermanian, Embedding and learning with signatures, Computational Statistics & Data Analysis 157 (2021) 107148. doi:10.1016/j.csda.2020.107148

    arXiv 2021

  22. [22]

    Krauss, Statistical arbitrage pairs trading strategies: Review and outlook, Journal of Economic Surveys 31 (2) (2017) 513–545.doi:10.1111/joes.12153

    C. Krauss, Statistical arbitrage pairs trading strategies: Review and outlook, Journal of Economic Surveys 31 (2) (2017) 513–545.doi:10.1111/joes.12153

    doi:10.1111/joes.12153 2017

  23. [23]

    J. F. Caldeira, G. V. Moura, Selection of a portfolio of pairs based on cointegration: A statistical arbitrage strategy, SSRN Electronic Journaldoi:10.2139/ssrn.2196391

    doi:10.2139/ssrn.2196391

  24. [24]

    Z. Guo, H. Jin, J. Kuang, Z. Qian, J. Wang, Signature decomposition method applying to pair trading, Journal of Futures Markets 46 (3) (2026) 582–603. doi:https: //doi.org/10.1002/fut.70075

    doi:10.1002/fut.70075 2026

  25. [25]

    Cuchiero, P

    C. Cuchiero, P. Schmocker, J. Teichmann, Global universal approximation of functional input maps on weighted spaces, Constr. Approx. 63 (2) (2026) 537–612. doi:10.1007/ s00365-025-09726-3

    2026

  26. [26]

    Vidyamurthy, Pairs Trading: Quantitative Methods and Analysis, John Wiley & Sons, 2004

    G. Vidyamurthy, Pairs Trading: Quantitative Methods and Analysis, John Wiley & Sons, 2004

    2004

  27. [27]

    Sirignano, R

    J. Sirignano, R. Cont, Universal features of price formation in financial markets: Perspectives from deep learning, Quan- titative Finance 19 (9) (2019) 1449–1459. doi:10.1080/ 14697688.2019.1622295

    arXiv 2019

  28. [28]

    Z T 0 B, S0,t(Z) dt−γ DX i F i 2 , S0,T (Z) E# =E

    R. J. Elliott, J. van der Hoek, W. P. Malcolm, Pairs trading, Quantitative Finance 5 (3) (2005) 271–276. doi: 10.1080/14697680500149370. correspondence:g.morbelli@uva.nlPage 12 of 21 Signature-Based Optimal Execution for Statistical Arbitrage A. Explicit Tensor Computation This appendix provides detailed algebraic expansion underlying the proof of Theorem...

    doi:10.1080/14697680500149370 2005