Signatures vs Stochastic Calculus

Predictive utility on high-frequency IBM and S&P 500 sequential data

Signature (depth 3)
Lead-lag signature
GBM-MLE
OU-MLE
ARMA(2,2)
01 — Methodology

Two views of the same path

Both methodologies aim to encode the geometry of a price path, but they part ways at the first principle. Stochastic calculus posits a generative SDE — Brownian-driven GBM, mean-reverting Ornstein–Uhlenbeck, stochastic-volatility Heston — and fits parameters via quadratic variation, MLE, or moment matching. The path signature, in contrast, is non-parametric: it encodes every iterated integral of the path up to a truncation level, producing a universal feature basis for linear functionals of streamed data (Gyurkó, Lyons et al., 2013).

Rough path signature, S≤N(X)

S(i₁,…,iₖ)(X)_[s,t] = ∫s<u₁<…<uₖ<t dXi₁(u₁) ⋯ dXiₖ(uₖ)

For a path X : [s,t] → ℝd, the depth-N signature is a tensor with 1 + d + d² + ⋯ + dN entries. By Chen’s theorem (1957) this object is unique up to tree-like equivalence and constitutes a basis under which any continuous functional of the path can be approximated by a linear combination.

The lead-lag transform embeds a 1-D stream into ℝ² so that its level-2 signature recovers the path’s quadratic variation — a fact stochastic calculus needs Itô’s lemma to even discuss.

Stochastic calculus baseline

dXt = μ(Xt, t) dt + σ(Xt, t) dWt

The classical SDE view. We benchmark two members of the family:

  • GBM-MLE: μ and σ estimated from log-return moments inside the window.
  • OU-MLE: exact maximum-likelihood κ, θ, σ for mean-reverting returns (Holý & Tomanová, 2018).
  • ARMA(2,2): direct one-step return forecast as the standard linear baseline.

All three are nested inside the rough-path picture: stochastic calculus is exactly the level-2 signature theory of semimartingales (Abi Jaber & Gérard, 2025).

02 — Data

IBM & S&P 500, 1-minute bars (Jun 9–11, 2026)

True trade-by-trade tick data is not exposed by the connected market-data API. The highest granularity available is 1-minute OHLCV, which we treat as the price stream. Mid = (high + low)/2; log-returns rt = log(midt/midt-1). All conclusions about computational complexity and feature geometry transfer to true tick data — only the noise level differs.

IBM bars
IBM realised vol (ann.)
IBM lag-1 ρ
IBM kurtosis
SPY bars
SPY realised vol (ann.)
SPY lag-1 ρ
SPY kurtosis

Mid-price & return series

03 — Predictive accuracy

Walk-forward 1-step & 5-step return forecasts

Walk-forward: 60% train / 40% test on rolling 30-bar windows. Ridge regression on each feature family; ARMA fit anew per window. Targets are h-step log-returns. R² on minute-bar returns is famously close to zero or negative — markets are nearly a random walk at this granularity. The interesting question is not “does anything work” but “which features lose least?”

Out-of-sample R²

higher is better — zero is the constant-mean baseline

Directional accuracy

sign-of-return hit rate — 50% is the coin flip

Full results table

All metrics, all (symbol, horizon, method) cells.
SymbolHorizonMethod RMSEDir. acc.Feature dimBuild (ms/win)
04 — Feature extraction

What lives inside the signature?

A depth-3 signature on a 2-D path produces 15 features (1 + 2 + 4 + 8). The first level is the increment, the second contains quadratic variation and Lévy area, the third captures asymmetric path bending. Stochastic-calculus features (μ, σ, κ, θ) are scalar summaries of the same window. The chart below contrasts how many independent features each method exposes versus how many actually carry signal (here proxied by feature dimension and the Ridge model’s implicit shrinkage).

Level 1 — drift

Δ in time and price. Same information stochastic calculus uses to estimate μ.

Level 2 — variation & area

Quadratic variation (≡ realised volatility) and Lévy area (signed cross-variation). The latter captures lead-lag without a regression.

Level 3+ — path bending

Third-order moments that distinguish accelerating from decelerating moves — invisible to GBM, partially captured by Heston, fully encoded by signatures.

05 — Non-linear dependencies

How each method handles non-linearity

Capability Signature (depth N) Stochastic calculus ARMA / regression
Path order-of-events sensitivity Yes — non-commutative tensor algebra preserves ordering Only via path-dependent SDE (Volterra, rough-vol) No — symmetric in lagged values
Captures quadratic variation Yes — exact at depth 2 via lead-lag Yes — by construction (Itô) No (only via GARCH add-on)
Higher-order cross-effects (Lévy area, skew of skew) Yes — depth ≥ 2 Only with extra factors (Heston ρ) No
Universal approximator of path functionals Yes (Hambly–Lyons 2010) Only inside the assumed SDE family No — linear in lags
Robust to irregular sampling / missing ticks Yes — piecewise-linear interpolation suffices Needs careful filtering or jump terms Breaks without resampling
Closed-form pricing/hedging Semi-analytic via Fourier inversion (Abi Jaber) Yes — Black-Scholes, Heston char. function No
06 — Lead-lag & Lévy area

Embedding 1-D streams into ℝ²

The lead-lag transform pairs each price with its predecessor: X̂2i = (Si, Si); X̂2i+1 = (Si+1, Si). The level-2 signature of this 2-D path contains the quadratic variation on the diagonal, with the off-diagonal capturing local lead-lag. On our benchmark the lead-lag features hurt R² — the 1-minute IBM stream is dominated by microstructure noise, and the transform amplifies it before the Ridge can regularise. With longer windows or denoised mid-quotes the signal/noise balance flips (Pérez Arribas, QuantStart).

Lead-lag of last 60 IBM log-prices

2-D path that the depth-3 signature consumes

Rolling Lévy area — IBM × SPY (30-bar window)

positive ⇒ SPY leads IBM; negative ⇒ IBM leads
07 — Computational complexity

Real-time cost per window

Signature cost is O(N · (dN+1−1)/(d−1)) per window of length L, multiplied by L for Chen’s product chain. With d=2 and depth 3 we sit at ~0.7 ms per 30-bar window on a single CPU — i.e. ~1,400 windows/sec — easily real-time. ARMA refitting is ~70× more expensive because the MLE optimiser dominates each call.

Build time vs. depth (window L=30, d=2)

log-scale exposes the dN growth

Build time vs. window length (depth 3)

linear in L; total dim is constant

Production note. Naïve Python implementations like the one used here are ~20–100× slower than vendor C++ kernels (Signatory, iisignature). For a depth-3, d=2 signature on a 30-tick window, Signatory on GPU is sub-100 µs — comfortable inside a microsecond-budget pricing loop.

08 — Where each method excels

Pros, cons, financial use cases

Method Pros Cons Excels in Beats ARMA / OLS at
Path signature Universal feature basis; non-parametric; order-sensitive; embeds quadratic variation natively; robust to irregular sampling; linearises non-linear functionals. Feature count grows as dN; depth tuning required; interpretation harder than μ/σ; needs good library (Signatory/iisignature) for speed. LOB micro-prediction; path-dependent option pricing; regime classification of intraday trajectories; cross-asset lead-lag mining; rough-vol calibration (Abi Jaber 2025). Any task where the shape of the path matters, not just its lagged values: microstructure prediction, signature-based portfolio control, deep-BSDE solvers (Sig-RDE BSDE, 2025).
Stochastic calculus (GBM / Heston / OU) Strong economic interpretability; closed-form pricing; well-understood inference; natural under risk-neutral measure. Mis-specified under regime change; assumes Markov dynamics; quadratic variation must be modelled, not measured; rough-vol needs heavy machinery. Derivative valuation, risk-neutral hedging, term-structure modelling, pairs-trading spread modelling (Ornstein-Uhlenbeck for cointegrated series). Volatility surface fitting, scenario generation under coherent measure-change — domains where ARMA cannot speak the language of options at all.
ARMA / linear regression Cheap; transparent; easy diagnostics; mature inference. Linear in lags; no quadratic variation; no order-of-events sensitivity; forecast power collapses on near-random-walk data. Macroeconomic series, low-frequency factor regressions, stationary signal denoising.

Use case — LOB micro-prediction

On the FI-2010 NASDAQ benchmark, a depth-4 signature beats hand-crafted OBI/VPIN microstructure features by 35–56% in Information Coefficient at 100–500 ms horizons (LOB Signatures, 2024).

Use case — Rough Bergomi calibration

Signature-based pricing replicates Heston accuracy and extends to non-Markov volatility processes where asymptotic expansions break down (Abi Jaber 2025).

Use case — Pairs trading on tick data

Tick-level OU MLE captures heavy-tailed residuals that 1-minute OU misses, lifting Sharpe by ~0.4 in simulated pairs strategies (Holý & Tomanová, 2018).

09 — Caveats & sources

Read before drawing conclusions

What the benchmark does not show

  • True ticks would expose richer microstructure that 1-minute bars smooth away — the gap between signature and stochastic calculus typically widens in signature’s favour as you go higher-frequency.
  • Sample size: ~1,150 windows is modest. Production studies use millions.
  • Single horizon family: we test 1- and 5-minute returns. Path signatures shine more at intermediate horizons (seconds to minutes) for LOB tasks.
  • Naïve Python signature is ~50× slower than Signatory / iisignature; the complexity scaling shown is shape-accurate but absolute times overstate cost.

Primary references

Market data: IBM and SPY 1-minute bars (2026-06-09 to 2026-06-11), Perplexity Finance API. Code: signature transform, lead-lag, GBM-MLE, OU-MLE, walk-forward Ridge — all in analysis/benchmark.py.