Noise estimation of SDE from a single data trajectory

arXiv:2509.25484 · q-fin.ST, math.PR · Submitted 2025-09-29 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.

Ruth: I'm Ruth.

Warren: And I'm Warren, here with Iris.

Ruth: We're looking at the paper: "Noise estimation of SDE from a single data trajectory".

Warren: Here's the gist of the paper. In this paper, we propose a data-driven framework for model discovery of stochastic differential equations (SDEs) from a single trajectory, without requiring the ergodicity or stationary assumption on the underlying continuous process By combining (stochastic) Taylor expansions with Girsanov transformations, and using the drift function’s initial value as input, we construct drift estimators while simultaneously recovering the model noise This allows us to recover the underlying P Brownian motion increments Building on these estimators, we introduce the first stochastic Sparse Identification of Stochastic Differential Equation (SSISDE) algorithm, capable of identifying the governing SDE dynamics from a single observed trajectory without requiring ergodicity or stationarity Problem Formulation and Core Idea The central challenge addressed is recovering the drift and diffusion functions from a high-frequency single realization of a continuous-time stochastic process governed by the Itô SDE dXt = µ(Xt)dt + σ(Xt)dB P t Classical approaches typically rely on ergodicity or stationarity, which are often unrealistic in finance The authors address this via a data driven pathwise approach based on a single trajectory of (1) The goal is to estimate simultaneously the underlying noise vector (B P t0, BP t1, · · · B P tN), the drift vector (µ(Xt0), µ(Xt1), · · · µ(XtN)), and the diffusion vector (σ(Xt0), σ(Xt1), · · · σ(XtN)) from a single discrete data vector (X t0, X t1, …, X tN) Stochastic Identification Methodology The proposed SSISDE algorithm involves a multi-step process to recover the SDE dynamics 1.

Ruth: Recovering Q-Brownian motion: This is achieved by taking the quadratic variation of the original data process Xt, which eliminates the drift µ and randomness B P from the system, enabling an estimate of σ as follows 2.

Warren: Constructing estimator for Xt via Stratonovich SDE: The authors use a second-order expansion of the process in Stratonovich form to obtain an estimator for the diffusion process, denoted as Ψ s,t 2 3.

Ruth: Applying Girsanov transformation: This is used to transform the approximated SDE into a representation involving the martingale Brownian motion B Q 4.

Paper summary: Iris: So we've just talked about how they are trying to get the drift and diffusion from a single data path, and now I want to summarize what the paper actually claims without getting into all those technical details of their specific methods.

Ruth: Right. The core thesis of this paper is that you can discover the model for a stochastic differential equation just by looking at one trajectory. They don't require any assumptions about how the underlying continuous process behaves over time, like stationarity or ergodicity.

Warren: So if I'm thinking about trading strategies, I usually need things to be stationary so my risk metrics stay consistent. What does this paper mean for that kind of analysis?

Ruth: It means you can analyze a single realization of market data and extract the underlying dynamics, even when those traditional assumptions don't hold. They are combining stochastic Taylor expansions with Girsanov transformations to do this.

Iris: What's the big takeaway here for someone who is looking at financial time series? Is it that we can model things without making big assumptions about the market behavior?

Ruth: It suggests that even in complex systems like finance, you can uncover the governing SDE dynamics using a single path. They construct drift estimators and simultaneously recover the noise component, which is the Brownian motion increments.

Warren: So they're saying we can figure out what's driving the process just by seeing how it moves once? That sounds pretty powerful if it works reliably.

Iris: But can you really trust that identification? What variations in the data might cause a problem for their identification process?

Ruth: The paper validates this approach by testing it on both linear and quadratic drift diffusion functions, including the Black Scholes SDE, which is a classic example of something that doesn't fit those traditional assumptions. They showed they can recover drifts and diffusions that match the true ones in some cases.

Warren: And what about when the model is really complicated? Say we're dealing with highly nonlinear dynamics, not just linear or quadratic ones?

Ruth: The paper shows that even for quadratic coefficient SDEs, where the structure is more complex, they can still reliably recover the main dynamics if they use a specific method to get those initial drift estimates.

Iris: So it seems like the identification process is quite flexible in terms of the type of functions it can find?

Ruth: It does. The final algorithm they propose is a sequence: observations, then estimating noise increments, then using that to identify the symbolic form of the SDE. It’s a systematic way to go from raw data to a model.

Conclusion: Warren: So we've gone through this paper on "Noise estimation of SDE from a single data trajectory" and what they claim, and where they found these results. What's the real-world implication for us?

Ruth: The main implication is that this method gives us a way to discover the underlying dynamics of a stochastic process just from one path, without needing those strong assumptions about stationarity or ergodicity.

Iris: So if we distill it down, what is the central contribution they are making in terms of how it helps people understand these models?

Ruth: It moves the field toward data driven model discovery from single trajectories. Instead of needing long-term time series to check for statistical properties, you can use a single path to infer the governing SDE dynamics.

Warren: So, when we're talking about applying this, what's the practical limitation we need to be aware of? Where does it stop working?

Ruth: One thing they mention is that the method relies on certain nominal assumptions on the drift and diffusion coefficients, specifically Lipschitz continuity and linear growth. If those conditions aren't met for your system, you can run into issues with existence or uniqueness of the solution.

Iris: So, to wrap up what we've discussed about "Noise estimation of SDE from a single data trajectory," it’s a method that attempts to bridge the gap between complex mathematical theory and real-world data.

Warren: It does seem like an interesting technical challenge for quantitative analysis, but the result is that they can identify the governing equations even in challenging scenarios.

Florida State University · King’s College London · University of Washington · Purdue University

q-fin.ST, math.PR

Submitted: 2025-09-29

Updated: 2026-11-09

Comments: 37 pages

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 88/100

The gist: In this paper, we propose a data-driven framework for model discovery of stochastic differential equations (SDEs) from a single trajectory…

Key concepts

Stochastic Differential Equation (SDE)
An SDE describes a continuous-time stochastic process, such as stock price movement. It is mathematically defined by how the process changes over time, involving both a deterministic drift component and a random noise component that drives the randomness.
Girsanov Transformation
This mathematical tool is used to change the probability measure under which a stochastic process is viewed. In this paper, it helps transform an approximated SDE into one involving a standard Brownian motion, which simplifies the subsequent estimation of the underlying noise.
Sparse Identification of Stochastic Differential Equation (SSISDE)
This is the core algorithm proposed to identify the exact functional form of an SDE. It treats the problem as a sparse regression task, using regularization to find a model with the fewest necessary parameters, which helps in recovering the true dynamics from noisy data.
Quadratic Variation
This concept measures how much a stochastic process's path changes over time. By calculating it from observed data, researchers can eliminate the unknown drift and noise terms to get an initial estimate of the diffusion coefficient ($\sigma$) of the SDE.

Terminology

Summary

In this paper, we propose a data-driven framework for model discovery of stochastic differential equations (SDEs) from a single trajectory, without requiring the ergodicity or stationary assumption on the underlying continuous process By combining (stochastic) Taylor expansions with Girsanov transformations, and using the drift function’s initial value as input, we construct drift estimators while simultaneously recovering the model noise This allows us to recover the underlying P Brownian motion increments Building on these estimators, we introduce the first stochastic Sparse Identification of Stochastic Differential Equation (SSISDE) algorithm, capable of identifying the governing SDE dynamics from a single observed trajectory without requiring ergodicity or stationarity

Problem Formulation and Core Idea

The central challenge addressed is recovering the drift and diffusion functions from a high-frequency single realization of a continuous-time stochastic process governed by the Itô SDE dXt = µ(Xt)dt + σ(Xt)dB P t Classical approaches typically rely on ergodicity or stationarity, which are often unrealistic in finance The authors address this via a data driven pathwise approach based on a single trajectory of (1) The goal is to estimate simultaneously the underlying noise vector (B P t0, BP t1, · · · B P tN), the drift vector (µ(Xt0), µ(Xt1), · · · µ(XtN)), and the diffusion vector (σ(Xt0), σ(Xt1), · · · σ(XtN)) from a single discrete data vector (X t0, X t1, …, X tN)

Stochastic Identification Methodology

The proposed SSISDE algorithm involves a multi-step process to recover the SDE dynamics

  1. Recovering Q-Brownian motion: This is achieved by taking the quadratic variation of the original data process Xt, which eliminates the drift µ and randomness B P from the system, enabling an estimate of σ as follows

  2. Constructing estimator for Xt via Stratonovich SDE: The authors use a second-order expansion of the process in Stratonovich form to obtain an estimator for the diffusion process, denoted as Ψ s,t 2

  3. Applying Girsanov transformation: This is used to transform the approximated SDE into a representation involving the martingale Brownian motion B Q

  4. Taylor expansion and ODE solving: The drift function µ(Xt) is then recovered by comparing coefficients of iterated integrals from Step 2 and Step 4, leading to an ODE (24) for µ(Xti)

  5. SSISDE identification: Finally, a stochastic symbolic regression problem is formulated to identify the exact functional form of the SDE dXt = µ(Xt)dt + σ(Xt)dB P t

Model Selection and Sparsity Control

The identification process is framed as a sparse regression problem (28), where the target is to minimize the residual L2-norm while promoting sparsity through regularization R α,ρ(Ξµ,Ξσ) The authors employ k-fold time–series cross-validation to select optimal hyperparameters (α⋆, ρ⋆) and then use a 1-SE rule to choose the simplest model with the smallest support size They also define optimally sparse sizes ˜nµ, n˜σ based on the validation error curves δ(n)

Validation and Application

Numerical experiments validate the method across various models, including linear and quadratic drift–diffusion functions The paper demonstrates performance on the Black–Scholes SDE, which does not satisfy ergodicity or stationarity assumptions In the Black–Scholes case, using Method 2 for estimating ψ21 leads to a reconstructed drift that overlaps the true drift when compared against regression-based methods For quadratic coefficients SDEs, the method recovers functions that align with the intended polynomial structure, achieving a small mean squared error MSE in some cases

Error Analysis and Convergence

The error analysis confirms that for the Black-Scholes model, where linear constant drift and diffusion coefficient assumptions hold for all u ∈ [s, t), the approximation errors ε µ˜ = ε σ = 0 in Step 2, and ε µ/σ = 0 in Step 4 The overall error terms ε µ,σ˜ + ε µ/σ have the same order as t − s2−ε, which is small if t − s is small The proof of Proposition 5.1 shows that the total approximation error of Xt for t ∈ [0, T] has the form E[∥Xt − L 3(t)∥2] ≤ Ctk+1 - tk2, which completes the proof

Conclusion on Performance

The method is shown to be robust, as evidenced by results where the reconstructed noise process tracks the driving Brownian motion quite well, and the path simulated under the learned model is practically indistinguishable from the observed trajectory The authors conclude that their technique can capture main dynamics even when ψ21 is only approximated

Numerical Settings

The experiments utilize k = 7 contiguous folds, a logarithmic grid Λ spanning several decades for α, and P = (0.3, 0.4, 0.5, 0.7, 0.85, 0.95, 1.0) for ρ Feature blocks are built as Θµ(Xti) ∆ti and Θσ(Xti) ∆B P ti with RMS column normalization applied separately to each block on the training portion of each fold The final model size is reported alongside the CV results to document the sparsity–accuracy trade-off achieved by (28)

Estimation of Drift and Diffusion

The estimation of µ(Xt) is performed recursively using the Euler scheme based on the estimated coefficients, where b(Xti) is calculated using ψ21 computed via Method 1 or Method 2 The noise increments are then predicted using the discretization scheme (25), assuming B P starts at zero The final identification of the functions µ and σ is achieved by applying SSISDE to recover the symbolic form dXt = Θµ(Xti)Ξµdt + Θσ(Xti)ΞσdB P ti

Black-Scholes Specific Results

For the Black–Scholes model, Case 1 using Method 1 for ψ21 yields a recovered drift µb(x) = 0.000265 x + 0.499733 and diffusion σb(x) = 1. p. 00002 Case 2 using Method 2 yields µb(x) = 0. p. 85, σb(x) = 1.0002 with a small mean squared error MSE = 3. p. 8497 × 10−31 The recovered noise now overlaps the driving noise very closely, and the simulated path matches the data path almost exactly

Quadratic Coefficients Case

For the quadratic-coefficient SDE dXt = 0.7 Xt + 0.3 X squared dt + 0. p. 85 Xt + 0.1 X squared dBP t, the method recovers functions µb(x) = 0. p. 0620512 x + 0.35905 x squared and σb(x) = −0. p. 80177782 + 0.535985 x + 0.0842435 x squared with MSE = 1. p. 8448 × 10−7 This demonstrates the method continues to recover main dynamics reliably when ψ21 is obtained via regression, even in the presence of quadratic drift and diffusion structure

Final Model Summary

The final algorithm is summarized as a sequence of steps: Observations: Xti N i=1 construct ∆B Q ti N i=1 Calculate µ(Xti) N i=1 using Euler scheme Step 7: apply SSISDE to identify the two functions µ, σ This concludes our algorithm to identify the stochastic governing equation from noisy, stochastic, and single-shot time-series observations

References

[1] Y. A¨ıt-Sahalia, Maximum likelihood estimation of discretely sampled diffusions: A closed-form approximation approach, Econometrica, 70 (2002), pp. p. 0 223–262

[5] R. Azencott, Formule de taylor stochastique et d´eveloppement asymptotique d’int´egrales de feynmann, in S´eminaire de Probabilit´es XVI, LNM921, Springer, 1982, pp. 237–285

[6] P. Kidger and T. Lyons, Signatory: differentiable computations of the signature and logsignature transforms, on both cpu and gpu, arXiv preprint arXiv:2001. p. 0 p. 000706, (2020)

[8] F. Comte, V. Genon-Catalot, and Y. Rozenholc, Penalized nonparametric mean square estimation of the coefficients of diffusion processes, Bernoulli, 13 (2007), pp. p. 4 514–543

[9] T. Hastie, R. Tibshirani, J. Friedman, et al., The elements of statistical learning, 2009

[10] M. Stone, Cross-validatory choice and assessment of statistical predictions, Journal of the royal statistical society: Series B (Methodological), 36 (1974), pp. p. 8 111–133

[28] Q. Feng and X. p. 0 Zhang, Taylor expansions and castell estimates for solutions of stochastic differential equations driven by rough paths, Journal of Stochastic Analysis, 1 (2020), p. p. 0 4

[37] T. Hastie, R. Tibshirani, J. Friedman, et al., The elements of statistical learning, 2009

[41] P. Kidger and T. Lyons, Signatory: differentiable computations of the signature and logsignature transforms, on both cpu and gpu, arXiv preprint arXiv:2001.00706, (2020)

[53] M. F. Osborne, Brownian motion in the stock market, Operations research, 7 (1959), pp. p. 0 145–173

[62] M. Wanner and I. Mezic´, On higher order drift and diffusion estimates for stochastic sindy, SIAM Journal on Applied Dynamical Systems, 23 (2024), pp. 1504–1539

[65] S. Zhang and G. Lin, Robust data-driven discovery of governing physical laws with error bars, Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences, 474 (2018), p. p. 0 20180305

[63] M. Yuan and Y. Lin, Model selection and estimation in regression with grouped variables, Journal of the Royal Statistical Society Series B: Statistical Methodology, 68 (2006), pp. 49–67

[11] F. Black and M. p. 0 Scholes, The pricing of options and corporate liabilities, Journal of political economy, 81 (1973), pp. p. 0 637–654

[20] C. Cuchiero, G. Gazzani, and S. Svaluto-Ferro, Signature-based models: theory and calibration, SIAM journal on financial mathematics, 14 (2023), pp. p. 0 910–957

[36] J. Harrison and S. R. Pliska, Martingales and stochastic integrals in the theory of continuous trading, Stochastic Processes and their Applications, 11 (1981), pp. p. 1 215–260

[47] T. J. Lyons, Differential equations driven by rough signals, Revista Matem´atica Iberoamericana, 14 (1998), pp. p. 0 215–310

[32] I. V. Girsanov, On transforming a certain class of stochastic processes by absolutely continuous substitution of measures, Theory of Probability & Its Applications, 5 (1960), pp. p. 1 285–301

[46] T. Lyons and N. Victoir, Cubature on wiener space, Proceedings of the Royal Society of London. p. 0 Series A: Mathematical, Physical and Engineering Sciences, 460 (2004), pp. p. 2 169–198

[52] B. Øksendal, Stochastic differential equations, in Stochastic differential equations: an introduction with applications, Springer, 2003

[31] P. K. Friz and N. B. Victoir, Multidimensional Stochastic Processes as Rough Paths: Theory and Applications, Cambridge Studies in Advanced Mathematics, Cambridge University Press, 2010

[54] S. H. Rudy, S. L. Brunton, J. L. Proctor, and J. N. Kutz, Data-driven discovery of partial differential equations, Science advances, 3 (2017), p. p. 0 e1602614

[58] R Stanton, A nonparametric model of term structure dynamics and the market price of interest rate risk, The Journal of Finance, 52 (1997), pp. 1973–2002

[38] A. A. Kaptanoglu, C. Hansen, J. D. Lore, M. Landreman, and S. L. Brunton, Sparse regression for plasma physics, Physics of Plasmas, 30 (2023), p. p. 1

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Data-driven SDE Discovery from Single Trajectory: The system can recover drift, diffusion functions, and noise increments simultaneously from a single trajectory without requiring ergodicity or stationarity assumptions on the underlying process.

  2. Simultaneous Parameter and Noise Recovery: The SSISDE framework allows for the identification of the underlying physical Brownian motion increments alongside the governing SDE dynamics by solving a stochastic symbolic regression problem (Equation 28).

  3. Robust Estimation via Initial Value Knowledge: The algorithm relies only on knowing the knowledge of the drift coefficient at the initial time point as its major assumption, enabling estimation even when prior knowledge of the diffusion function is unavailable.

  4. Handling Non-Ergodic Models: The method can successfully identify dynamics for systems that do not satisfy ergodicity or stationarity, such as the Black-Scholes SDE (Equation 30), by utilizing Taylor expansions and Girsanov transformations to approximate the solution.

  5. Sparse Model Selection: The system implements a model selection procedure using k-fold time-series cross-validation with an elastic net penalty (Equation 28) to select the optimal sparsity level for the drift and diffusion functions, ensuring compact and stable models.

Sources

Related papers