Noise estimation of SDE from a single data trajectory
summary
The gist
In this paper, we propose a data-driven framework for model discovery of stochastic differential equations (SDEs) from a single trajectory…
In short
The framework proposes a data-driven method to discover stochastic differential equations (SDEs) from just one trajectory, bypassing traditional requirements like ergodicity or stationarity. It combines Taylor expansions and Girsanov transformations to simultaneously estimate the noise process, drift function, and diffusion function. This allows for the identification of governing SDE dynamics directly from single-shot data.
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 used across episodes
This episode discusses
- Noise estimation of SDE from a single data trajectory · Paper Radio
- Derivatives pricing using signature payoffs
- A note on the group lasso and a sparse group lasso
- Signatory: differentiable computations of the signature and logsignature transforms, on both CPU and GPU
- Learning Dynamical Systems and Bifurcation via Group Sparsity
The paper
Noise estimation of SDE from a single data trajectory · Read on arXiv
Florida State University · King’s College London · University of Washington · Purdue University
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.
More episodes
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- The day in conversation for 2026-11-18
- Daily Summary for 2026-11-18
- 2601.05274-On the use of case estimate and transactional payment data in neural networks for individual loss reserving
- 2009.00868-Loss-Given-Default Modeling by Post-Last Passage Time Process
- 2606.30070-Financial Resilience Evaluation: From Conditional Expectations to Dynamic Convex Risk Measures
- 2212.12687-Measuring price impact and information content of trades in a time-varying setting
- Daily Summary for 2026-11-17
- 2401.08323-Dynamic portfolio selection under generalized disappointment aversion