Neural SDE

Unlike deterministic neural ODEs, Neural SDEs learn two distinct neural networks: a drift network that models the deterministic trend of the system, and a diffusion network that captures the state-dependent volatility or noise intensity. This separation allows the model to explicitly represent both the expected direction of a price trajectory and the uncertainty surrounding it, making it ideal for assets where volatility is not constant but varies with price levels or market regimes.

By integrating a Brownian motion term directly into the model architecture, Neural SDEs generate full probability distributions over future paths rather than point estimates. This provides a mathematically rigorous framework for computing risk metrics like Value at Risk (VaR) and Conditional Value at Risk (CVaR) from the model's own internal representation of randomness, without needing external post-hoc calibration.

Neural SDEs are often trained adversarially, where a discriminator network evaluates the realism of generated paths against real market data. This setup, closely related to SigCWGAN and Wasserstein GANs, forces the SDE to replicate complex stylized facts such as:

• Volatility clustering: large moves followed by large moves
• Fat tails: extreme events occurring more frequently than normal distributions predict
• Leverage effects: negative correlation between returns and future volatility

To generate long, coherent trajectories that avoid mode collapse, Neural SDEs can be conditioned on path signatures—mathematical objects that capture the sequential and geometric structure of a time series. The Signature Wasserstein GAN (SigCWGAN) architecture leverages this by computing the expected signature of generated paths and matching it to the signature of real data, ensuring that the generated sequences respect the chronological ordering and lead-lag relationships critical in market microstructure.

The initial condition of a Neural SDE is often sampled from a learned latent distribution, typically parameterized by a Variational Autoencoder (VAE) or a Normalizing Flow. This latent variable encodes the macro-context of a generated trajectory—such as whether the market is in a trending, mean-reverting, or high-volatility regime—allowing a single model to produce diverse, regime-specific synthetic data without explicit regime-switching logic.

Neural SDEs can be extended to model not just price paths but the full Limit Order Book (LOB) dynamics. By parameterizing the arrival rates of orders at different price levels as a system of coupled SDEs or by using a Hawkes process-driven diffusion, these models generate synthetic order books that replicate the queue dynamics, bid-ask bounce, and order flow autocorrelation observed in real exchanges, providing a sandbox for testing execution algorithms.

The foundational mathematical framework that Neural SDEs extend. An SDE models the evolution of a system over time with a deterministic drift component and a random diffusion component driven by Brownian motion. In finance, the classic Black-Scholes model is an SDE where asset prices follow geometric Brownian motion. Neural SDEs replace the fixed functional forms of drift and diffusion with neural networks, allowing the model to learn complex, non-linear dynamics directly from data rather than imposing rigid parametric assumptions.

The continuous-time stochastic process that serves as the source of randomness in Neural SDEs. Key properties include:

• Independent increments: Future movements are independent of the past
• Gaussian increments: Changes over any interval are normally distributed
• Continuous but nowhere differentiable paths: Captures the jagged nature of financial prices Neural SDEs use Brownian motion to inject calibrated noise into the generative process, enabling the simulation of uncertainty and tail events that deterministic models miss entirely.

The numerical scheme used to simulate sample paths from a Neural SDE on a computer. It approximates the continuous-time SDE by taking small discrete time steps. The method evaluates the neural network at each step to compute the drift (predictable direction) and diffusion (random volatility), then adds scaled Gaussian noise. This is the workhorse algorithm that makes Neural SDEs computationally tractable for training and inference, analogous to how Euler's method solves ordinary differential equations.

The learning paradigm often used to train Neural SDE generators. A discriminator network is trained to distinguish real market data from synthetic paths produced by the Neural SDE. The Neural SDE acts as the generator, learning to produce increasingly realistic trajectories through a minimax game. This adversarial pressure forces the model to capture subtle statistical properties like volatility clustering and leverage effects that maximum likelihood estimation alone might miss. The SigCWGAN architecture applies this specifically to path-valued data.

A mathematical tool from rough path theory that transforms a time series into an infinite sequence of iterated integrals. Signatures provide a unique, compact representation of the geometric shape of a path, capturing the order and interaction of events over time. In the context of Neural SDEs, path signatures are used to define discriminator features in SigCWGANs, allowing the model to compare entire trajectories holistically rather than point-by-point. This yields superior performance on long-term statistical fidelity.

A modern stochastic volatility paradigm where the volatility process is driven by a fractional Brownian motion with a Hurst parameter less than 0.5. This produces volatility paths that are rougher than standard Brownian motion, accurately matching empirical observations in high-frequency financial data. Neural SDEs can be architected to incorporate rough volatility dynamics, either by parameterizing the diffusion with a neural network that learns the roughness or by explicitly modeling the fractional noise structure.