Neural ODE Velocity Field & Coupling Configuration
Continuous Probability Flow ODE: $x_t \in [p_0, p_1]$
Displacement paths from standard Gaussian prior $p_0$ ($t=0$) to target empirical density $p_1$ ($t=1$).
Velocity Vector Field $v_\theta(x, t)$
Instantaneous phase direction $\frac{dx}{dt}$ across state-time space.
$$\partial_t p_t(x) + \nabla \cdot (p_t(x) v_t(x)) = 0, \quad \psi_t(x_0 \mid x_1) = (1 - t) x_0 + t x_1$$
Synthesized Geometric Asset Price Paths $S(t)$
Generated multi-path asset trajectories using continuous flow field.
Terminal Log-Return Distribution $p_1(r)$
Generated sample histogram vs Empirical heavy-tailed density.
Stylized Facts Fidelity Radar
Comparison of Flow Matching Synthetic vs Empirical Ground Truth.
Microstructure & Distribution Metrics
Statistical tests verifying preservation of non-linear financial phenomena.
| Quantitative Stylized Fact | Flow Matching | Empirical Target | Status |
|---|
Synthetic Limit Order Book (LOB) Depth Profile
Multi-level order queue depth generated via continuous flow field with bid-ask spread replenishment.
Generative Paradigm Arena: Flow Matching vs Diffusion (DDPM / DDIM)
Why Flow Matching solves generative finance: straight Optimal Transport paths vs curved Brownian drift.
Straight displacement velocity field $v(x_t, t) = x_1 - x_0$. Zero trajectory intersections, deterministic RK4 solver.
Curved score-matching drift $\nabla_x \log p_t(x)$ requiring hundreds of noise-injection reverse steps.