Skip to main content
Back

Importance Sampling Lets Rough Volatility Hedging Skip Full Retraining

A new framework lets one Monte Carlo dataset train stochastic control policies across many candidate models by reweighting samples instead of resimulating trajectories.

Paper, Full Access
  • New method trains one Monte Carlo dataset once, then adapts to new model parameters by reweighting.
  • Targets fully non-Markovian control: path-dependent SDEs, rough volatility, fractional Brownian motion.
  • Explicit Radon-Nikodym weights constructed for three representative classes of controlled systems.
  • Error bounds cleanly separate Monte Carlo sampling error from model-risk error.
  • Rough volatility P&L variance drops from 6.22 to 0.54 as discretization refines from k=4 to k=6.
  • Paper on arXiv; no code repository released.

Training a stochastic control policy when the state process depends on its entire history, and when the model parameters are themselves unknown, is one of the messiest problems in quantitative finance. A new paper on arXiv proposes a way around the usual bottleneck: generate one synthetic dataset under a fixed reference law, then recover the dynamic programming operators for any target model by importance sampling.

The setup covers continuous-time stochastic control problems whose controlled states are fully non-Markovian and depend on unknown model parameters. These arise in path-dependent stochastic differential equations, rough-volatility hedging, and systems driven by fractional Brownian motion. Concretely, think of hedging an option when the volatility is driven by fractional Brownian motion with Hurst parameter near zero, or an SDE whose drift and diffusion depend on the entire past trajectory of the state.

Why the usual toolkit breaks

In the Markovian case, dynamic programming over a finite state gives you tractable Bellman recursions and, more recently, deep learning solvers. In the non-Markovian case, the value process cannot in general be reduced to a finite-dimensional deterministic equation, and naive discretization on refining time grids ends up sub-optimal while suffering from the curse of dimensionality.

The paper builds on an earlier discrete skeleton approach that projects the original continuous-time system onto a sequence of Brownian hitting times. Rather than sampling the Brownian motion on a fixed grid, you sample the times at which the Brownian motion crosses a barrier of size epsilon. As epsilon shrinks, solving the embedded backward dynamic programming equation gives you near-optimal controls for the original problem. That still leaves the numerical problem of actually solving the embedded equation, and doing so repeatedly as you update your estimate of the model parameters.