Neural networks help find factors for derivative pricing
Learned factors can reduce variance in Monte Carlo simulations for path-dependent claims.
Neural networks can be used to construct stratification factors for derivative pricing. This method uses proportional allocation in continuous-time models.
Stratification can reduce estimator variance in Monte Carlo simulations. It works by partitioning the sample space according to a low-dimensional factor. Sampling then happens separately within the resulting strata.
The performance of stratification depends on the choice of the stratification factor. Effective low-dimensional factors are generally difficult to identify in complex continuous-time models.
The study establishes continuity results for associated conditional expectations and conditional distributions. Several counterexamples show the limits of this continuity theory.
Findings yield a universal approximation theorem for admissible stratification factors. Numerical experiments show learned factors outperform considered benchmarks in specific models.
Experiments involved a constant-maturity-swap range accrual in a two-factor Gaussian short-rate model. They also used a worst-of autocallable in a Wishart stochastic covariance model.
Key facts
- Stratification partitions the sample space using a low-dimensional factor.
- Neural networks provide a flexible approximation class for finding stratification factors.
- Learned factors outperformed benchmarks in a two-factor Gaussian short-rate model.
- The work provides a universal approximation theorem for admissible stratification factors.