Operator

Status: in development. Operator is a new Numerical Method product that has not been released yet. This page describes what it is being built to do. There is no download, price or release date yet; to follow its progress or to discuss early access, contact us.

The problem

Pricing under realistic models and calibrating those models to market quotes means solving partial differential equations or running Monte Carlo simulations thousands of times, each time with slightly different parameters. Classical solvers come with convergence guarantees but are slow when used inside an optimizer. Learned solvers are fast but come with no guarantees.

The approach

Operator uses machine learning to accelerate classical methods rather than replace them. Two techniques are central:

• Neural operators that learn the map from model parameters to solution, used to give the classical solver an excellent starting point or to screen candidate parameters during calibration;
• Learned preconditioners that make iterative linear solvers converge in fewer iterations while the solver itself, and therefore its accuracy guarantee, stays unchanged.

In both cases the classical method remains the ground truth. A poor learned component can make Operator slower, but it cannot make the final answer wrong.

What it is being built to do

• Neural-operator surrogates for parameter-to-price maps of common pricing PDEs.
• Learned preconditioners for the sparse linear systems that arise in finite-difference and finite-element pricing.
• Calibration workflows for stochastic volatility models that use the learned components to search and the classical solver to confirm.
• Benchmarks against the plain classical solver on the user’s own problems, reported honestly, including the cases where learning does not help.

How it fits

Operator builds on the differential-equation, optimization and linear algebra methods in SuanShu, and is aimed at quantitative teams for whom calibration time is a bottleneck.