Certus
Status: in development. Certus 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
Neural networks trained to reproduce an expensive pricing or risk computation can be orders of magnitude faster than the method they imitate. They are also opaque: a surrogate that is accurate on average can be badly wrong on exactly the inputs that matter, such as deep out-of-the-money options, stressed markets or parameters near the edge of the training data. “It worked on the test set” is not an acceptable standard for a number that feeds a risk report.
The approach
Certus pairs every surrogate with a way to check it. For each input it computes an a-posteriori error estimate from quantities that are cheap to evaluate: residuals of the governing equation, consistency checks such as put–call parity and no-arbitrage bounds, and the distance of the input from the training distribution. When the estimate is within tolerance, the fast answer is returned together with its bound. When it is not, Certus falls back to the reference method, so that the fast path never silently replaces a correct answer with an incorrect one.
This follows the same principle as backward error analysis in classical numerical computing: do not ask whether an answer looks right, ask how far it is from an exact answer and whether that distance is acceptable.
What it is being built to do
- Train surrogates for prices and Greeks from a reference pricer supplied by the user.
- Return a per-input error estimate with every evaluation.
- Route inputs that fail the check to the reference method automatically.
- Produce a validation report: coverage of the input space, error distribution, fallback rate.
How it fits
Certus is intended for desks and research teams that already have a trusted pricing method and want a faster one without giving up the ability to say how wrong it can be.