Machine Learning Surrogates
Fast learned approximations to expensive pricing maps.
State the empirical or computational motivation.
Pay a training cost once, then approximate repeated model evaluations very quickly inside a controlled domain.
Fast learned approximations to expensive pricing maps.
Expose the proposed mathematical object.
Notation and units
Decimal rates and volatilities, year-fraction time and continuous compounding unless stated otherwise.
Separate evidence from modelling choice.
The proposed method is compared with an established baseline on held-out scenarios.
Parameter uncertainty and extrapolation are reported rather than hidden by one fit metric.
Production use requires independent validation, monitoring and a documented fallback.
“An unstated convention is a future reconciliation break.”— THEQUANTBATEMAN
Define a falsifiable validation target.
Active deployment area, but error control and extrapolation governance are essential.
Treat governance as part of the method.
Keep an established Frontier baseline beside the new method and define the scenario in which the fallback takes control.
Ask Bateman about this topic →