Real World Simulation for Long Time Horizons with Generative Neural Networks (Webinar)
Limit management, asset/liability management, insurance, and macro investing each require simulation in real world measure for time horizons of 5y to 30y and beyond. At time horizons that long, samples cannot be drawn directly from the historical data while Monte Carlo simulation encounters well documented challenges.
In this webinar, Alexander Sokol, CompatibL Executive Chairman and Head of Quant Research, proposes the use of generative neural networks as an alternative to Monte Carlo simulation in real-world measure for long time horizons. For market risk, historical simulation is preferred by regulators and risk managers over Monte Carlo simulation because exogenous assumptions involved in selecting the Monte Carlo model SDE proved to be an important source of variability in calculation results. Generative machine learning models, which also do not require an explicit SDE, provide a natural extension of historical simulation to long time horizons and share many of its attractive properties. We describe two generative interest rate model frameworks: one for the term rates and the other for the forward rates.
To overcome the challenge of insufficient time series length compared to model horizon, we use unsupervised learning to calibrate the model to multi-currency datasets. Using historical datasets for government bond yields and swap rates, we show that incorporating data from additional currencies into the model does not result in a single model for the “average currency”. Rather, it reduces model error compared to single currency calibration while preserving unique calibration for each currency.