Real-Time Bayesian Forecasting with State-Space Models

Abstract

Nonlinear and/or non-Gaussian state-space models are a very general and expressive class of data-generating processes that can be useful for modeling all kinds of streaming time series data. However, this generality comes at a price. Both parameter inference and filtering are challenging when using these models, and unfortunately, real-time forecasting combines both of these tasks into one. In this talk, I provide a high-level overview of the algorithms that are available for this task, and I describe the three-way tradeoff between bias, variance and computational cost. I then introduce, in a non-technical way, the particle swarm filter, and provide an example on “pricing” options on the S&P 500 index.