Robust Sequential Online Prediction with Dynamic Ensemble of Multiple Models: A Review

This paper presents a concise introduction to a generic theoretical framework termed Bayesian Dynamic Ensemble of Multiple Models (BDEMM), which has been widely used for robust sequential online prediction with time series data. This framework has three major features: (1) it employs a model pool, rather than a single model, to capture possible statistical regularities underlying the data; (2) the model pool consists of multiple weighted candidate models, wherein the model weights are adapted online to capture possible temporal evolutions of the data; (3) the adaptation for the model weights follows Bayesian formalism. These features together define BDEMM. To make the introduction comprehensive, we describe BDEMM from five perspectives, namely the basic theories, its different forms of algorithmic implementations, its applications, its connections to related research, open resources for algorithm implementations, followed by a discussion of practical issues for applying it and some open problems that are worth further research.