A0632
Title: Recent advances in fast estimation of high-dimensional hidden Markov models
Authors: Jordan Rodu - University of Virginia (United States)
Xiaoyuan Ma - University of Virginia (United States)
Jordan Rodu - University of Virginia (United States) [presenting]
Abstract: Hidden Markov models (HMMs) are powerful models commonly used for modelling time series. HMMs are typically estimated using a special case of the E-M algorithm called the Baum-Welch algorithm. However, the Baum-Welch algorithm can be slow, is prone to finding local optima, and can struggle (or even outright fail) with high-dimensional data. An alternative approach to estimating HMMs, called spectral estimation of HMMs (sHMM), which overcomes these challenges, is discussed. While sHMM is ideal for real-time, high-dimensional scenarios where fast estimation and adaptation are crucial, it can suffer from instability in some cases. Recent developments that address this issue are highlighted, making sHMM a crucial tool in the time series toolbox. Examples from high-frequency trading are provided.
