142 – Frames and Other Census Issues
On Estimating Multiple-regime Threshold Autoregressive Models
Chun-Yip Yau
Chinese University of Hong Kong
Chong Man Tang
Chinese University of Hong Kong
Threshold autoregressive (TAR) models have been widely used in many areas including financial data analysis. When the number of thresholds is large, the estimation of the thresholds is often computationally infeasible. In this work we employ the Minimum Description Length (MDL) Principle to develop a criterion function to estimate the number of thresholds and the corresponding order and parameter values of the AR model in each regime. A genetic algorithm is implemented to efficiently solve this optimization problem. This can be interpreted as the "space" version of the AutoPARM of Davis, Lee and Rodriguez-Yam (2006).