Bayesian Statistics Seminar
North Carolina State University
presents
Dr. John Aston
Institute of Statistical Science, Academia Sinica
"Change-point Distributions in Hidden Markov Models"
ABSTRACT
The locations and distributions of explicit or implicit change-points are of fundamental interest in statistics as well as in numerous applications. One approach to modelling change-points is through the use of Hidden Markov models (HMMs) and extended versions such as Markov switching models. However, previous work has relied on the conditional sampling of the state sequence in order to build empirical change-point distributions from these models. This talk will develop a new method to assess change-points based on the theory of runs and patterns in HMMs (Aston and Martin, 2007). It will be shown that the distributions of where change-points occur can be easily calculated from the models in an exact fashion without the need for state sequence sampling. This allows for both accurate and fast characterisation of the distributions. It will also be seen that these methods allow easy extension to Bayesian change-point distribution analysis. The methods will be applied to two diverse data sets. The first is the analysis of CpG islands for gene detection. Change-points here indicate switches to and from CpG islands. The second involves the analysis of GDP macroeconomic data, and here change-points indicate switches to and from recessionary periods in the economy. It will also be demonstrated that the method can easily be extended to find confidence intervals for where recessions may have occurred.
Tuesday, September, 18, 2007
4:00 - 5:00 pm
208 Patterson Hall