Abstract:

Traditionally, statistics uses digital characters, such as mean and variance, instead of distributions for simple. Especially, classical time series methods are based on second moment properties, such as autocorrelation function (acf) and partial autocorrelation function (pacf). This is a pragmatic approach when we have not enough data and powerful computing tools. Now, as the information technology develops so rapidly, we have "big data" and unprecedented processing power to analyze them. This situation stimulates new approach in statistics. In this talk, we will introduce some methods in this direction for time series analysis, including extended Hidden Markov Models. We also will discuss how to adopt the emerging machine learning methods in time series analysis.