Multiple change-point autoregressive moving average model

Abstract

This thesis presents a new method to obtain multiple change-points. These change-points are used in a Multiple Change-point AutoRegressive Moving Average (MARMA) model for a fluctuated time series prediction. A change-point is captured by using a sample reduction strategy. The statistical residual normality test is used to validate the change-point detection in our strategy. If the residual series is not normally distributed, the initial point will be removed until the residual series has normality or not enough sample is left. The in-sample dataset is selected from the dataset in the last cluster. In this framework, we examine one-step ahead prediction of an annual sunspot data between the year 1920 and 2008 and ten specific daily closing prices of Thai Stock Exchange, during the year of 2010 to 2012. These ten indices are ADVANC, AOT, BANPU, CPALL, DTAC, JAS, KBANK, LOXLEY, PTT and STANLY. The result of the multiple change-point autoregressive moving average models obtains a better performance than the threshold autoregressive models; the autoregressive integrated moving average models and the generalized autoregressive conditional heteroskedasticity models using R programming. The prediction accuracy is reported by the mean absolute error, root mean square error, mean absolute percentage error and mean square error.