Assessing Bayesian Covariance Structure Modelling for Time Series Data

Time series prediction has important implications for psychological research, for instance when assessing longitudinal change in outcomes or modelling electroencephalogram (EEG) data. One relatively new tool that can be used for this purpose is the Bayesian Covariance Structure Model (BCSM), which can model serial dependences within time series through a covariance model. The suitability of the BCSM for predicting time series has been investigated and in performance compared to the commonly used autoregressive integrated moving average (ARIMA) model. For this purpose, three simulation studies were designed, in which the sample size and properties of the simulated time series were varied. The results show that for time series of shorter lengths, the BCSM provides more accurate predictions than the ARIMA model according to which the data was generated. However, the autocorrelation function (ACF) of the BCSM’s residuals indicated that there was still significant autocorrelation, suggesting that not all available information in the time series is used efficiently. The BCSM also failed to describe MA dependences in the data. Nonetheless, the BCSM has shown high predictive accuracy, while remaining relatively simple to interpret.