This research investigates alternative Bayesian approaches to the existing static smooth transition auto-regressive (STAR) models for auto-regressive non-linear time series. The approaches we propose are dynamic and analytic being thus suitable for non-stationary real-time AR processes such as those exhibiting asymmetric cycles and/or sudden changes of level and variability. Examples include unemployment rates, industrial production and hourly electricity consumption and prices data.
An application to a time series of hourly electricity consumption (in Mega Watts) in southern Brazil, during the South-Africa Football World Cup in June 2010, showed the forecasts produced by the dynamic model (24-hours in advance) to quickly adapt to calendar effects and to be quite accurate. The figure below shows the forecasts, the 95% prediction intervals and the observed hourly electricity consumptions for the 13-19 June 2010 week. Note the sudden drop in consumption on Tuesday, 15 June 2010, afternoon when Brazil played its first match against North Korea.
More details about dynamic Bayesian STAR models can be found here.