Computer Intensive Forecasting Methods
The Centre has been very active in researching computer intensive forecasting methods. Primarily the focus has been on the application of neural networks in forecasting. These models are able to capture complex and dynamic effects and interactions in the time series. In particular they are able to model both linear and nonlinear information making them a very potent forecasting tool. At the same time this increases their modelling complexity substantially. The Centre is actively investigating several aspects of how to best model and implement such models in an accurate and robust way.
Crone and Kourentzes (2010) and Kourentzes (2012) have looked at how to automatically identify input variables for neural network forecasting models. This is one of the key problems in neural network modelling, as there is lack of consensus on how to best identify the relevant inputs, specifically when they are interacting nonlinearly. Crone and Kourentzes (2009) investigated how to model seasonal time series with neural networks. A key advantage of these models lies in their ability of modelling multiple overlaying seasonalities in an elegant and simple way. The focus of the centre on neural network model specification is demonstrated in the NN3 and NN5 competitions, lead by Crone, which identified several novel computational intelligence methods (Crone et al. 2011). These competitions demonstrated major improvements in computer intensive forecasting methods and highlighted best practices.
Crone and Kourentzes (2011) applied neural networks on electricity load forecasting, investigating whether the common practice of time series segmentation is beneficial. They found that properly specified neural networks did not require such complex pre-processing and outperformed substantially other benchmark models. Furthermore, they were able to naturally model nonlinear effects of temperature on electricity load. Such massive and high frequency datasets introduce several new problems in modelling and forecasting, in particular for data understanding, analysis and monitoring. Kourentzes and Crone (2011) proposed a semi-supervised automatic methodology for identifying and monitoring outlying days of electricity load, while Kourentzes (2011) investigated how to best model those.
Barrow et al. (2010) looked at the use of neural network ensembles for time series prediction and found that they can provide very accurate results, while simplifying the model selection step. Following this stream of research, Barrow (2013) extensively explored the use of boosting to improve the forecasting performance of neural networks.
A full list of publications can be found here.