Predicting Time Series with Space-Time Convolutional and Recurrent Neural Networks

Convolutional neural networks (CNNs) with their ability to learn useful spatial features have revolutionized computer vision. The network topology of CNNs exploits the spatial relationship among the pixels in an image and this is one of the reasons for their success. In other domains deep learning has been less successful because it is not clear how the structure of non-spatial data can constrain network topology. Here, we show how multivariate time series can be interpreted as space-time pictures, thus expanding the applicability of the tricks-of-the-trade for CNNs to this important domain. We demonstrate that our model beats more traditional state-of-the-art models at predicting price development on the European Power Exchange (EPEX). Furthermore, we find that the features discovered by CNNs on raw data beat the features that were hand-designed by an expert.