Trend Prediction in Finance Based on Deep Learning Feature Reduction

One of the main features of Deep Learning is to encode the information content of a complex phenomenon in a latent representation space. This represents an element of definite interest in Finance, as it allows time series data to be compressed into a smaller feature space. Among the different models that are used to accomplish this task are Restricted Boltzmann Machines (RBM) and Auto-Encoders (AE). In this paper we present a preliminary comparative study in the use of these techniques in predicting the trend of time series finance. We attempt to outline the impact of architectural and input space characteristics have on the quality of prediction.