Recurrent Neural Networks Approach to the Financial Forecast of Google Assets

A huge quantity of learning tasks have to deal with sequential data, where either input or out-put data can have sequential nature. This is the case,e.g., of time series forecasting, speech recognition,video analysis, music generation, etc., since they all require algorithms able to model sequences. Duringrecent years, recurrent neural networks (RNNs) architectures have been successfully used in one as well as for multidimensional sequence learning tasks, quickly constituting the state of the art option for extracting patterns from temporal data. Concerning financial applications, one of from the most important examples of sequential data analysis problems is related to the forecasting the dynamic in time of structured financial products. To this end, we compare different RNNs architectures. In particular we consider the basic multi-layer RNN, long-short term memory (LSTM) and gated recurrent unit (GRU) performances on forecasting Google stock price movements. The latter will be done on different time horizons, mainly to explain associated hidden dynamics. In particular, we show that our approach allows to deal with long sequences, as in the case of LSTM. Moreover the obtained performances turn out to be of high level even on different time horizons. Indeed, we are able to obtain up to 72% of accuracy.