On stock return prediction with LSTM networks

Sammanfattning: Artificial neural networks are, again, on the rise. The decreasing costs of computing power and the availability of big data together with advancements of neural network theory have made this possible. In this thesis, LSTM (long short-term memory) recurrent neural networks are used in order to perform financial time series forecasting on return data of three stock indices. The indices are S&P 500 in the US, Bovespa 50 in Brazil and OMX 30 in Sweden. The results show that the outputs of the LSTM networks are very similar to those of a conventional time series model, namely an ARMA(1,1)-GJRGARCH(1,1), when a regression approach is taken. However, they outperform the time series model with regards to direction of change classification. The thesis shows significant results for direction of change classification for the small Swedish market, and insignificant results for the large US market and the emerging Brazilian market. When a trading strategy is implemented based on the direction of change, a deep LSTM network vastly outperforms the time series model.