Employing deep learning for stock return prediction on the Oslo Stock Exchange : a comprehensive test of deep learning models for predictive purposes and applicability to trading
MetadataShow full item record
- Master Thesis 
We predict daily out-of sample directional movements of the constituent stocks of the Oslo Stock Exchange Benchmark Index (OSEBX) using Long Short-Term Memory (LSTM) net works, benchmarked against other machine learning and econometric techniques. Our results unambiguously show that the LSTM model outperforms all benchmark models in terms of pre dictive performance. When testing simple long trading strategies utilizing the predictions, we ﬁnd that the LSTM model outperform all other methods with a Sharpe ratio of 3.25 prior to transaction costs from 1999 - 2017. In comparison, the OSEBX had a Sharpe ratio of 0.30 over the same period. We ﬁnd that the LSTM model seems to follow a short-term mean-reversion strategy. While seeing somewhat diminishing excess returns in the last years, the excess returns are still present in the three latest years, which diﬀers from similar studies on other indexes where the excess returns have been found to be absent in recent years. When total transaction costs are implemented we see that the excess returns are lost in the bid-ask spread. Training the model on spread-adjusted returns and imposing advanced strategies leads to a modest Sharpe ratio of 0.37 over the whole period. Even though the trading performance after total transaction costs is not statistically signiﬁcant better than the OSEBX, we see that LSTM networks have predictive properties that can make it a great tool and complement to other trading strategies.