Artificial Neural Networks in Stock Return Prediction: Testing Model Specification in a Global Context

• Main Author: Buxton-Tetteh, Naa Ayorkor
• Other Authors: van Rensburg, Paul
• Format: Dissertation
• Language: English
• Published: Faculty of Commerce 2021
• Subjects: Investment Management
• Online Access: http://hdl.handle.net/11427/32567

This research investigates whether artificial neural networks which make use of firm specific fundamental and technical factors can accurately predict the returns of a sample of several large-cap stocks from various markets across the globe. This study also explores which hidden layer configuration leads to the best network predictive performance. Furthermore, this research identifies which firm-specific factors predominantly influence the predictions made by the artificial neural networks. Five artificial neural networks are designed, trained and tested on a sample of 161 stocks from the Russell 1000 and the S&P International 700 stock indices. The investigation period extends over a 166-month period from January 2001 to October 2014 with a 70:30 split for training and testing subsamples respectively. Eighteen firm-specific factors, based on prior research about the presence of style effects or anomalies on the cross-section of global equity returns, are used as the input variables of the artificial neural networks to forecast one-month forward returns of all the stocks in the sample. The five artificial neural networks investigated in this research differed in hidden layer size. Specifically, the number of hidden neurons examined were three, nine, 13, 18 and 30. All five networks train significantly well, with each network's training error indicating a good model fit. Each network also achieves the desirable information coefficient of 0.1 between its predicted returns and the actual returns in the training sample. It is interestingly discovered that network performance generally improves as the number of hidden neurons in the hidden layer increases until a specific point, after which network performance weakens. In the context of avoiding overfitting, the best-trained network in this research is that with 13 neurons in its hidden layer. This is the primary network used for the out-of sample testing analysis. This network achieves an average prediction error magnitude of approximately 7% and an information coefficient of 0.05 during out-of-sample testing. These results underperform their respective benchmarks moderately. However, further analyses of the network's performance suggest an overall poor out-of-sample predictive ability. This is illustrated by a significant bias and a considerably weak relationship between the network's predicted returns and the actual returns in the testing sample. Global sensitivity analysis reveals that growth style effects, particularly, the capital expenditure ratio, return on equity, sales growth, 12-month percentage change in non-current assets and six-month percentage change in asset turnover were the most persistent factors across all the ANN models. Other significant factors include the 12-month percentage change in monthly volume traded, three-month cumulative prior return and one-month prior return. An unconventional result of this analysis is the relative insignificance of the size and value style effects.