| dc.description.abstract | We investigate the predictability of abnormal stock returns using artificial neural networks, and examine whether Google search volume data can enhance such predictions. Our results show that the neural network models significantly outperform linear and semi-parametric methods for abnormal stock return prediction. We implement a trading strategy, buying the 50\% of stocks with highest predicted abnormal return, and selling the 50\% of stocks with lowest predicted abnormal return. The neural network trading strategy with a quarterly trading horizon has an average annual return that is 5 percentage points higher than the equally weighted portfolio, after accounting for trading costs. The performance is also improved in terms of portfolio volatility, and thus the risk-adjusted return is exceptional. We find that both the horizon of the input data and the prediction horizon impact the accuracy of predictions, with enhanced performance for longer horizons. Further, we examine the impact of Google search volume data on predictions, by comparing the neural network trading strategy with a benchmark excluding Google search volume data. We find that Google search volume has significant predictive power of abnormal stock returns for a quarterly trading horizon. Our results suggest that a decrease in Google searches over a period of one quarter is associated with significantly decreased abnormal return over the next quarter. However, for an increase in searches there is no such pronounced effect. | |
