Neural Networks and Investor Sentiment Measures for Stock Market Trend Prediction

Soft computing methods and various sentiment indicators are employed to conduct out-of-sample predictions of the future sign of the stock market returns. In particular, we assess the performance of the probabilistic neural network (PNN) against the back-propagation neural network (BPNN) in predicting technology stocks and NYSE up and down moves. Genetic algorithms (GA) are employed to optimize the topologies of the BPNN. Our results from Granger causality tests show strong evidence that all stock returns are strongly related to at least one of the sentiment variables. In addition, the results from simulations show that the GA-BPNN is more capable of distinguishing between market ups and downs than the PNN. Finally, the simulations show that trading given decision rules (for example; buy stock if predicted return is higher than a given threshold) yields to higher accuracy than predicting the stock market ups and downs.