Title: How stock returns predictability using a simple neural network changes over time
Authors: Adam Chudziak - Szkola Glowna Handlowa w Warszawie (Poland) [presenting]
Abstract: Scientific predictions in financial markets are commonly based on theoretical finance foundation. Nonetheless, many techniques used by practitioners do not come from theoretical considerations. Successful trading strategies based on Artificial Neural Networks (ANN) have been reported and are used by leading hedge funds. Although many ANN methods still used today, such as the multi-layer perceptron, have origins in 1950s and 1960s, the interest in them was rather small for almost half a century. Recently, more results on stock predictability using the ANN has been published, many of them use only past market behavior as predictor variables. However, they are usually constrained to specific time periods. The changes in the US stock prices monthly returns predictability are studied by using Artificial Neural Networks since the 1970s to 2015. We show that over this time there were periods when using past market data, even a simple feedforward Artificial Neural Network has some predictive capabilities. The investment strategy based on the predictions is tested. The predictability and profitability of trading varies between securities, but generally tends to decrease in time, with noticeable drop in performance in the 21st century. The results indicate that the stock returns predictability using ANN is not stable over time and that changing characteristics of the market require adaptation of forecasting techniques.