Title: Directional predictability of daily stock returns
Authors: Janis Becker - Leibniz University Hanover (Germany) [presenting]
Christian Leschinski - Leibniz University Hannover (Germany)
Abstract: In contrast to the monthly or quarterly case, daily stock returns are generally regarded as unpredictable. While this may be true for the level of daily returns, we focus on the signs of these returns. Using various machine learning techniques, such as neural networks and general additive models, we show that meaningful directional forecasts can be generated. The analysis is carried out using pseudo out-of-sample forecasts for a data set consisting of all stocks in the Dow Jones Industrial Average from 2003-2016. Relevant predictor variables are chosen beforehand - in a separate model selection window from 1996-2003. This model selection procedure is carried out using a cross-validation procedure with forward chaining that is applicable in a time series context. Since the forecast period and the model selection period are strictly separated, the procedure mimics the situation a forecaster would face in real time. Applying common statistical tests, our forecasts are shown to be statistically significant. Therefore, we draw the conclusion that the sign of daily stock returns is (to some extent) predictable. Moreover, trading strategies based on these forecasts suggest the possibility to outperform the market index in terms of return and Sharpe ratio.