Empirical investigation on the performance of a feed-forward artificial neural network on the Nordic stock markets

Abstract

In this paper, the authors have made an empirical investigation on the performance of a feed-forward artificial neural network (ANN) on the four main Nordic stock markets, Sweden, Norway, Denmark, and Finland. First, a benchmark OLS regression model is compared against an ANN model to see which model performs best in terms of predictive accuracy and has the least amount of error. From the results of the study, it cannot be concluded that an ANN model outperforms the benchmark OLS regression model in predicting risk premia in terms of out-of-sample predictive forecast accuracy on the Nordic stock markets, when using the same dataset (for 5% confidence level). However, it is logical for practitioners to consider ANN for predictions of risk premia since ANN models can accommodate more features and variables than regression models. Thus, the authors of this paper further implemented an ANN model to predict excess returns for the Nordic stock markets and check their performance. The ANN model was further developed by trying five different ANN architectures for each one of the Nordic countries. The best ANN model of the five different ANN architectures for each country was used to evaluate the predictive performance. However, it was concluded that the performance in terms out-of-sample R-squared of an ANN model that should predict market returns in the Nordic stock markets resulted in negative out-of-sample R-squared for all the four Nordic countries and stock markets in this study. The economic interpretation of this result is that a model with a negative out-of-sample R2 cannot explain the data or the trend. An important note here is that our dataset may not be an ideal one to use for predicting returns in the Nordic stock markets. This is because advanced index metrics were not available for dates before 2001, so with this limited amount of data the ANN model cannot train itself properly.