Comparison of machine learning methods for financial time series forecasting at the examples of over 10 years of daily and hourly data of DAX 30 and S&P 500

This article conducts a systematic comparison of three methods for predicting the direction (+/−) of financial time series using over ten years of DAX 30 and the S&P 500 datasets at daily and hourly frames. We choose the methods from representative machine learning families, particularly supervised versus unsupervised. The methods are support vector machines (SVM), artificial neural networks, and k-nearest neighbor (k-NN). We explore the influence of different training window lengths and numbers of out-of-sample predictions. Furthermore, we investigate whether kernel principle component analysis (KPCA) improves prediction, through reducing data dimensionality. Additionally, we verify whether combining machine learning methods by bootstrap aggregating outperforms single methods. Key insights from the experiment are: All machine learning methods are in principle useful to predict the direction of (+/−) financial time series. But to our surprise, increasing the window size only helps to a certain extent for hourly data, before it actually reduces the performance. The number of out-of-sample predictions had a small impact, while KPCA made a strong difference for SVM and k-NN. Finally, backtesting selected machines with a trading system on daily data revealed that the lazy learner k-NN outperforms the supervised approaches.