| Abstract [eng] |
It is commonly agreed that the level of financial asset returns is hardly predictable. Hence, in this thesis, instead of focusing on the level, we explore the direction of the return. Therefore, the aim of this thesis is to perform direction-of-change forecasts of exchange traded fund returns and find the method that produces the most accurate forecast results in an out-of-sample environment. For that purpose, we use several classification methods: logistic regression, support vector machines, and random forests. Additionally, a combination of several classification models is considered by constructing ensemble models. In this thesis, the daily returns of 141 ETFs are considered. For modeling purposes, several types of independent variables are considered: technical indicators, financial market indicators, and measures of moments. Modeling is performed on a train sample ranging from 3 March, 2005, to 31 December, 2014. Model comparison is performed on a test sample that ranges from 1 January, 2015 to 9 November, 2017. For model comparison, Diebold-Mariano, DeLong, and Pesaran and Timmermann tests are used. Additionally, models are compared against benchmarks: optimistic and pessimistic forecasts. According to the empirical calculations, the following conclusions were made. The most accurate out-of-sample forecasting results are obtained with the random forests method when the overall average accuracy is 52.2%. That implies that the direction of the daily returns is to some degree predictable and based on the statistical tests performed it was shown that for some ETFs prediction is statistically significant in out-of-sample environment. |
