Jeremias Lukas Aechtner

Abstract
In the financial world various machine learning techniques are applied to different tasks. They are used to manage risk and predict price movements for every kind of asset. Forecasting accuracy as well as especiency thereby play a major role. Therefore in this thesis we wanted to evaluate the performance of three machine learning algorithms that were trained on stocks and ETFs with different volatility. This thesis compares their performance and investigates what impact the volatility in the training data has. The models we used are vanilla LSTM, CNN and XGBoost. A variety of technical and macro-economic indicators were added as features. The XGBoost model performed outstandingly. With approximately one hour execution time it was 32 times faster than the LSTM and about 4.5 times faster than the CNN model while having the best overall error. The LSTM is close behind the XGBoost in terms of predictive accuracy and the CNN has a much higher error. The thesis was aimed to show relations of predictive accuracy and certain characteristics in the data while focusing on the volatility as a characteristic. This can help to clarify strengths and weaknesses of certain models when applied to more or less volatile assets. Some further research questions are suggested in the end.