Utilizing Machine Learning for Trading Algorithms Exploiting the Time Series Momentum Anomaly

Momentum or trend following investing refers to trading strategies constructed around the idea that in financial markets, the current trend will, more often then not, prevail. In the context of asset prices, this means that previous returns or the price development of an asset is indicative of similar future returns and price development. While at odds with established theory such as the weaker form of market efficiency, as defined by the efficient market hypothesises, the pricing anomalies have proven robust enough for an industry of funds, using systematic trading, to rely heavily upon them. This thesis aims at building a profitable trading strategy around the momentum anomaly by using machine learning and common momentum indicators. The underlying assets will be futures contracts due to their frequent use in the industry and generally high liquidity. Following the exploration of different machine learning algorithms, Random forest was chosen and subsequently optimised on training data by cross-validation using the model evaluation metric Matthews Correlation coefficient. These fitted models were backtested in three ways and benchmarked against simple trading strategies as well as buy-and-hold strategies using several performance metrics. The final result indicates great performance during strong negative trends, such as the financial crises, and an ability to lessen drawdowns. However, the models ultimately fail to act as robust and profitable strategies over a longer time horizon