Trading based on classification and regression trees

This thesis investigates whether stock picking based on classification and regression trees can be implemented as a successful algorithmic trading system, if only based on technical analysis. To evaluate the performance of this method a fictional portfolio was constructed from the Stockholm Stock Exchange OMX30, traded on a five-year period. By means of implementation, classification of the assets in the portfolio was initially conducted. By using threshold values of the weekly returns and comparison with the index of the portfolio, every asset was classified as either outperforming, neutral or underperforming. With a satisfactory classification, each asset that is considered as outperforming is held over a period of one week and at the end of the period the position is terminated and a rebalancing of the portfolio is made. If no assets are classified as outperforming, the portfolio is liquidated and invested at a risk-free rate, defined as the STIBOR 1 week rate. When backtesting the model we find that the hit ratio of the overall classification is slightly larger than 50 %. During backtesting over the complete trading period it is found that an immense increase of portfolio value is generated. However, since the model is used in sample no predictive validity outside the range can be made. For this reason, 10-fold cross-validation and resubstitution techniques are employed in order to increase the validity if used in an out-of-sample test. Further, a rolling Sharpe ratio is introduced to evaluate the risk-adjusted returns for both portfolios and it is found that the rebalanced portfolio exhibits greater values. It is concluded that algorithmic trading based on classification and regression trees can be effective in finding patterns that influence the stock prices and that it can form the foundation for an algorithmic trading system. Acknowledgements I would like to thank Karl Hallberg at Nordea for sharing his experience as a quantitative analyst and for his knowledge in algorithmic models. I would also like to thank Professor Boualem Djehiche, my supervisor at the Royal Institute of Technology (KTH), for his efforts and experience of portfolio theory. Further, I would like to thank Gert Engman for providing me with invaluable comments regarding my thesis. Moreover, I would like to thank Johan Obermayer for our discussions regarding mathematical statistics and portfolio theory. Finally, I would like to thank my family for their support as well as my friends at the Royal Institute of Technology for making my time there a great and memorable experience.