Applying Machine Learning on Momentum: An Empirical Study Testing Whether Machine Learning Can Exploit Time-series Patterns to Generate Abnormal Profits

Martin Brenøe

Student thesis: Master thesis


For decades scholars and practitioners have tried to create trading strategies that “beat the market”. A more recent addition to the numerous trading strategies trying to create abnormal profits are the so-called quantitative trading strategies based on algorithmic trading. Through an empirical study, this paper is investigating whether some of the most commonly used machine learning classification algorithms including k-NN, Random Forest and Naive Bayes Classifier are able to generate abnormal returns by exploiting time-series patterns. The thesis is firstly presenting the results of 16 momentum portfolios, constructed from stocks listed on NYSE from January 1993 to December 2016, based on Jegadeesh and Titman (1993)’s framework. The results presented are in favor of the existence of momentum, as the momentum portfolios realize significant abnormal returns. Our findings therefore support that it is possible to generate abnormal returns based on time-series patterns. Secondly, a training protocol is conducted where the stocks listed on NYSE from December 1925 to December 2016 are labelled with a class to separate the 25% best- and worst performing stocks each month. The algorithms are then trained on data from December 1925 to December 1992 to predict whether a stock is a ”loser”, ”winner” or ”neutral” stock based on its returns in the previous 12 months. Finally, the algorithms’ prediction abilities are tested on the same data-set applied for the momentum strategies. The thesis presents the results of 4 long-short equally weighted portfolios for each classification algorithm over 1-, 3-, 6- and 9-month holding periods. Of the investigated algorithms it is concluded that Random Forest is the best algorithm to predict winners and losers using time-series recognition. Its long-short portfolio with a holding period of 1-month yielded an average monthly return of 0.96% and a Sharpe ratio of 0.91. Moreover, our findings show that the portfolio generated a significant Jensen’s alpha, which could not be explained by the additional three factors included in Carhart (1997)’s four factor model. The results from Carhart’s four factor model, indicated that both the winner and loser portfolio mainly consisted of small stocks however. The results of a further investigation show, that the long-short portfolio with a holding period of 1 month still yielded abnormal returns when Random Forest was restricted to pick stocks with a market-cap above the 30%-percentile, which increases the robustness of our results. The thesis concludes that it is possible to apply machine learning algorithms to generate abnormal returns based on historical return data only. However, due to the low prediction accuracy, combined with the fact that the advanced algorithms are ’black boxes’, it is also concluded that it would require further research and tests of the algorithms before it would be realistic to apply them for portfolio management.

EducationsMSc in Finance and Investments, (Graduate Programme) Final Thesis
Publication date2018
Number of pages89