Building Technical Analysis Strategies Using Multivariate Longitudinal and Time-to-Event Data in Stock Markets

Stock market prediction by machine learning techniques has been attracting more and more attention. Technical analysis trading strategies with binary classifiers that can predict market moving direction are such typical applications. However, there exist two deficiencies in applying binary classifiers. First, they only predict whether instead of when the interested event will occur. Second, they usually only use the cross-sectional information, without taking account of the longitudinal evolution of the features. In this paper, we propose to utilize multivariate functional principal component analysis (MFPCA) to overcome the second deficiency and obtain better trading strategies. MFPCA is used as a data augmentation tool to systematically extract longitudinal informative features that can replace technical indicators. Technical analysis trading strategies enhanced by survival models with MFPCA are built, with backtesting on the daily trading data of S&P 500 stocks. The experimental results show that MFPCA can significantly improve both the performance of the survival models and the trading strategies. We further show the contribution and effectiveness of MFPCA by interpreting the deep learning survival model and comparing with the long short-term memory model that can process multi-timestep information.