[Research Seminar 2018.09.17]Machine Learning Based Stock ClassiﬁcationSpeaker : Chulwoo Han(Assistant Professor in Finance, University of Durham)
In this paper, we develop a neural network model for multi-class stock classiﬁcation using input features derived from widely known momentum factors and apply it to long-short portfolio construction. A naive approach to construct a long-short portfolio, i.e., buying the stocks in the highest return class and selling those in the lowest return class poses two problems in the ﬁnancial context: the two classes can be signiﬁcantly imbalanced; stocks predicted to be in the highest (lowest) return class do not necessarily carry the highest (lowest) expected return. We address these problems by reclassifying stocks utilizing the probabilities of the stocks to be assigned in each class. Empirical ﬁndings suggest that our model can create a long-short portfolio generating a signiﬁcant proﬁt and high Sharpe ratio. We also ﬁnd that economic performance of a model is not always consistent with its statistical performance.