Quantitative Stock Selection Based On Machine Learning

In recent years, the quantitative investment acquires more and more attention with its boundless energy, relying on the advantage of probability to win. Compared with the western mature market, quantitative investment in our country is still in its infancy. The quantitative investment products are in a small scale, investment strategy is lack of diversity, the investment performance is lack of differentiation. In spite of this, seeing from the truth in our country, the quantitative investment will still have a very broad prospects for development. So, studying the theory and practice of stock selection method, choosing a strategy suitable for China Stock Market, and guiding the investors to make quantitative investment have very important practical significance.The stock market is a low signal-to-noise ratio, complex nonlinear system. Machine learning in many fields such as searching and speech recognition are proved to be powerful on modeling the nonlinear fuzzy data, using machine learning method to build a quantitative investment strategy has a certain natural advantages. Stock selection is essentially a sorting problem, investors are hoping to sort out the stocks that have relative better performance to other stocks in the future of the stock, based on this, this article attempts to apply the two sorting algorithm GBDT and GBRank which are relatively mature in the field of machine learning to stock selection problem. Based on the analysis of technical theories, this article has constructed the short-term stock selection strategy based on pattern recognition and long-term stock selection strategy based on momentum and reversal effect. The former build characteristic vector according to individual stocks’ price movements of morphological structure in the past month, use machine learning algorithm for automatic pattern recognition; The latter construct characteristic vector based on the momentum and reversal factor in different time period, using machine learning algorithms automatically learning momentum and reversal effect on the time scale of distribution from a large amount of data. The experimental results show that the two strategies using GBDT sorting algorithm, in the past four years all can significantly outperform the CSI 300 index, have certain reference significance to the investment decision-making of the traders.The innovation of this paper is combining machine learning and technical analysis to construct quantitative stock selection strategy based on pattern recognition and the momentum and reversal effect, solve the disadvantages of traditional stock selection method such as models are difficult to discover, parameters are difficult to determine. This paper the first time apply machine learning algorithm GBDT and GBRank which are better sort of information retrieval algorithm to the field of quantitative stock selection. The experiment proves that the two strategies using GBDT algorithm have strong profitability. This paper makes some data processing to improve the performance of the machine learning when extracting feature vector from the strategy, such as make some noise reduction, introduces quantile.