Data Augmentation Based Stock Trend Prediction Using Self-organising Map

Abstract

Stock trend prediction has been of great interest for both investment benefits and research purposes. Unlike image processing or natural language processing, where the amount of data could easily reach a million order of magnitude, the application of artificial intelligent models is however limited in the domain of stock prediction because of insufficient amount of stock price data. This article seeks to ameliorate the stock prediction task from a different angle and provides a novel method to enlarge the training data by firstly clustering different stocks according to their retracement probability density function, and then combine all the day-wise information of the same stock cluster as enlarged training data, which is then fed into a recurrent neural network to make stock trend prediction. Experimental results show that this data augmentation technique suits for deep learning methods and notably improves the stock trend prediction task.