Data augmentation for stock return prediction

In the last decade, there have been advances in machine learning performance various domains, including image classification, natural language processing, and speech recognition. The increase size of training data is essential for improvement these domains. two ways to larger sets are acquiring more original employing effective augmentation techniques. However, stock prediction studies, sizes datasets not changed much no accepted technique. Consequently, has similar progress prediction. This paper proposes an intuitive technique return New synthetic stocks generated from linear combinations stocks. Unlike previous our mimics actual financial asset creation processes. Our significantly improves accuracy. Moreover, we investigate how characteristics affect performance. We find a U-shape relationship between accuracy improved correlation data.