Stock Price Anomalies Analysis and Detection Based on Machine Learning

Abstract

Through long years of development in the financial industry, more people and investments have been immersed here. However, the highly volatile stock market often presents investors with significant surprises. Therefore, identifying anomalies within stock prices in a timely manner is a widely-discussed problem, successfully detecting anomalies can help investors avoid loss, even gaining profit in some cases. In this essay, the performances of the Autoencoder and Light Gradient Boosting Machine are compared in predicting anomalies, employing supervised learning and unsupervised learning respectively. For Autoencoder, we construct layers and use a target loss to make decisions. As for the Light Gradient Boosting Machine, we split the original data into a test set and training set and trained the model. To assign labels, a statistical approach is employed to pre-test the data for normality. From an overall picture, a threshold of 4 achieves the best accuracy performance, proving that the number around four would be a good choice while detecting the strange turning trend in stock price.