Detection of Anomaly Stock Price Based on Time Series Deep Learning Models

Abstract

Anomaly detection is a critical task for financial market, investors, and regulatory authorities, where conventional methods employ rule-based models. With the development of machine learning and deep learning techniques, it becomes more promising to detect anomalous trading behaviors from data. Here we present a deep learning model based on time series LSTM model to detect anomalous behaviors in Chinese stock market. The model is composed of 1dConv-LSTM neurons, which can predict time series stock price data from historical data. We analyzed the price of 14 stocks variations ranging from 2015/01/05 to 2019/12/31 and used univariate and multivariate time series models to generate MAE less than 4.0 consistently. The proposed method improved MSE to 0.0171 on validation datasets. Our model successfully predicts the anomalous price behaviors of `601318' stock in the range of 2019-02-13. Our method provides an automatic way of predicting anomaly stock price behavior in Chinese stock market.