With the vigorous development of economy and the continuous innovation of financial technology,the intelligence of the securities market is gradually improved,quantitative trading has a bright future.Stock exchanges in China are taking order-driven trading mechanism,stock limit order book(LOB)is an important reference for investors to explore effective information and make investment decisions.Mid-price of stock LOB is proposed to describe the dynamics of short-term stock price behaviors,so that accurately forecasting the movements of mid-price has great significance.In the recent literature,extensive research has been only conducted on stock LOB data to predict mid-price movements.Obviously,the information source used by these research is relatively single.Electronic trading platform can accurately record every single order in real time,producing a wealth of investor trading behavior data.Trading behavior reflects the psychological factors and the market information processing of investors,which will have a direct impact on the price.Submitting large orders can reflect the participation of institutional investors,while canceling orders can reflect the intensity of investors’ commitment.Therefore,how to identify hidden features and patterns of trading behavior effectively is worth exploring for mid-price movements forecasting.Deep learning models are perfect candidates for analysing such amounts of data,since they perform significantly better than the conventional machine learning methodologies when facing a large amount of high-dimensional,non-linear and volatile stock time series data.This thesis quantifies the transaction behavior of investors submitting large orders and canceling orders.At the same time,considering that the historical information of different time has different influence on the predictions,we introduce temporal attention mechanism into deep GRU network,and a mid-price movements forecasting model of stock LOB based on deep learning is designed.Firstly,in order to provide more valuable information for stock LOB mid-price movements prediction,we mine the trade size of large order and cancellation on buy and sell sides to construct large order features and order cancellation features.Secondly,we construct a deep GRU network with dual temporal attention mechanism to learn more complex dynamics in time series and adaptively assign weights to historical data at different times.Finally,we conduct an empirical analysis using intraday stock trading data sets of SZSE.It’s experimentally demonstrated that the proposed large order features and order cancellation features significantly improve the prediction results,and indicate that large order features has a greater impact on results than order cancellation features.A novel combined model is also proposed for stock LOB movements forecasting which can effectively discover the value information of time series and leads to better results than other models.In this thesis,we extract the valuable trading information implied in the behavior of investors submitting large orders and canceling orders.Besides,we develop a deep learning model to predict mid-price movements of stock LOB.To a certain extent,this paper improves related research in the field of financial market microstructure,enriches the price formation mechanism and price discovery mechanism of order driven market at a micro level,and has some practical reference value in the field of quantitative trading. |