Deep Feature Fusion Methods For Trend Prediction Of Limit Order Books

With the rapid development of the financial trading market,Internet finance and mobile finance have gradually become the new financial business patterns of investment trading,which not only brings great convenience to investors,but also makes the transaction data in the financial market grow explosively.In the trading process of financial assets,limit order books(LOBs)can be formed from limit orders in order streams.Investors use LOBs to predict the price trend of financial assets and make the trading decision according to the prediction result.At present,the mainstream methods for LOB trend prediction are deep learning-based ones.Most of deep learning-based methods adopt only the factual information of LOBs,while a few of them take into account the distribution information and the dynamic information of LOBs.The factual information reflects asset prices and volumes directly,the distribution information provides a relationship between supply and demand of assets,and the dynamic information represents dynamic changes of transactions.Therefore,the three kinds of information in LOBs would be helpful to predicting the trend of LOBs.If these information could be fully utilized by a deep learning-based method,it would generate an efficient model for trend prediction.However,the existing methods do not distinguish the three kinds of information,but simply merge these kinds of information at the input,which may result in an insignificant improvement in the prediction performance.In order to make full use of these three kinds of information,this thesis studies the deep feature fusion methods for LOBs trend prediction.The main contributions of our work are as follows:(1)This thesis proposes a trend prediction model based on voting feature fusion for LOBs(VFF-LOB).To deal with three kinds of LOB information,we use three deep learning sub-models according to the characteristics of information,where a sub-model of convolutional neural network(CNN)is used to deal with the factual information to extract local features,and two sub-models of long-term memory(LSTM)are adopted to process the distribution information and dynamic information to capture the global temporal dependence.Finally,a simple voting method is used to integrate the three sub-models.Experimental results show that VFF-LOB can achieve better prediction performance.(2)This thesis proposes a trend prediction model based on hierarchical feature fusion for LOB s(HFF-LOB).To deal with three kinds of LOB information,we use three deep learning sub-models,where CNN sub-model is used to deal with the factual information,and two gated recurrent unit(GRU)sub-models are used to the distribution information and dynamic information.Compared with VFF-LOB,HFF-LOB has two main differences.First,HFF-LOB uses GRU to displace LSTM,which reduces the number of model parameters.Second,the new model uses a hierarchical fusion way according to the output properties of sub-models to improve the fusion efficiency.In addition,the LOB data in different prediction intervals is unbalanced with different degrees.In order to alleviate this issue,all samples are weighted to make the model pay more attention to the categories with fewer samples.Experimental results show that HFF-LOB reduces the computational cost and improves the prediction accuracy.(3)This thesis proposes a trend prediction model based on position attention feature fusion for LOBs(PAFF-LOB).To deal with three kinds of LOB information,we use three deep learning sub-models,where CNN sub-model is used to deal with the factual information,and a residual GRU and a stacked GRU sub-models are used to process the distribution information and dynamic information,respectively.Compared with HFF-LOB,PAFF-LOB not only improves the structure of the GRU sub-model by the residual or stacked connection,but also designs a vector-based location attention feature fusion scheme to enhance the fusion efficiency.Experimental results show that the three sub-models have a strong adaptability to the three types of information,and make PAFF-LOB generate better prediction performance.