| Limit order book data,as the most direct portrayal of traders’ trading behavior in the financial markets,records in detail information about the price and quantity of assets ordered.With the development of computer hardware,fast reading and writing of financial data and storage of large amounts of data has become possible,while the order book data provided to investors by exchanges is becoming more frequent.This high frequency data describes how market prices respond to changes in supply and demand and implies traders’ price expectations for the asset.The study of limit order books can,on the one hand,provide a better understanding of market microstructure,such as price formation and market liquidity.On the other hand,it can help market participants discover the best trading strategies and reduce trading costs.The research on multi-asset volume and price relationship based on high frequency data will provide some ideas for the mining of order book event factors and the discovery of price impact mechanisms.With more studies show that assets other than the asset’s own order book information can also play a role in explaining the price movement of the target asset,the price impact of multi-asset order book information has gained widespread attention.In order to make full use of order book data,this paper attempts to build a multi-asset cross-interval price impact model by using price and volume information of multiple assets.For the introduction of a large amount of information,feature selection is considered to filter key information to increase the explanatory power of the model while reducing the model complexity.Since the order flow imbalance(OFI)is an classical order book event factor with strong price explanatory power,OFI is used as an order book information variable for each asset,and the price change of the target individual asset is used as the dependent variable to select important assets based on LASSO regressions using sample data.In addition,a combination of LASSO regression and lightGBM algorithm is used to determine the fixed important assets of each asset,which is used as an empirical parameter in the later model implementation to omit the initial feature selection and thus improve the operational efficiency.Recurrent neural networks provide a method to extract features from time series and capture complex dependencies,so a long and short-term memory model is used to mine the order book event factors of multiple assets after feature selection,and finally a multilayer perceptron is used to establish the association between order book event factors and price movements in different intervals.This paper takes the high-frequency data of Dalian Commodity Exchange futures as the research object,empirically analyzes that the price explanatory ability and short-term forecasting ability of multi-asset order flow imbalance on the target asset,and gives the fixed futures varieties with significant influence on the price movement of target futures based on LASSO regression and lightGBM algorithm.The multi-asset cross-interval price impact model is further explored for the coking coal future,and the validity of the model is demonstrated by verifying that the multi-asset cross-interval price impact model can better explain contemporaneous and forecast partial over-range price movements than the cross-interval price impact model in the regression and classification cases. |
