Research On Intertemporal Arbitrage Model Of Nonferrous Metals Based On Ensemble Learning

Under the guiding objectives of deepening financial supply-side structural reform and improving modern financial system,China’s financial system has been improving day by day and financial derivative products are becoming more and more abundant.With the continuous upgrading of artificial intelligence and big data algorithms,the field of quantitative investment is also developing and growing step by step.Among them,data mining has become an important means to study big data,and ensemble learning plays an important role in data mining.In complex situations,it is extremely difficult to construct efficient single classifiers and there may be instability in the output,and ensemble learning can effectively overcome this drawback.Futures market has the functions of price discovery,risk avoidance and hedging.Intertemporal arbitrage trading is a very popular trading method for investors,traditional intertemporal arbitrage models such as Bollinger bands model,which are limited by fixed parameters,are no longer able to adapt to the ever-changing capital markets,while emerging artificial intelligence algorithms such as ensemble learning are better able to dynamically explore various market characteristics.In this paper,we take the Shanghai Futures Exchange non-ferrous metal futures as the research object,and select the Random Forest,Adaboost,XGBoost and Light GBM models of ensemble learning to predict the spread between the main futures contract and the secondly continuous contracts,develop the corresponding intertemporal arbitrage model and compare with the traditional Bollinger bands model,find the differences between the ensemble learning model and the traditional arbitrage model are compared based on various evaluation indicators.The results show that,firstly,the traditional Bollinger bands arbitrage strategy has significant differences under different period and standard deviation multiples settings,and the traditional model can still obtain certain stable returns by establishing a training method similar to machine learning and rolling the optimal parameters to achieve an adaptive Bollinger bands strategy;secondly,four ensemble learning models are established to set a time series rolling training window for the spread between contracts to complete the prediction,and all four models show good prediction accuracy,among which the XGBoost model has the highest prediction accuracy and the Light GBM model has the fastest training speed;then,the intertemporal arbitrage strategy is established according to the prediction model,and the changes in the effects of the four ensemble learning models tend to be consistent under different thresholds,they all perform significantly better than the traditional arbitrage model with adaptive parameters,and pass the stability test under different shock costs and minute-level data.The findings of this paper validate the superiority of ensemble learning models in nonlinear sample forecasting and their ability to mine information from time series data,and also provide new ideas for the development of low-risk intertemporal arbitrage strategies.