Research On The Design And Application Of High-Frequency Trading Strategy Based On Machine Learning

In recent years,with the development of domestic financial market,more and more professional participants,and the improvement of the regulatory system,quantitative trading technology has developed rapidly.Among many quantitative trading strategies,highfrequency trading still has great potential and development space.At the same time,the recent outbreak of the conflict between Russia and Ukraine has led to dramatic volatility in the international crude oil futures market,although from a macro perspective,the oil price volatility has brought instability to the international and domestic economic environment.But from the perspective of investment practice,the turbulent market environment precisely creates the conditions for the profit of high-frequency trading.Based on the above background,this thesis constructs a futures high-frequency trading strategy based on machine learning model,and applies it to the investment in crude oil futures during the Russia-Ukraine conflict,in order to explore the profitability and robustness of the strategy.The research of this thesis is divided into three parts: first,the construction of factor pool.In this thesis,the basic volume price data of minute crude oil futures are cleaned.Then,the quantitative financial analysis library Ta-lib was used to calculate the volume and price data,and eight types of technical index factors were constructed.Then this thesis filters the factors by Spearman correlation coefficient,Kendall grade correlation coefficient and other indicators,and finally gets 102 effective factors.Second,model training.In this thesis,the data of two years before the Russia-Ukraine conflict event was taken as the training set,and the model was adjusted by the method of 3-fold cross test.This thesis compares the performance of four integrated models,namely random forest,gradient lift tree,XGBoost and Light GBM,and uses accuracy rate,recall rate and f1 score for comprehensive investigation.Finally,Light GBM model is selected to construct the trading strategy.Meanwhile,to solve the problem of unbalanced positive and negative samples in the data,this thesis compares two methods:random undersampling and Borderline SMOTE oversampling,and finally chooses the latter to adjust the sample proportion.Third,strategy construction.The execution logic of the strategy in this thesis is based on the prediction results of the model.In order to make the strategy take into account long-short trading,two models are trained at the same time to predict whether the crude oil futures will rise or fall one minute later.If the prediction is "yes",the strategy executes the corresponding long and short trades and automatically closes out the position one minute later.The backtest results show that the annualized return rate of the strategy can reach162% in the non-extreme market environment and 128% in the extreme market environment during the Russia-Ukraine conflict,which is much higher than the return performance of the full position and linear regression strategy.On this basis,this thesis appropriately increases the strategic leverage ratio,and the strategic return rate can reach 369% under 2 times of leverage.The innovation of this thesis is mainly in three aspects.First,the selection of data set is special.The data in the period of Russia-Ukraine conflict is used to carry out strategy backtest,which not only combines current events and is close to the actual application environment,but also can test the robustness of the strategy in the extreme market environment.Second,the factor selection range of the model is more extensive.In addition to the indicators commonly used in high-frequency trading,such as moving average and momentum,the model also includes indicators such as cycle recognition and shape recognition.Thirdly,the construction method of the model is more suitable for practical application.This thesis follows the idea of integrated learning to model the rise and fall forecast respectively,so that the strategy can flexibly adjust the super parameters according to different commission fees and the rise and fall requirements.