Prediction Of TF Index Based On Random Forest Regression

With the artificial intelligence coming into the people's vision,more and more investors will combine machine learning and securities market investment in order to obtain good returns.Random forest is an important tool for classification and regression in machine learning.This paper regards the 5-year Treasury bond futures index(code:TF0000)into a learning issue,using the TF index day closing price from September 6,2013(the offering day of 5-year Treasury futures market)to December 31,2015,totally 564 samples,were used to select technical indicators and macroeconomic indicators,and then the selected indicators was used to forecast 445 test sets,the TF index daily closing price,from January 1,2016 to October 31,2017.The main work of this paper is:1.We find that the TF index market is inefficient by the Hurst index based on the fractal theory.Therefore,according to the market effective hypothesis,technical analysis and fundamental analysis are useful in this market.2.We select several common technical indicators,with the indication selection set to select their parameters which have the best performance in the history.We also select some macroeconomic indicators which have a high correlation with the TF index.3.By using sole indicators and the combination of technical and macroeconomics indicators as input variables,we forecast the yield of TF index through random forest regression.Furthermore,the dimension of input variable is compressed by principal component analysis(PCA).Therefore,we get four models in total.4.For each of the models above,we multiply the volatility calculated by the GARCH model and the historical volatility calculated by the ordinary variance(the volatility is expressed by~s)by the constant k to be k~sas the threshold.Once the prediction result is greater than the threshold k~s,we buy the TF index and the position is held until the predicted result is less than the threshold k~s.Once the forecast is less than the threshold-ks,we short the TF index and the position is held until the predicted result is greater than the threshold-ks.The results show that from the view of the error between prediction and the real value(judging by RMSE and MAE),technical indicators combined with the macroeconomic indicators work better than sole technical indicators.This may be attributed to the relatively more information contained in the model,and the importance of fundamental analysis to bond investment.Moreover,the processing of input variables by PCA does not necessarily reduce the prediction error,and this also confirms a statement mentioned in the literature review that we need not consider multiple regression multicollinearity problems involved in the random forests model,and sometimes we can even input thousands of independent variables into the model.The results also show that from the view of investment according to the prediction(annualized Sharp ratio),technical indicators combined with the macroeconomic indicators work better than sole technical indicators.This is similar to the results of the analysis of the error(judging by RMSE and MAE).However,different from the error analysis results,the input variables can improve the investment to a certain extent through PCA processing.The innovation of this paper is using rolling windows to test whether the TF index market is efficient,which enhances the reliability of the test results.And we originally combine the random forest regression and the TF index forecasting,.We combine the macroeconomic indicators and technical indicators as input variables,and the output variable is not a simple"up"or"down"classification,but a specific rate of return.Finally,we not only explore the accuracy of the model forecast,but also test whether it makes sense in real investment through a specific threshold filter signal.