The Combination Forecasting Of Term Structure Of Interest Rates Based On Neural Networks

Term structure of interest rates refers to the relationship between the maturity yield and the residual maturity of a Treasury bond at different maturities at a certain time.It is not only the benchmark of financial product pricing,but also an important reference element in macroeconomic research.It is of practical significance to discuss how to obtain the precise term structure of interest rates.Previously,term structure of interest rates is deduced from certain assumptions,which will be inevitably affected once the real market does not conform to the assumptions of the model.As a big data algorithm,neural network completely relies on market data to fit the resulting model,which is less subjective and can effectively avoid the errors caused by the improper parameter setting of the traditional model.In this paper,four kinds of neural networks are used to predict term structure of interest rates,on the basis of which,the forecast results are compared and the combination optimization is carried out.The main work of this paper are as follows:(1)The optimal parameter values of the four neural networks were determined,respectively:the number of hidden nodes of BPNN was determined by rule of thumb ?,and the number of iterations was 5000.The number of hidden nodes of WNN is determined by rule of thumb III,and the number of iterations is 10000.The number of hidden nodes of RBFNN is determined by rule of thumb IV,and the smoothing parameter is 0.25.The smooth parameter of GRNN is the optimal value of the year for 7 test samples.(2)By comparing the prediction results of four kinds of neural networks,the mean of error of WNN is the smallest and 1.17;the standard deviation of BPNN is the smallest and 0.79.(3)Based on four models of neural networks,five kinds of combined prediction models were constructed,and the five combination models were compared and analyzed with each kind of neural network and five combination models.The error and the volatility of forecast result of the fifth combination model is the worst of five kinds of combined prediction models,and the mean of error is 0.66,the standard deviation is 0.67.Meanwhile,this result is lower than the optimal value of the mean error and the standard deviation in the four types of neural networks.It is proved that any of the combination forecast models can improve prediction accuracy and stability.The error and the volatility of forecast result of the fifth combination model is the best of five kinds of combined prediction models.The mean error is 0.37 and the standard deviation is 0.37.