Empirical Analysis Of China’s Treasury Yield

The bond market is an important part of the overall financial system.Chinese bond market has developed into the second largest bond market in the world.In addition to bond as an important asset class itself,bond yields can have a significant impact on the prices of other assets.Among the yields of various maturities,the yield to maturity of10-year government bonds is the most important indicator in the market and is an important reference for pricing RMB assets.Deep neural networks have achieved excellent results in recent years in the fields of speech recognition,text classification and financial data analysis.As a variant of recurrent neural network,the hidden layer of LSTM neural network consists of forgetting gate,input gate and output gate.The gradient vanishing problem can be effectively solved by this delicately designed structure.The neural network model can better portray the nonlinear relationship among time series data and is suitable for the field of prediction of financial data.In this thesis,I model Chinese bond yields based on Long-Short Term Memory networks to construct more accurate yield forecasting models and compare them with ARIMA and other models to test the applicability of deep neural networks in the bond field.According to the classical framework,economic growth,inflation and liquidity are the three most important factors affecting bond yields.This thesis constructs eight high frequency daily factors: DR007,Shanghai Composite Index,Brent crude oil price,cement price index,Baltic Dry Index,average price of pork,average price of 28 key monitored fruits and average price of 7 key monitored vegetables,to forecast daily frequency bond yields.In this thesis,multiple linear regression model,ARIMA model and LSTM model are constructed with the ten-year maturity yield of Chinese bond from 2014 to 2021 as the sample data.The forecasting accuracy of different models is compared and analyzed based on MSE.The empirical analysis finds that the eight high frequency factors can explain 67.13% of the variation of the Treasury yield;the LSTM singlefactor model performs better in forecasting the yield for the next two weeks using oneweek lagged data as input;the LSTM multi-factor model,which integrates external factors and its own time-series characteristics,has the best forecasting result.The LSTM multi-factor model’s MSE in the validation set is 0.0032,which is about 50%better than the LSTM single-factor model.When forecasting bond yields,data selection,processing,feature extraction and model construction are very important aspects,and the research in this thesis has some reference value for these contents.