Bond return predictability and its economic value have always been a hot but controversial topic． Using regression models，this paper examines both the statistical and economic significance of bond return predictability in Chinese markets，and analyzes the non-Markov and stochastic volatility properties of bond yields．On the basis of the above analysis，a systematic method is proposed for constructing non-Markov dynamic term structure models ( DTSMs) under a generalized Heath-Jarrow-Morton ( HJM) framework with stochastic volatility，which is then used to investigate the roles of the non-Markov property and stochastic volatility in bond return predictability and its economic gains realizing． Finally，this paper analyzes the economic drivers of bond return predictability． Empirical results show that bond return predictability in Chinese markets is statistically significant，which can also be converted into significant economic gains． The non-Markov property and stochastic volatility are of critical importance in the converting． Moreover，time-varying risk premia driven by the economic environment are the main sources of the bond return predictability in Chinese markets，while unspanned stochastic volatility factors also contain much information for future bond returns．