This paper proposes a new dynamic Nelson-Siegel (DNS) model with time-varying factor loadings and stochastic volatilities, in which the factors are instantaneously correlated. The proposed model is evaluated statistically and economically. For the statistical evaluation, we examine its out-of-sample yield curve density forecasting. The economic value of the model is analyzed in terms of the bond portfolio choice of a Bayesian risk-averse investor. According to our out-of-sample U.S. monthly yield curve density forecasting and bond portfolio optimization, our model results in substantially more accurate density forecasts and better portfolio performance than standard DNS models do. This finding suggests that incorporating time-varying factor loadings, stochastic volatilities, and factor shock correlations is essential for improving predictive accuracy and bond portfolio strategy.