By means of the time-discretization approach, a stochastic difference equation from the geometric mean-reversion process was derived, and then a nonlinear regression model was established. In this way, the distribution and estimation for each parameter were obtained with Bayesian inference. Monte Carlo simulation results proved the effectiveness of this model. Lastly, the one-step forward forecasting was performed for the daily data of YEN/RMB by means of three models:ARMA-GARCH model, nonlinear ARI model and geometric mean-reversion model. The results have shown that the geometric mean-reversion model is a bit better than the nonlinear ARI model but much better than the ARMA-GARCH model.