In this study, an adaptive probabilistic algorithm was developed for the selection of the most plausible Kalman filter based time-varying statistical model by using Bayesian approach. The method was validated by a case study involving the prediction of the sample time series recorded in Macau between 2001 and 2005. Two types of statistical models, namely the time-varying autoregressive model of order p, abbreviated as the TVAR (p) model, and the timevarying autoregressive model with exogenous inputs, abbreviated as the TVAREX model, were adopted to predict the sample time series. By judging upon the model occurring plausibility conditional on the measured data, it was found that using a longer past history did not guarantee better prediction performance. On the contrary, inclusion of explanatory variables which reflect the mechanism of the physical problem can better capture the actual system.