To address the issue that the traffic load effects on bridges are neither independent nor identically distributed due to the influence of multiple events, the predictive method of composite generalized Pareto distribution (CGPD) was improved using joint threshold, which is robust to predict the extrema of mixture peaks over threshold (POT) using any tail approximation function. The autocorrelation coefficient informed sampling interval method, K-S test based threshold selection, and probability weight moment method were proposed to resolve the time-independent test of POTs, threshold selection, and parameter estimation in CGPD, respectively. Theoretical solutions were conducted to verify the accuracy of the improved CGPD model and its critical techniques, and the CGPD method was implemented in the extrema prediction of realistic traffic load effects on short to long span bridges. Results indicated the proposed practical techniques showed good application effect and generated accurate results, which provide strong support for the implementation of CGPD. Numerical examples showed the CGPD could precisely predict extreme values of multi-event driven samples with relative error below 3%. In contrast, the conventional single generalized Pareto distribution (SGPD) exhibits a significant deviations compared with CGPD. The realistic traffic load effects on short and medium span bridges can be categorized based on the events of number of trucks involved, but that on long span bridges can be categorized based on events in rush hours or normal hours according to the ratio of hourly traffic volume to hourly truck. The CGPD method is convenient to predict extreme load effect in any return period, whereas the SGPD method would produce significant deviation in extrema with a maximal relative error of 13.7% compared to that of CGPD.