Chinese cities are experiencing severe air pollution in particular, with extremely high PM2.5 levels observed in cold seasons. Accurate forecasting of occurrence of such air pollution events in advance can help the community to take action to abate emissions and would ultimately benefit the citizens. To improve the PM2.5 air quality model forecasts in China, we proposed a bias-correction framework that utilized the historic relationship between the model biases and forecasted and observational variables to post-process the current forecasts. The framework consists of four components: (1) a feature selector that chooses the variables that are informative to model forecast bias based on historic data; (2) a classifier trained to efficiently determine the forecast analogs (clusters) based on clustering analysis, such as the distance-based method and the classification tree, etc.; (3) an error estimator, such as the Kalman filter, to predict model forecast errors at monitoring sites based on forecast analogs; and (4) a spatial interpolator to estimate the bias correction over the entire modeling domain. One or more methods were tested for each step. We applied five combinations of these methods to PM2.5 forecasts in 2014–2016 over China from the operational AiMa air quality forecasting system using the Community Multiscale Air Quality (CMAQ) model. All five methods were able to improve forecast performance in terms of normalized mean error (NME) and root mean square error (RMSE), though to a relatively limited degree due to the rapid changing of emission rates in China. Among the five methods, the CART-LM-KF-AN (a Classification And Regression Trees-Linear Model-Kalman Filter-Analog combination) method appears to have the best overall performance for varied lead times. While the details of our study are specific to the forecast system, the bias-correction framework is likely applicable to the other air quality model forecast as well.