The logistic discriminant model (LDM) is used in the heavy rainfall forecasting with three different schemes. In Scheme 1, 14 impact factors are imported into the model directly so that the model has high simulation skill but low forecast skill because of the colinearity effect among the factors and noise signals. In Scheme 2, the principal component analysis is applied to all the impact factors and only the first 6 leading principal components are used in building the model. Compared to Scheme 1, the simulation skill in Scheme 2 is lower but the forecast skill is higher due to the elimination of both colinearity effect and noise signals. In Scheme 3, the Bootstrap sampling technique is applied to get subsamples in order to obtain the model parameters. Thus the fluctuations in the original time series have been disturbed and only the stable information is remained. Though both the degree of freedom in the fitting and the simulation skill in this scheme are lower than in Scheme 2, the forecast skill is the highest of all the three schemes. Based on above results and by using the forecast data of ECMWF (European Centre for MediumRange Weather Forecasts), an objective LDM heavy rainfall forecasting system has been established and used in the forecasting operation at the National Meteorological Centre of China. Verification results in 2013-2014 indicate that the TS skill using the LDM scheme is generally higher than using the numerical model outputs directly.