16 kinds of linear regression prediction models are developed based on precipitation data of 160 stations in China and ECMWF reanalysis data (ERA-Interim). The models include station direct/indirect precipitation prediction model group (in indirect models, the East Asian monsoon index is firstly predicted, and based on which, station precipitations are predicted.), region-station direct/indirect precipitation prediction model group (regional precipitations are firstly predicted, and then distributed to the stations). In addition, an ensemble precipitation prediction model is composed with those 16 models as members. There are totally 4 factors in the regression equations, including two types of ENSO index, North Atlantic Oscillation（NAO）index and the mean snow depth on Tibetan plateau．According to the number of factors, the models can also be grouped into 3 factors (including the ENSO indices and NAO index) models or 4 factors (contains all the factors) models. In addition, the differences of the models also lies in their predictor, precipitation or its logarithm. Hindcast for 2005—2016 shows that the average PS score for the station prediction model group is higher than that of region-station prediction model group. The performance of the models taking logarithmic precipitation as predictor is better than those using precipitation itself as predictor. In the station prediction model group, the indirect precipitation prediction models are usually superior to the direct ones, while it is opposite for the region-station model group. For precipitation prediction, the deviation caused by the uncertainty in snow depth of ERA-Interim reanalysis data is greater than the contribution of the factor of snow depth. Among all the models, the direct station prediction model with 3 factors for logarithmic precipitation (MDS-3Ln) get the highest PS score on average, reaching 71 point, which is higher than that of the ensemble model (MEM). It is indicated that the real performance of a linear regression precipitation model is not always consistent with its design strategy, since new factors or processing methods may induce new uncertainties or deviations. Therefore, it is necessary to evaluate the Cost-Effectiveness in model development.