According to the literature, election prediction markets have excellent accuracy rates of prediction. However, one can only acknowledge the prediction results after the elections and cannot discriminate the accuracy rates of particular election predictions prior to the elections. This paper constructs four models to discriminate the accuracy rate of each election contract prior to the election. According to the information of forty original variables collected from the election contracts in the prediction markets, the Logit model can precisely discriminate which election contracts with the highest price criteria of predictions will be likely correct. In addition to the complete sample model, this paper uses election contracts of the 2008 presidential election, the 2009 magistrate and mayoral elections, and the 2010 five-metropolis mayoral elections as out-of-sample tests. In terms of prediction accuracy, the Logit model using forty original variables is the best among the four discrimination models. The accuracy rates of discrimination of the Logit model for correct predictions are all 100%. Nevertheless, the Logit model's prediction ability for discriminating incorrect prediction groups needs to be improved.