In election forecast, the ability to predict the intentions of undecided voters plays an important role in determining its accuracy. This article us-es discriminant analysis with demographic variables, issue positions, and attitudinal variables, respectively, to classify voters who stated their vot-ing preferences. We found that comparatively speaking, attitudinal vari-ables can correctly classify voters with the highest percentage, followed by issue positions and demographic variables. However, none of the above models displayed satisfying results, not even the combined model including all three types of variables. Therefore, there are reasonable doubts employing these variables to construct models predicting the vote intention of undecided voters. The author then attempts to use feeling thermometer scores to classify voters who expressed their vote intentions and to predict the inclinations of undecided voters. After contrasting the results with the actual election outcome, we found that feeling thermome-ter scores have better discriminant power than previous models. The au-thor also found that among the supporters of the three major candidates, ideologically speaking, Chen Shui-Bian's and Lian Chan's supporters are furthest away from each other on their average positions, while James Soong's supporters' average ideological position is closest to that of Lien Chan's supporters.