Correctly predicting upsets in the NCAA March Madness basketball tournament is an annual goal for fans and gamblers alike. Considering the period 1985-2005, we find that NCAA seeding is consistently biased and that first-round upsets are predictable to a statistically significant level. We further find that our model can consistently better the oft-cited RPI statistic as a means of predicting first-round March Madness games. Data available since 2005 in prediction markets and related venues offer opportunities for future research to determine whether "crowds" or investors make similarly biased mistakes in their predictions.