I compare unemployment expectations from the Michigan Survey of Consumers to VAR forecastable movements in unemployment. I document three key facts: First, one-half to one-third of the population expects unemployment to rise when it is falling at the end of a recession, even though the VAR predicts the fall in unemployment. Second, more people expect unemployment to rise when it is falling at the end of a recession than expect it to rise when it is rising at the beginning of a recession even though the VAR predicts these changes. Finally, the lag change in unemployment is almost as important as the VAR forecast in predicting the fraction of the population that expects unemployment to rise. Professional forecasters do not exhibit these discrepancies. Least squares learning or real time expectations do little to help explain these facts. However, delayed updating of expectations can explain some of these facts, and extrapolative expectations explains these facts best. Individuals with higher income or education are only slightly less likely to have expectations which differ from the VAR, and those whose expect more unemployment when the VAR predicts otherwise are 8-10 percent more likely to believe it is a bad time to make a major purchase.