In this paper, we provide a novel performance analysis of three widely used survey forecasts, by investigating how well they predict the stock market. Judging by their out-of-sample R2s, a standard measure in the market predictability literature, we find that none of the survey forecasts outperforms a simple random walk forecast!, which predicts the future returns simply by their past sample mean. Our study raises important questions on how to well-guard the use of the survey forecasts and how to properly interpret the results from the large literature that rely on the survey forecasts. On the other hand, we find that the stock market short interest, which are profit-oriented and are backed with money, performs well in predicting the market, beating the random walk significantly. Our study suggests that the short interest is by far a much better measure of investor subjective beliefs than any other measures used today.