Analysts’ target prices have received limited attention in academic research. In this paper we try to fill the gap by developing an innovative multi-layer accuracy metric that we test on a novel database. Our analysis shows that forecasting accuracy is very limited: prediction errors are consistent, auto-correlated, non-mean reverting and large (up to 36.6%). The size of forecasting errors increases with the predicted growth in the stock price, the size of the company and for loss making firms. Additionally, the intensity of research and the market momentum negatively affect accuracy. These results suggest that analysts' research is systematically biased which supports theoretical predictions by Ottaviani and Sorensen (2006). Since stock price forecasting is largely an unmonitored activity, market participants may fail in fully understanding this behavior, thus not arbitraging away these inefficiencies.