We examine short-horizon return predictability using a novel proprietary dataset of institutional traders with known identities. We estimate investor-specific short-term trading skill and find that there is pronounced heterogeneity in predicting short-term returns among institutional investors. Incorporating short-term predictive ability, our model explains much higher fraction of variation in asset returns. Ignoring the heterogeneity in short-term trading skill has major implications in modeling price impact. We analyze the differences between trading characteristics of skilled and unskilled investors. A simple trading strategy exploiting our skill estimates yields statistically significant abnormal returns supporting the skill-based interpretation.