We apply state-of-the-art financial machine learning to assess the return-predictive value of more than 45,000 earnings announcements on a majority of S&P1500 constituents. To represent the diverse information content of earnings announcements, we generate predictor variables based on various sources such as analyst forecasts, earnings press releases and analyst conference call transcripts. We sort announcements into decile portfolios based on the model's abnormal return prediction. In comparison to three benchmark models, we find that random forests yield superior abnormal returns which tend to increase with the forecast horizon for up to 60 days after the announcement. We subject the model's learning and out-of-sample performance to further analysis. First, we find larger abnormal returns for small-cap stocks and a delayed return drift for growth stocks. Second, while revenue and earnings surprises are the main predictors for the contemporary reaction, we find that a larger range of variables, mostly fundamental ratios and forecast errors, is used to predict post-announcement returns. Third, we analyze variable contributions and find the model to recover non-linear patterns of common capital markets effects such as the value premium. Leveraging the model's predictions in a zero-investment trading strategy yields annualized returns of 11.63 percent at a Sharpe ratio of 1.39 after transaction costs.