We use machine learning to efficiently combine a broad set of signals and produce a testing set of portfolios sorted by ex-ante estimates of expected returns. None of the well-known factor models can explain the returns of the testing set, and we observe monotonically increasing realized risk-adjusted excess returns. A long-short value-weighted portfolio produces significant realized risk-adjusted excess returns above 1\%. We also provide an even more troublesome testing set: ex-ante covariance neutral portfolios sorted on ex-ante estimates of expected returns. A long-short covariance-neutral portfolio produces a Sharpe ratio well above one and no statistically significant covariation with any of the well-known factors, posing notable challenges to both reduced-form and consumption-based asset pricing models.