1: \begin{abstract}

2: Though out-of-sample forecast evaluation is systematically employed with

3: modern machine learning methods and there exists a well established classic

4: inference theory for predictive ability, see, e.g., West (1996, Asymptotic

5: Inference About Predictive Ability, \textit{Econometrica}, 64, 1067-1084),

6: such theory is not directly applicable to modern machine learners such as

7: the Lasso in the high dimensional setting. We investigate under which

8: conditions such extensions are possible. Two key properties for standard

9: out-of-sample asymptotic inference to be valid with machine learning are (i)

10: a zero-mean condition for the score of the prediction loss function; and

11: (ii) a \textquotedblleft fast rate\textquotedblright\ of convergence for the

12: machine learner. Monte Carlo simulations confirm our theoretical findings. We recommend a

13: small out-of-sample vs in-sample size ratio for accurate finite sample inferences with machine learning. We illustrate the wide

14: applicability of our results with a new out-of-sample test for the

15: Martingale Difference Hypothesis (MDH). We obtain the asymptotic null

16: distribution of our test and use it to evaluate the MDH of some major

17: exchange rates at daily and higher frequencies. \newpage

18: \end{abstract}

19: