This paper studies the out-of-sample predictability of the monthly market as well as size, value, and momentum premiums. We use a sample from each the US and the Swiss stock market between 1989 and 2007. Our Swiss sample provides an important new perspective as the repeated evaluation of the same (US-) dataset leads to the problem of data mining. To exclude data mining in our predictability study, we test both statistical significance and robustness in the two samples. Our key results are as follows: We find no robust indication that the market premium is predictable. The same is true for the momentum and the value premiums: Our statistically significant results from the US sample look like data mining in light of the results from the Swiss sample. However, the size premium seems to be predictable to some extent, due to the credit spread. We assume three reasons for this - in comparison to the literature - rare evidence for predictability: first, predictability could have evaporated in the last decade as academic research made the respective information public. Second, predictability is, as we demonstrate, not robust to the choice of the methodology. Third, robustness tests unveil many statistically significant interrelations as random data structures, known as data mining. Therefore, we think that the future discussion of predictability should address the issue of data mining by applying robustness tests.