Out of sample tests

I am a big fan of so called "out of sample" tests. When researchers find some anomaly within a data set and then others test for the presence in the same data set, we really do not learn much if they find the same thing. But when a new data set is used for the test, we have a much better understanding of the possible anomaly.

In the current JFQA there is just such an article by Richard Grossman and Stephen Shore. Using a data set that goes from 1870 to 1913 for British stocks, the authors find no small firm effect, and only a limited value effect.

In their own words:

"Unlike modern CRSP data, stocks that do not pay dividends do not outperform stocks that pay small dividends during this period. But like modern CRSP data, there is a weak relationship between dividend yield and performance for stocks that pay dividends. In sum, the size and reversal anomalies present in modern data are not present in our historical data, while there is some evidence for a value anomaly."

Which makes me wonder how many other things we think we "know" we really don't.

The current version of the paper is not listed on SSRN, but a past version of the paper is available (at least right now) here.