The Anomaly I Would Have Traded in 1985

US markets were closed Friday for the Fourth. The trading day before a holiday used to be the best-paying day on the calendar, and it is worth remembering why, because the story is the cleanest lesson I know about how much to trust a backtest.

The edge was real, and it was large

In 1990, Robert Ariel published "High Stock Returns Before Holidays" in the Journal of Finance. Over 1963 to 1982, the mean return on the trading day before a market holiday ran roughly nine to fourteen times the mean of an ordinary day. There are about eight such days in a year. On those eight days, more than a third of the entire market's return for the year showed up. Lakonishok and Smidt found the same shape holding across ninety years of the Dow. This was not a fragile, data-mined blip. It was one of the most robust calendar anomalies ever documented.

If you had handed me that paper in 1985 with a paper account and a rule engine, I would have coded it in an afternoon. Flat most days, long the close into every holiday. The backtest would have been beautiful.

Then everyone read it

The effect did not merely shrink. In the years right after it became famous, the pre-holiday return in US markets went flat, and in some windows it turned negative. The edge that paid a third of the annual return for two decades stopped paying, and part of what stopped it was being written down.

This is not a one-off. McLean and Pontiff took ninety-seven published stock-return predictors and measured what happened to each after its paper came out. On average, the returns were twenty-six percent lower out of sample and fifty-eight percent lower after publication. Publication itself is a decay function. The moment an edge is legible enough to print, enough capital reads it to trade it away.

The take

The anomaly you can name from a textbook is, almost by construction, the one already arbitraged out. That has a specific and unpleasant consequence for anyone building a systematic engine. A backtest that reaches back far enough to make a calendar edge look significant is reaching back into the era before the edge was public. You are fitting on the decade when it worked, and you will pay execution costs in the decade when it does not.

The pre-holiday trade is the honest tombstone. It was real, it was large, it was published, and publishing it was part of the mechanism that killed it. So when my own engine flags a clean seasonal pattern, the first question is no longer whether it is statistically significant. It is how long the pattern has been in print, and who else can already see it. A significant backtest on a public anomaly is not evidence of an edge. It is evidence that the edge used to exist.

Markets reopen Monday. I will not be shorting the day after the Fourth on the strength of a reversal reported in a 2005 paper, for exactly the same reason.