Simple trading rules lose their edge on the Dow after 1986

In 1992, William Brock, Josef Lakonishok and Blake LeBaron tested two families of simple rules on the Dow Jones index from 1897 to 1986: moving-average crossovers and trading-range breaks. Returns after buy signals were higher than returns after sell signals, and the result became one of the most cited papers on technical analysis.

I wanted to know two things. Does the result survive when you count how many rules a researcher could have tried? And does it hold on the years the paper never saw?

The method

I re-ran the paper’s 26 rules on the MeasuringWorth copy of the Dow and compared the averages with the paper’s printed tables. I then added 370 rules in four families (moving average, support and resistance, filter and channel), with every grid fixed before any run. Halbert White’s Reality Check, a bootstrap test, gives the best rule’s p-value after it adjusts for the size of that search. I ran it on the paper’s sample and on two later spans.

The result

Span	Rules	Best rule’s p, adjusted for the search
1897 to 1986, the paper’s sample	370	0.000
1987 to 1996	364	0.162 to 0.254
1987 to 2025	369	0.112 to 0.165

The test level is 0.05, and each range covers six settings of the bootstrap’s block length.

Inside the paper’s sample, the edge is consistent with the data even after the search is counted. After 1986 it is in tension with the data. On 1987 to 2025 the best single rule has a nominal p-value of 0.001 to 0.007, and the adjustment takes it to 0.112 or more. One good-looking rule out of 369 is what chance alone produces. Against buy and hold, no rule comes close on 1987 to 2025: the adjusted p-value is 0.997 to 1.000.

This agrees with Sullivan, Timmermann and White, who found little evidence for these rules on 1987 to 1996 in their own test of a much larger universe.

What failed

For the moving-average rules, two of the three averages in the paper’s Table III did not match my run within tolerance, so that half of the replication rests on weaker ground. The trading-range-break rules matched all three cells of Table IV. The paper used a different Dow series (Pierce, 1991), so no figure here is an exact copy of theirs.

One day carries much of the later result. On 1987 to 1996, 19 October 1987 alone gave 39.5% of the total for the best of the paper’s 26 rules. Without that day, the adjusted p-value on 1987 to 2025 falls to 0.059 to 0.088. When the rules may also read closes from before 1987, it is 0.054 to 0.073. Both are near the level and neither is below it.

What was not tested

The paper reports no transaction costs, and I applied none. Costs can only make the later result weaker. Thirteen of the paper’s sixteen testable claims are not run yet, which is no evidence against them. No second person has reviewed how I read the claims out of the paper, so treat each verdict as provisional.

A second, later run of the 1987 to 1996 test gave the same report byte for byte.