Exactly. Skyrefuge, if you understand it, maybe you can give us a coherent explanation of why it works and will continue to do so. :)
I thought
mdmd's first answer that brooklynguy quoted explained it sufficiently non-vaguely, but I guess I can try some more.
The main belief is that humans (in general) are overly-optimistic when they see something improving, and overly-pessimistic when they see something declining. Hopefully this isn't a shocking statement to anyone? If I was mdmd, after the 2nd or 3rd round of this, I would have actually linked to some research (likely from Kahneman) that explicitly reveals this human failing, but since it's such a baked-in part of Boglehead Investing 101 ("don't be like the masses, who buy high and sell low"), he probably figured a citation wasn't required. And since it's not my argument, I'm not going to look it up either.
If there were no delays or limits on human society's ability and interest to move stock prices, then seeing a 100% rise or fall in a millisecond would not be unexpected. But in fact, the US stock market has never risen or fallen 30% in a one day period. On the other hand, the market has never failed to rise by more than 30% in any 20-year period. These two combined suggest that there is some "stickiness" in the market that causes price movements to be spread out over a time-period longer than "instantaneous".
In my mind, some of this lack-of-instantaneousness comes from the fact that the "information" that drives market prices comes from the real world, and the real world is constrained by the laws of physics, which gives a fundamental speed limit to information, as well as a practical speed limit that's far lower. 100 customers parking, walking into a Best Buy, picking up an iPad, walking to the cashier, and paying eventually results in a "iPads are selling great!" bit of information, but that information is not created instantaneously.
So, unless the laws of physics change, or human behavior (which is likely built into our DNA and brains) changes, it seems reasonable that these non-instantaneous market price changes which we have observed in the past will continue in the future. And if a signal can be found that reliably indicates when a non-instantaneous directional price-movement has started, then that can be exploited for profit above market returns.
The challenge then is finding that signal. mdmd seems to believe it's not that difficult, and (given his belief in a range of lookback periods) does not have to be particularly precise in order to work. I get the impression he would admit that this portion of his system comes solely from backtesting.
But it doesn't have to. You could theoretically create a model that you would use to predict the average duration of a directional price movement. It would use inputs like average speed of delivery trucks, frequency of reporting by publicly-traded companies, duration of Congressional terms, vacation times of stockbrokers, time for a wheat crop to go from seed to harvest, and typing speed of participants posting on Internet investing forums. That way it could define a duration somewhere between the one-day and 20-year periods mentioned earlier, without relying on any backtesting.
That sounds like a giant pain in the ass, so in lieu of such a model, it's not shocking to me if a convincing-looking backtest is used in its stead.