Yes, I remember being very intrigued by this, but it is so different from what I do that I didn't underline much of anything in that chapter.
So with daily OHLC data for a product, you can use that data to derive many other price-derived variables. Two he mentions are volatility and acceleration.
Perhaps through rigorous data mining (which he discusses), you find that when the market makes three consecutive higher closes (maybe that variable is called "momentum") and the high-low range expands each day (maybe "volatility" which is expanding), and the rate of change of each close is increasing with each day (maybe "acceleration"), that you expect the next day to close down. Said plainly, the market gets too overbought, too quickly, and you expect a pullback.
So, in this case you would be examining historical data, determining the likelihood of where the market is going in the next 24 hours, and then you take that trade.
It has nothing to do with any notion of mean, or trend--it simply looks at what is most likely to happen next, and then executes based on that. As far as systematic approaches, if I were going to do some kind of computerized method, I would try to work with this type of thinking, as it is very logical, breaks the mold, and is very intriguing.
He does a very good job of being non-specific enough that he doesn't really tell you anything, so the above is just my best guess on the type of scenario he might work with. He talks about hundreds of variables though, and millions of data points, so no doubt it is much more involved than the hypothetical I described above.