Mean-reversion and trend-following strategies have some kind of a theory behind them that explains why they might work, if implemented well. Pattern-recognition, on the other hand, seems like nothing more than data mining and overfitting. Could patterns possibly have any predictive value? In other words, is there any theoretical reason why a pattern observed in historical data would be repeated in the future, other than random chance or self-fulfilling prophecy where the pattern "works" because enough traders use it?
Weak form market efficiency says that you can't predict prices based on past prices. Or that technical analysis doesn't work. I think that the tests of weak form market efficiency are pretty conclusive and show that the US stock market is weak-form efficient; at least on a a timeline longer than a few minutes.
That's not to say that markets are "efficient". The tests of semi-strong form efficiency (i.e., can't predict prices from all public info) are still debated a little, but I think most would say that markets are not semi-strong form efficient. You can do fundamental analysis have a chance a determining winners and losers. And markets are definitely not strong form efficient (i.e., can't predict prices from all info, public and private).
So does technical analysis work? I don't think so. Some may be earning abnormal returns, but they're likely taking risks that aren't obvious from the charts. Or in the context of mean reversion, yes, most of the time things revert to the mean, but they may not revert to the mean within your tolerance for pain. I think the best light read on the mean reversion topic is "When Genius Failed". Their convergence trades on off-the-run and on-the-run Treasuries were "right", but they went further away from the mean before converging after insolvency.
General answer to a very general question:
If you find a significant pattern which distinguishes between structure and noise you understand something about that system. You have a model about it so you can extrapolate and forecast. On that basis you can use this model to make money. In that sense mean-reversion and trend-following are also "only" strategies that use a model derived from the data (or where ever from).
Take evolution as a proxy: The living organism also have a model about their environment (which is partly stochastic, too). The successful organisms use that to survive and breed.
As an aside: In terms of a philosophical basis you can say that it is the belief that the past has meaning for the future (inductive argument) - but this is only a belief which cannot be justified in itself (saying that it worked well in the past is itself inductive and therefore we have a catch-22 here)
To my point of view there are never any reason that a pattern or anything else would repeat in the futur. Actualy I don't see any difference between pattern recognition and mean reversion/Trend Following in term of theoretical proof. One can read Pr Andrew Lo : "Foundations of Technical Analysis". It tries to give a theoretical background to TA by studying the enpirical distribution of stock returns conditionned or not on the presence of predefined chart pattern. His result is that there is a diference that justifies the use of TA.