“All of us need to be wealthy by having a pc simply generate piles of cash for us,” writes long-time Slashdot reader TekBoy. “Here is one man’s try at utilizing AI to foretell the market.
From the article (by tinkerer/author/community man Jason Bowling):
Fashions that did nice throughout their preliminary coaching and validation runs would possibly do okay throughout runs on later information, however may additionally fail spectacularly and burn all of the seed cash. Half the time the simulation would generate income, and half of the time it might go broke. Typically it might be just some share factors higher than a coin toss, and different occasions it might be far worse. What had occurred? It had regarded so promising. It lastly dawned on me what I had accomplished.
The outcomes biking round 50% was precisely what you’d anticipate if the inventory worth was a random stroll. By letting my program hunt via a whole lot of shares to search out ones it did nicely on, it did stumble throughout some shares that it occurred to foretell nicely for the validation timeframe. Nonetheless, just some weeks or months later, throughout a distinct slice of the random stroll, it failed. There was no delicate underlying sample. The mannequin had merely gotten fortunate a couple of occasions by sheer likelihood, and I had cherry picked these situations. It was not repeatable.
Thus, it was pushed residence — machine studying will not be magic. It could actually’t predict a random sequence, and you need to be very cautious of your individual biases when coaching fashions. Cautious validation is essential.
I’m positive I can’t be the final to fall sufferer to the decision of the previous treasure map within the attic, however train warning. There are far much less random time sequence to play with in case you are seeking to study. Simulate, validate fastidiously, and concentrate on your individual biases.
Learn extra of this story at Slashdot.