Hacker News Reader

Comment by gs17 | Hacker News Reader

Yeah, e.g. someone who can, for example, consistently predict single bits of output from true random number generators 51% of the time would be evidence of something very strange going on, but not necessarily usable for anything practical. (Maybe if you could gather enough of these people to overcome the "weak" ability?)

View on HN

If you could consistently predict stock moves 51% of the time eventually you’d have all the money in the universe.

It depends on what you mean by 51% of the time. If you’re just making a directional prediction, “will the next market day be up or down?”, that’s nowhere near enough.

A little over 53% of days are up (green) days and yet you can’t use that 3x edge over a 51% predictor to make all the money.

Try "will the X go up or down next second".

> A little over 53% of days are up (green) days and yet you can’t use that 3x edge over a 51% predictor to make all the money.

You can make money, just not ALL the money. And that's because other people know that too.

I don’t have access to tick-by-tick feeds any more to test this, but I’d predict that “X will go up or down next second, [ignoring UNCHs]?” is predicted more than 51% of the time by predicting the same trend as the current/previous second and that you’re going to win on percentage of guesses and lose on money (because the changes in direction seconds [where your guess is wrong] are higher vol than the trend continuation seconds).

Again the difference is that everybody (who cares enough) know about the trends, but only you know about the 51% predictions.

Assuming these information sources are independent - even if the trends predicted the market in 99% and your prediction was only 51% - you can still combine them to get a better estimate than the generally known one.

Fair point, maybe it applies more to categories like dowsing, where doing marginally better than chance (or general unaided human ability to detect whatever) would have huge implications, but might not be profitable to employ when compared to modern technological methods, and definitely doesn't allow for easy reinvestment. Who wants to rely on a dowser who is wrong often and can't explain why when you could have a geological survey done that's more reliable and explainable?

what's the point in a dowser that's wrong often? isn't that like having a "slightly better than chance" super power? ie, can't fly but can jump higher than average (some times)

Yes, exactly.

Just because something isn't particularly useful doesn't mean it cannot exist.

Wait, what? I don’t have a specific cite, but I thought that predicting a random number generator that well is good enough to break common crypto algorithms?

IIRC, doing better than chance by 2^-32 is enough for the randomization system to be thrown out, and marching the output 51% is way better than that.

But yes, I’d agree with a steelmanned version of your point with a much smaller improvement over chance. But then, you would have to do a realllllly long test and it wouldn’t actually look that impressive.

Thanks, replace 51% with (50 + small non zero)%, anything more than chance, assuming the RNG has no bias to exploit, is "doing the impossible" but not necessarily feasible to profit off of.

Hence my last paragraph: that would mean that psychic powers aren’t all that impressive and just have effects on the 32nd most significant bit.

AdaBoost, etc. were invented precisely to make good use of weak predictors.

Reply on news.ycombinator.com