News for hackers. Stuff that matters.
Re:Gambler’s Fallacy and the Regression to the Mean by coldtea on Friday January 07, 2022 @08:11PM (#29847448)
>That’s a great question, and the answer is a definite no – they are not in conflict. Again, the pressure to think that the past influences future independent events is powerful. Regression to the mean is not a power in the universe that ensures that statistics work out in the end, it is purely a probability.

I find TFA's argument about the gambler fallacy not being associated with regression to the mean quite hand wavy.

"Regression to the mean is not a power in the universe that ensures that statistics work out in the end, it is purely a probability."

Who said the opposite?

The gambler making the gambler's fallacy basically says that after successive black streaks it's more likely to see a red. They aren't saying that there's some "power in the universe" ensuring it. They're saying it's just more probable.

And in a way it is: assuming a non-biased roulette, each successive repeat N-same-streak should be increasingly less likely.

Taking all the groups of N rolls and showing that any N-same-streak going is just one of the possible permutations same as any other (and thus any N+1, N+2, N+3, etc. would be too) doesn't really cut it.

If you happen to actually see a 1000th black roll streak in a casino, you'd better suspect the roulette is rigged/biased, than naively think that "well, even with random independent events, there are bound to be large streaks".

Re:Gambler’s Fallacy and the Regression to the Mean by rout39574 on Friday January 07, 2022 @08:28PM (#29847552)
But "streaks" are irrelevant. Reversion to the mean doesn't tell you things across trials, it tells you things about individual trials.

Try this on: What is the "mean" red-black on the roulette wheel? There isn't one.

The only way you can reasonably expect the past streak to have an impact on the next spin, is if you've concluded the streaks are sufficiently unlikely to cause you to judge the wheel not to be fair.

If it's random, then you have no recourse but to post-hoc description. There is -no- predictive weight.

... And editing your comment to refer to groups of throws is disingenuous. It would have been more honest to modify your case in a reply.

Re:Gambler’s Fallacy and the Regression to the Mean by gexla on Saturday January 08, 2022 @12:20AM (#29848859)
Right, I have never heard "mean" being used in the context of casino games.

I get what they are trying to say though, at some point the variance should start evening out to reflect the odds of the game. But we should be practical and remember we're in a casino. A red/black bet on roulette is one of the few games where you're going to get that low of variance. With slots, you have to play a ton of games before you can get an idea of what the variance is just from your game play. In poker, the rule of thumb is 10,000 hands before the player can be confident of having an idea of how effective a change of strategy is.

Overall, whoever wrote this article is out of their element. Like this bit...

> So we think we can use our amazing powers of pattern recognition to determine that black is “due” and use that power to win big. Casinos love this delusion, because they know that math wins out in the end.

This is contradictory. They're saying the player is wrong, because of math. By that line of thinking, the casino would also be wrong to love it, because of math. In a typical casino game, gambler's fallacy won't make your choice more or less correct, unless the choice is to keep playing or to walk away. The author doesn't specify why the casinos would love this though. As an aside, gambler's fallacy could affect more decision making in poker, but poker is a game of skill and the casino doesn't have a house edge on poker (they take a rake.)

Re:Gambler’s Fallacy and the Regression to the Mean by coldtea on Saturday January 08, 2022 @10:13AM (#29852373)
>If it's random, then you have no recourse but to post-hoc description. There is -no- predictive weight.

Well, a "post-hoc description" of empirical observations is how we make predictions. We gather cases (and their probabilities) post hoc, and predict future results based on that.

Why shouldn't it be applied in this case, especially when here not only we can calculate the probabilities post-hoc, but even use probability to pre-calculate those post-hoc probabilities for every series of streaks?

>The only way you can reasonably expect the past streak to have an impact on the next spin

Regression to the mean is a reason to expect past streaks to have an impact on future spins. Doesn't have to be some force that specifically affects the individual next spin, but it's a probability that affects (and moves toward the mean) the outcome of next spins.

Intuitively speaking, why shouldn't the gambler use it?

If they shouldn't, "because spins are independent" seems too hollow as an answer, because we have independent spins but also a statistical property for their aggregation.

I guess we could test it empirically: in a fair sequence of coin tosses, would a gambler always betting on tails after seeing N (for e.g. N > 4) consequtive heads have an advantage, over one always betting randomly on the same N+1 toss?

But regardless of that, my point is that intuitively is quite justified, and the explanations why it's not seem more hand-wavy.

>Try this on: What is the "mean" red-black on the roulette wheel? There isn't one.

Well, (ignoring non red-black case), there's not a mean value, but for the distribution there's an expected long term tendency of a balanced red and black outcomes.

>... And editing your comment to refer to groups of throws is disingenuous. It would have been more honest to modify your case in a reply.

I didn't modify my case, I expressed it more clearly. What exactly was disingenuous? I more often than not post a quick comment (HN habbit, as sometimes old edit forms expire), and then re-read it and quickly improve it as I re-read it, find typos, better wording, etc. Didn't even think anybody would read it and answer in the time it took me to refine it.

And, how about it, you can always answer to my final, improved, formulation. Does it have any historical interest if you answer to a worse worded case?

Re:Gambler’s Fallacy and the Regression to the Mean by psyc on Friday January 07, 2022 @11:46PM (#29848704)
HH is less likely than H (.25 vs .5), but HT is also less likely than H and by the same amount. HH and HT are equally likely. This extends to HHHHHH and HHHHHT, and so on. There’s no way to frame the gambler’s fallacy that makes it a better bet.

The probability that there will be one T somewhere in the sequence is a lot higher than the probability that there will be no T. That's only because there are a lot of somewheres for the T to be, but only one way for there to be no T. But once you have already seen 99 H, you still have precisely zero information about the next fair flip, and there's no way around it.

If you must make predictions about the streak as a whole, you have to lock them in before the first flip. Otherwise you piss off both Claude Shannon and Tony Soprano.

Re:Gambler’s Fallacy and the Regression to the Mean by coldtea on Saturday January 08, 2022 @11:26AM (#29853154)
>HH is less likely than H (.25 vs .5), but HT is also less likely than H and by the same amount. HH and HT are equally likely. This extends to HHHHHH and HHHHHT, and so on.

Well, kind of. There are 2^6 patterns with 6 tosses, and only a few of them have some recognizable pattern (are compressible). HHHHHH is the worst of them. It might be "equally likely" to any other 6-toss pattern seen from a fair coin, but it's also a great example to make one suspect a totally rigged (same sided H) coin.

Now, on average a fair coin would have roughly balanced H and T. So the (much more numerous and thus probable) series where H and T are mixed (e.g. HTTHHT..., HTHTHH..., etc...), are more probable than a large streak of H.

>The probability that there will be one T somewhere in the sequence is a lot higher than the probability that there will be no T. That's only because there are a lot of somewheres for the T to be, but only one way for there to be no T.

A, my sentiments exactly!

>But once you have already seen 99 H, you still have precisely zero information about the next fair flip, and there's no way around it.

That's my argument: you don't have zero information.

You know that the next fair flip is random - sure.

But you also have the information that large series of tosses end up balanced in the long run, which is a kind of push (given fairness) towards making a T after tons of Hs increasingly likely.

Why should this be discarded?

Re:Gambler’s Fallacy and the Regression to the Mean by psyc on Saturday January 08, 2022 @12:28PM (#29853855)
> Why should this be discarded?

Well, it's a restatement of the Gambler's Fallacy, but I'll taboo using that as an argument because it begs the question (but for reals!)

I thought about this for a while, and here's the best I can do for right now. It's like winning the MegaMillions lottery on Saturday, and then winning $1 on a scratch ticket on Sunday. You'd go running around screaming on Saturday, but you wouldn't on Sunday. (Would you?) Because the impossible already happened. Hitting $1 on a scratch-off is never remarkable (unless you're 8), including on the day after you won the MegaMillions. It's just a somewhat anti-climactic coincidence.

Re:Gambler’s Fallacy and the Regression to the Mean by chrismcb on Saturday January 08, 2022 @12:15AM (#29848829)
After successive black streaks about the only thing you can say is maybe the wheel is biased toward black. Assuming no green, and a perfectly unbiased wheel, black if a fifty fifty chance. Which means the next infinity sounds half will be black. If there were three blacks in a row, it doesn't mean red is more likely, as the rest of the spins half will be black.
Re:Gambler’s Fallacy and the Regression to the Mean by coldtea on Saturday January 08, 2022 @10:45AM (#29852701)
>Assuming no green, and a perfectly unbiased wheel, black if a fifty fifty chance. Which means the next infinity sounds half will be black. If there were three blacks in a row, it doesn't mean red is more likely, as the rest of the spins half will be black

Well, the point is:

"black if a fifty fifty chance" and "the next infinity sounds half will be black"

doesn't sit well with:

"if there were three blacks in a row, it doesn't mean red is more likely".

The roll is independent, but in some sense either the roulette is unfair or red will increasingly be more likely.

In other words, since there is a statistical property working across many/infinite rolls (regression to the mean, black/red equally likely), this property makes sense to also somehow affect the possibility of individual rolls (as individual rolls are still rolls within an aggregate).