Quote:
Originally Posted by #1ThunderFan
Thank you Mpethy. This was much needed and very very helpful. Awesome work as usual.
Any chance of discussion on at what point (# of hands wise) should we start trusting our hud stats? Mostly for the commonly used ones. I am trying to figure out the amount of hands I need for each stat to determine when it will be significant enough to be useful and should be considered in decisions.
Maybe a math guy can come in here and answer this question with hard numbers, but to my way of thinking, the important point here is that the question iyou asked is the wrong one. It betrays an unreasonable desire for certainty in a game based on uncertainty.
The better way to approach HUD stats is to ask this question: When I have a small sample, how do I use my HUD stats to best advantage?
You have to start considering your HUD stats as soon as you have
any sample.
Using them and
relying on them are two completely different things. Any information is better than no information, so even a small sample is better than nothing, provided that you are considering them, and not acting in blind reliance on them.
With that in mind, here are my rules of thumb.
VPIP: I read a post a few years ago, that I have been unable to find ever since, that proved mathematically that VPIP was reliable enough to use after 40 hands--it has to do with bell curves and confidence intervals. I am not a hard core math guy--I took a minor in stats 20 years ago, lol, so I can't replicate the math, but this makes intuitive sense. Sure, it is possible that the player is an 8/6 who just got dealt AA-JJ 14 times in the last 40 hands, but how often does that actually happen? Realistically, not very often.
Moreover, we don't really care that much whether the player is a 9/7 or an 8/6; similarly, we don't care that the player is actually a 27/10 while our HUD is showing him at 32/8. In either case, we can fairly accurately categorize and adjust to the player despite the discrepancy between what we see on our HUD and his actual VPIP.
Also, you need to use a little common sense, because VPIP is useful even before you have 40 hands on the villain. There's an example of this thinking in my well. I raised on the button first in and a 2+2er in the BB, against whom I had no hands (it was his first hand at the table) 3 bet me. Now ask yourself, what happens more often--that a 2+2er will make a play against a fairly obvious steal or that a 2+2er will will get dealt a top 5% hand his first hand at the table? the answer is obvious, so in this case, a sample size of 1 hand was sufficient information to use his VPIP.
So when you see a player who his splashing around and his vpip is 80/60 after 10 hands, ask yourself--what is more likely--that this guy is getting slapped in the face by the deck or that he is a live one just splashing around? The answer is obvious.
So this is going to be a common refrain throughout this post--mathematically, if that poster years ago was right, VPIP is fairly normalized after 40 hands; but like every other stat, it is useful long before it normalizes. And if anybody is looking for mathematical certainty at the table, they're going to be disappointed A LOT, and I suggest a different hobby.
OK, so VPIP becomes reliable fastest, because it is a stat based on every hand the player is dealt. Thus, all of the other stats will normalize more slowly, since they are not based on every hand the player is dealt.
PFR: PFR is also described as a percentage of every hand the player plays, but really it is more useful to consider as a proportion of the hands the player volunteers to put money in the pot. So after 40 hands, you're going to have a reasonable amount of confidence in this stat, too, but its relationship to vpip probably takes a while longer to stabilize. Recently, I had a player in my DB against whom I had ~800 hands and his VPIP/PFR was 14/4 in that sample. By the time he hit 1000 hands, though, he was at 14/9 or 10, so weird things do happen, and I had to mentally recategorize that player by the end of the session in which I watched his PFR skyrocket from 4 to 10 or whatever. When I have 100 or 200 hands on a payer, i feel very comfortable assuming that his VPIP and PFR in my sample are reasonably close to his actual long term VPIP and PFR. The recent example is the only one I can recall drastically changing after I had a good sample on the player (excluding players who changed their styles).
3 bet% Every time a player has the opportunity to 3 bet, it is an opportunity included in his 3 bet % in HEM. So this stat is very volatile in small samples, and must be used with caution. It probably takes a big sample to normalize--no way will I hazard a guess on just how big a sample. But there are some ways to think rigorously applying common sense, rather than math. You're faced with an unknown villain, against whom you have fewer than 200 hands. He 3 bets a raise. What do we know? Well, we know that a typical value 3 betting range is AK and QQ+; depending on the situation, it may be wider. This range is 2.6% of all hands; phrased more usefully, he will get dealt a value 3 betting hand 2.6 times in every 100 hands. So suppose you have 10 hands on this guy. We know that a sample of 10 random hands will contain .26 value 3 betting hands. So we know that this guy's 3 betting frequency is higher than random chance accounts for*. Does that mean we know he is light? No, of course not. If we get dealt a value 3 betting hand 2.6 times per 100, we are getting one every ~40 hands or so. So if a guy 3 bets once in his frst 20 hands, it's a coin flip as to whether he is value 3 betting or not. But this isn't bad; using the remainder of the situation as a tie breaker, you can easily decide which is more likely--If he is UTG+1 3 betting a 12/9 UTG, he's probably not light, but if he is in the BB facing an apparent steal from a guy with a 30% ATS, he probably is.
So we don't know that he is light, but it does make it more likely that he is light than if this was the first 3 bet you saw in 100 hands. So if you are the one he 3 bet, you factor this into your decision--then you look at the situation, your image, his vpip/pfr (if he is 90/80 at that point you'll make a different decision than if he is 10/10) you make a read and you go with it (or I suppose you could play it safe and fold and keep your eye on him, alert to the possibility that he is 3 betting light).
*3 bet % is based on the number of opportunities you have to 3 bet; thus, to determine somebody's 3 bet % who only ever value 3 bets these hands you have to do 2.6 * % of hands raised in front of him. So if we get dealt a value 3 betting hand 2.6 hands per 100 hands dealt, and, on average, 40% of hands are raised in front of us we will be value 3 betting 40% of 2.6, roughly 1% of the time (I have no idea how accurate that 40% figure is, it's just a for instance). But since we know that the percentage of hands raised in front of us is less than 100, we know that a 2.6% 3 bet % is a fairly tight 3 bettor, who is rarely 3 betting very light at all.
The lesson here is that something in the neighborhood of 2.6% is
sort of a magic number. If a villain is 3 betting much more than this, he is representing getting hit in the face by the deck. You need to decide based on other factors whether you believe this representation or not. But you can start using his 3 bet % as important information the very first time you see someone 3 bet--just consider it compared to a value 3 betting range and the number of hands you have on the guy, factor in the situation and act accordingly.
If you are looking for mathematical reliability, first, don't bother. Second, if you ignore that advice, then just do some easy filtering for the stat you are studying.
For example, to figure out how many stealing opportunities you get OTB, just look in HEM on the position page to see how many hands you had OTB, then go to main filters/preflop action facing player and click off everything but "unopened." This usually runs at about 25% in my DB. So 25% of 1/9th of your total hands are opportunities to steal OTB. In a sample size of 1000 hands, then, you have maybe 27 stealing opportunities. if i want a sample size of 50 steals before I
rely on the stat, I need about 2000 hands.
If you're a TAg, you raise about 12% of your hands. You see the flop about 40-45% of the time you raise (filter PFR=True, note the total hands, then add saw flop = true, note the fraction of pfr hands, easy, ldo). You will have the opportunity to c-bet maybe 85% of the time you see a flop (add filter "could c-bet = true for the actual number) as the preflop raiser. So if you want to have a sample of 40-50 c-bet opportunities before you rely on a TAg's c-bet %, you'll need to have over 1000 hands on him--120 raises, ~51 see the flop, ~40 opportunities to c-bet.
You can go through this thinking for every stat you want to, but it is a waste of time, because you should be considering these stats long before that, and you should have specific notes on players long before you have 1000 hands on them.