Nice try. But there's not one lick of science in any of the "methods" or "strategies" that people on this forum come up with as proposed alternatives to Kelly. That's not to say that such alternatives don't exist (see, for example, Ganchrow's explanation in this prior thread), but simply asserting that "I've found flat betting to be a much better way to increase my bankroll" is not science.
-
Comment -
isn't there a newer version of kelly?Comment -
OK, I'll kick this off with an initial, simple thought pattern. Feel free to laugh, or add constructive ideas.
The underlying question is: how do we deal with small sample sizes, and fluctuations (changing chapters) within a season? If this question is of no interest to you, please stop reading here.
The soft spot in Kelly -with regards to everyday sports betting - overlaps with the weakness of Z-scores. Unless a play is purely mathematical, the sample size must be so large that it typically covers several seasons. For this, the world of sports betting is simply too changeable. Many opportunities will pass us by if we have to wait that long. When a season transits into its next 'chapter', such as, for instance, the NBA after the All Star Break (until the playoffs), the environment changes. Z-scores are too slow to keep up with such changes, and Kelly becomes inaccurate (with tremendous risk to the player). Action points can be part of the solution to solve this problem.
Action points have the advantage (over Z-scores) of giving added depth to short term results. This is important, because it allows you to be far more nimble than if you had to wait for a sample size to become 'Z-score acceptable' (which is often too slow for the changing environments during a season).
Example.
Overlays for NBA totals.
I had previously determined that I needed a 20 pt overlay (or technically, in this case, underlay) for a play. So far, this season, for UNDERS, I have a 10-7 result for overlays 15-19 pts and a 7-4 result for overlays 20 pts and higher. Both 10-7 and 7-4 are up 3 units. But the first sample (10-7) is MINUS 23.5 action points. The second sample (7-4), for the predetermined 20+ pt overlays, is PLUS 78.5 action points. In the first sample, the three excess wins are 'borrowing' almost 8 pts per win against the spread. In the second sample the three excess wins are beating the spread by 26 pts. In other words, while both samples are small, it is reasonable to expect the first sample to start evening out, and the second sample to improve.
To the extent that we can weed out luck, we develop a better understanding of our true edge. Once we can do this, we can be much more specific about bet sizing.
For the above 7-4 sample the Z-score is 0.90. No bet.
But if we change the formula from excess wins/sqrt sample size to: (Action Pts/excess wins)/SqRt sample size, the result for this tiny sample is: 7.89.
The same calculation for the 10-7 result gives: -1.90.
Just same early ideas. The final result, at the end of the road, could be league-specific scales, for sides and totals, that translate into bet sizes. The other final result could be ... nothing. Either is fine with me.
(odds of sect members to mention data mining: -425)Last edited by Dark Horse; 12-31-10, 08:14 PM.Comment -
i learned one thing from this thread. if you ignore flying dutchman, he will go away and try to find another thread to trollComment -
Oh yeah, almost forgot; DH, I've found that action pts are very useful when evaluating teams, not really sure about techniques. I use a combination of z-score, avg line and average score dif or total for each technique. Based on my super secret formula, I can get an idea of how much the market is on to each situation. I'm adding in action pts based on your post to see how it works. Thanks.
Comment -
Kelly (I am including fractional Kelly) used by pro ap's in all forms of gambling.
When I started reading the thread, was going to post "In before Wrecktangle throws the term 'Bayesian priors' into the discussion ", but I see that I'm way too late.
Flying Dutchman - "BTW, bonehead: fractional Kelly IS NOT OPTIMAL by Kelly's def, so it's just another betting system that you blowhards are waxing on about."
Fractional Kelly is calculated by first calculating Kelly. So without getting into the semantics of "it's just another betting system", if you are acknowledging the usefulness of fractional Kelly, you are acknowledging the usefulness of Kelly.Last edited by Scooter; 01-01-11, 02:19 PM.Comment -
I'm looking at this to see if there's a way to combine small samples with the idea of maximizing profit (Kelly). That is an area that has been overlooked, in my opinion. Small samples, and quickly changing environments, are a reality in sports betting (although the sect seems to find the idea of different chapters within a season to be amusing). In the example I mentioned earlier, one could treat the sample as 17-11 (10-7 and 7-4), and come up with an entirely wrong bet size. The added definition provided by action points could help prevent such mistakes.
I understand the risks of attaching too much value to small samples, but simultaneously am interested to see if there could be corresponding reward if that field is treated with greater precision.Comment -
Again, I applaud the effort, but it's obvious you have a misunderstanding of basic statistics. Yes, the 7.89 number you computed is a lot bigger than the conventional z-score 0.90. But that doesn't mean it's significant (in a statistical sense), which would give you a basis for believing your model -- because you haven't figured out the correct variance around your "action points" (which in practice would require a large data set, something you're apparently not willing to construct over time). The variance around action points is NOT the same as it is around mean wins, which is what your calculation implies.
You can't just make shit up because you don't like the fact that a larger sample than you would like is needed to make statistical inferences.Comment -
OK, I can see we have a bunch of business majors here. If any of you "pointy-heads" had had a lick of classical Math (you gotta prove some theorems, folks), YOU WOULD KNOW that the mountain is high when you state that you have a PROOF of anything and it is higher yet for optimality.
Fractional Kelly IS NOT KELLY!!!!!!!!! Fractional Kelly is NOT OPTIMAL because it is NOT KELLY!!!!!!!!!
Therefore, it IS JUST ANOTHER BETTING SYSTEM!!!!!!!! Fractional Kelly NOT OPTIMAL BECAUSE IT'S NOT KELLY, get it Business Major??????
One other thing, wizards. None of you guys seem to have noticed I mentioned Kurt Gödel, so I'll repeat. Visit this link:
I doubt you clowns'll understand it, but it tells you that there is a lot you CAN'T prove. Add in all the variability of sports stats, and "FULL" Kelly is there (even if it were provably optimal, which I contend it was not).
Comment -
You are so full of shit it's not even funny. I'll say it again, go read Ganchrow's post that I mentioned earlier (#10 in this thread). Long and short of it is that the original full Kelly formula that everyone here takes as gospel is optimal UNDER A VERY SPECIFIC ASSUMPTION about risk aversion, while as Ganch stated:
" "fractional" Kelly utility expands on full Kelly to include the entire class of CRRA bettors with nonnegative relative risk aversion coefficients". I'm not going to translate for you, if you can't understand the gist of what he's saying, you have no business commenting on Kelly optimality.
Quit pulling crap out of thin air when you have no clue what you're talking about.Comment -
I just love it when you Business Majors hide under Momma Ganch's skirts.
Go read Kelly's proof. But better yet, read Gödel.
...besides you wusses couldn't find a "non-negative relative risk aversion coefficient" if your bankroll depended on it, and then you expect the noobs to figger it out?
Comment -
Here's a "non-negative relative risk aversion coefficient": 1 (equals full Kelly)I just love it when you Business Majors hide under Momma Ganch's skirts.
Go read Kelly's proof. But better yet, read Gödel.
...besides you wusses couldn't find a "non-negative relative risk aversion coefficient" if your bankroll depended on it, and then you expect the noobs to figger it out?
Here's another: 0.5 (equals half Kelly)
With virtually every message you post, it's more and more clear that you don't even know what you don't know.
I think this thread has reached the end of its useful life.Comment -
As mentioned, it was just a first outline.
I do have plenty of models with long term action points. This effort is intended to see if there is something more 'nimble'. I believe there might be. But it requires more thought.
I expect variance around action points. One possible thing to look at is to keep only the most recent small sample size (size to be determined), in order to be more 'awake' to seasonal fluctuations. (when we look for imbalances, we can expect self-adjusting mechanisms that seek to restore balance).
As to 'making things up'. Imagination? I'm willing to be wrong 99 out of a 100 times. But that 1 little overlooked gem is what I'll write a report about. And those reports, together, give me my edge. It may be unorthodox, but it's more fun (to me).Last edited by Dark Horse; 01-02-11, 12:25 AM.Comment -
Comment -
Every Kelly thread that Flying Dutchman enters deserves to have this quote attached to it to show his omniscient comprehension of the Kelly Criterion. Quite the mathematician.Bullshit. Have you read Kelly's proof? It's about win percentage ONLY. AND there is only ONE formula. There is not one for Blackjack, and one for football, and one for baseball, etc. Get your shit straight.
...now as for folks getting cocky, you never do, do you?
Comment -
Monkey, I just got a PM from Justin7 telling me to be nice to you and the other members of the sect or he'll sanction me.
Apparently you and your fellow members can get away with all kinds of personal attacks, whereas I get threats from mods.
So, I'm backing off only because you've called in your protector.
...and I'm out of the Tank.
...have a nice life.
Comment -
I talked to a true math genius and friend of mine (by that I mean someone who does make millions off this knowledge in the business arena). He encouraged the idea of using action points, in the manner suggested above, and pointed me towards Chi Squares (somewhat ironic, because Ganch did the same a few years ago) as well as a program that is better than Excel (the name of which I already forgot... lol).
FD, hang around. Just laugh at the territorial instinct and predictability of the insults by the sect. Let's just say they're good at poisson props and march madness brackets.
On a sidenote, I'm looking for a good biography or background on Siméon-Denis Poisson. My friend explained to me how widespread his distribution is used, in a huge variety of fields. Incredible! A giant of science.Last edited by Dark Horse; 01-03-11, 03:36 PM.Comment -
To be fair, I thought there was one equation that is suitable for Blackjack and sports betting. It changes based on the payout odds, but it is pretty good in both.Comment -
Action points is something reasonable to have in your toolbox for testing purposes. It's not a perfect magic bullet answer or anything though.Comment -
From Wikipedia, this is the one I recognize:
f* = (bp-q)/b
where:
f* is the fraction of the current bankroll to wager;
b is the net odds received on the wager (that is, odds are usually quoted as "b to 1")
p is the probability of winning;
q is the probability of losing, which is 1 − p.
Which is the other that you use?Comment -
f* = a/v
a = Advantage
v = Variance
Typically, you simply multiply the inverse of the variance (something like 0.76) with your advantage to easily calculate your stake. I'm not quite sure how you'd plan on calculating the standard formula at the tables unless you force yourself to memorize a staking table.Last edited by MonkeyF0cker; 01-05-11, 09:21 AM.Comment -
I never thought of that as a separate Kelly formula, but you are correct. Blackjack variance is roughly 1.3, and I always divided the max stake by that amount... But the theoretical discussion on maximum bet-sizing on Blackjack is relevant in very few casinos. For most serious players, longevity dictates max bet more than Kelly.Comment -
Better out than in-sect. Nevertheless, a sad comment on the state of affairs in this place. Haven't seen Wreck in these parts for a while either.
FD, you do not have to be 'nice' to the sect members. They take a peculiar pride in treating everybody outside of their little group like horse sh*t. I hope that non-in-sects understand that this is not called the math tank, but the think tank. And if anybody thinks that math people are good thinkers outside of their expertise, think again. They cover a narrow margin of the overall field of sports betting, and confuse it for the whole. That's fine. To each their own. But if their misplaced arrogance starts to result in more original thinkers leaving or not posting here, something is not right.Last edited by Dark Horse; 01-05-11, 07:29 PM.Comment -
For blackjack to be worth it, you aren't gonna be able to get away making sporadic max plays especially in person if you're a pro (unless you have the perfect act to go along with the game), but more importantly maintaining an element of ruin of 1/20th or less...or am I on the wrong boat?Comment -
You're correct. The risk of getting tossed goes up to an unacceptable level long before you're betting your max kelly bet (in many outs, including Vegas). There are always exceptions -- like New Year's Eve or other crazy times when it is crowded any anything goes... But most days, even a 100k bankroll can't bet it's kelly max bet at most Vegas casinos.Comment -
Ronald Fisher puts Poisson to shame. His radical thinking was second to none (he was essentially what inspired Zahavi's "handicap theory" for "good genes"). He had a grasp of the mutation rate of genes before anybody knew about genetic mutations, which most scientists would say follow a Poisson distribution. (More specifically it would be related to generation turnover and the relative tradeoffs between metabolic copying mechanisms and possible deleterious mutations, therefore would not necessarily be random but part of a chaotic system, which are fundamentally two different ideas, but still virtually impossible to pinpoint at one given moment, fascinating stuff)
For football, once the season moves along, I find the aggregate z-scores for pythaogrean points for and against over against the spread record, and see where each team stands, along with a league wide correlation of point differential vs against the spread record as an additional component to serve as a check to any pythagorean discrepancies. Then use that as a base filter along with supplemental filters (maybe average line and points per play or point), but such calculations are hard to define an edge for a given game, rather the intensity for which a particular side deserves an investment.
IMO, its easier to find a quantified edge for baseball, but harder to determine the Kelly Stake because of the number of simultaneous events every single day. Scarcely is a person sophisticated, deliberate, patient, or sedentary enough to sit there and calculate the Kelly Stake for a game that starts at 1:30 based on the outcome of the first 30 minutes of the Nationals game at 1 and the current Pinnacle line, then log on to Matchbook and 5Dimes to compare lines to play the best available line. Unless you have written a script to automatically take care of all of that; gathering the data, calculating edge and simultaneous Kelly Stakes at each point in time based on wagers previous, then sign on to whichever website automatically to place a wager that is in cell C56 in Excel.
The opportunity cost would have been exhausted in the first 5 minutes had you not written a program to automate everythingComment
