Kelly idea is important but kelly formula less so

[specialronnie29]

all this talk of kelly has been pretty silly

the idea of maximizing expected growth is obviously important but people slaving over the correct kelly formula and worrying about whether to bet 3.4% or 3.5% is dumb. the reason is you never know your ev exactly. there is never one perfect model. one model might predict of a total of 40 and another a total of 40.5 or 41. If the market total is 43 you know you have a +ev bet but how ev is it exactly? depends which you take to be the true value - 40 or 40.5 or 41.

real pros bet in the kelly area and bet some amount on premium plays and some other amount on regular plays. if on average the market is off by 2 pts they bet $x and if the market is off by 3-4 points or more then they bet $y

and for those where limits are effectively binding then kelly is even more irrelevant

[turnip]

real pros bet in the kelly area and bet some amount on premium plays and some other amount on regular plays. if on average the market is off by 2 pts they bet $x and if the market is off by 3-4 points or more then they bet $y

Betting $x, where x is a fixed number independent of bankroll amount, doesn't just ignore the Kelly formula, it ignores the whole Kelly concept.

[That Foreign Guy]

> and for those where limits are effectively binding then kelly is even more irrelevant

Huh? If the limit is smaller than the kelly fraction it doesn't matter what the kelly fraction is.

[specialronnie29]

Betting $x, where x is a fixed number independent of bankroll amount, doesn't just ignore the Kelly formula, it ignores the whole Kelly concept.

you're right but i meant this for the short run. as your br grows, $x increases. but day to day or week to week or maybe even between months it doesnt.

[specialronnie29]

> and for those where limits are effectively binding then kelly is even more irrelevant

Huh? If the limit is smaller than the kelly fraction it doesn't matter what the kelly fraction is.

"Limits binding" means you are constrained by the limit which means you want to bet more than the limit which means kelly is irrelevant in the sense that you can only bet the limit. this assumes you're sure the kelly amount is above the limit. for some smaller markets this is obviously the case for some bettors (props for example)

[Thremp]

Your post is very bad.

[specialronnie29]

Your post is very bad.

idiot

why do you bother posting this when you have nothing to say

[roasthawg]

I agree with the op in that it's more important to understand the concept of compounded returns rather than utilizing the exact kelly formula.

[specialronnie29]

even if you have the exact perfect model, there is still the question of what is the right time period of data to use. this is a problem that is talked about in the think tank every second week it seems. it is particularly problematic for NFL

If your model says the total should be 40 according to last 5 years of data, and it should be 41 with the last 10 years of data, then you know U43 is +ev but cannot measure by exactly how much.

like i said real pros dont run to a calculator and calculate the kelly stake using each model and then go in between, or pick the smaller one or the bigger one. their bankroll doesnt change much from day to day and they know vaguely what the right amount is. then as their br increases they start betting more, but it is not a continuous exact increase

i can imagine some idiots here taking an extra 20 seconds to calculate the right kelly stake on a bet only then to see the line move against them.

[Inspirited]

real pros

[JustinBieber]

If you ignore kelly then most likely you will be a lot poorer than you should be.

[Thremp]

So what exactly is your point? To use two tiered fixed betting?

How does this jibe with the idea of KC and maximizing median geometric means? (Hint: It doesn't.)

Basically you posted a mathematically and logically unsound critique of KC and then made the arbitrary conclusion to use a two tiered fix % stake system based on some circular logic utilizing Bayes Theorem (or the Pokerjoe Uncertainty Principle).

How is that anything but "very bad"? I suppose very may be an exaggeration, but we're just quibbling now.

[MonkeyF0cker]

even if you have the exact perfect model, there is still the question of what is the right time period of data to use. this is a problem that is talked about in the think tank every second week it seems. it is particularly problematic for NFL

If your model says the total should be 40 according to last 5 years of data, and it should be 41 with the last 10 years of data, then you know U43 is +ev but cannot measure by exactly how much.

like i said real pros dont run to a calculator and calculate the kelly stake using each model and then go in between, or pick the smaller one or the bigger one. their bankroll doesnt change much from day to day and they know vaguely what the right amount is. then as their br increases they start betting more, but it is not a continuous exact increase

i can imagine some idiots here taking an extra 20 seconds to calculate the right kelly stake on a bet only then to see the line move against them.

Do you normally model the same game with two different datasets? I think you may have more problems than attempting to deduce the viability of Kelly with pseudomathematics.

[specialronnie29]

So what exactly is your point? To use two tiered fixed betting?

How does this jibe with the idea of KC and maximizing median geometric means? (Hint: It doesn't.)

Basically you posted a mathematically and logically unsound critique of KC and then made the arbitrary conclusion to use a two tiered fix % stake system based on some circular logic utilizing Bayes Theorem (or the Pokerjoe Uncertainty Principle).

How is that anything but "very bad"? I suppose very may be an exaggeration, but we're just quibbling now.

On paper no one is criticizing the Kelly criterion. When you know your exact EV and bankroll and there is no hurry to make a bet of course you use kelly.

This is different. you ignored everything i said about modeling discrepencies making it unclear what is your best prediction of exact EV. i said in REALITY due to PRACTICAL considerations, a lot of people bet in a 'two tiered' way. The reason is that often their edge on games they bet is about the same (lets say 2 pts), and they know that the optimum stake is somewhere around 3-5% of BR. Then if a line pops up that gives them a 4 pt advantage, they know that approximately the best bet is to double their normal bet. And from week to week, the term BR typically hasnt changed much. And if it does we all know to increase our bet size proportionately, but again it is usually a ballpark figure.

syndicates run into this problem all the time. with runners everywhere with a number theyre looking for they often get too much or too little action on a game just due to coordination problems. so why do they allow these mistakes to happen? because the opposite problem - of hesitating and missing a great number - is way more costly.

this has nothing to do with proving kelly does not maximize EG. it is all about the fact that calculating an optimal bet to the decimal point is silly and impractical. Try doing this with 2h lines changing lightning fast and re-running your kelly calculator over and over

[specialronnie29]

Do you normally model the same game with two different datasets? I think you may have more problems than attempting to deduce the viability of Kelly with pseudomathematics.

I like to know where my 'specific' number is coming from. Let's go with an NBA 2h total. It could be with 09-10 data, my number for an NBA total is 100 and with 07-08 and 08-09 data it is 99. I think the 09-10 data is most predictive, but feel more confident when there is a consensus or the market line is significantly different than both 99 and 100. If the posted total is 103, what is my EV?

Usually in my experience, the total is too close to my model for me to bet it (which means no play), or it's 2 pts off (which means a standard $x bet), or it's 3 or 4 pts off on occasion (which either means no play because i may not have all info, or it means a $1.5x or $2x bet). Whether the under 103 is -110 or -105 or -106 doesn't change my bet size much, it makes such a small difference on EG it isn't worth the hassle and it can just make you miss a bet at a juicy number.

again, i am NOT SAYING KELLY IS WRONG IN THEORY. just think some guys here are using it to make them feel like scientists who know everything about the right number and its distribution, etc., and are solving a pen and paper textbook problem. remember all we know is the BEST GUESS at the true number not the true number.

[bztips]

I like to know where my 'specific' number is coming from. Let's go with an NBA 2h total. It could be with 09-10 data, my number for an NBA total is 100 and with 07-08 and 08-09 data it is 99. I think the 09-10 data is most predictive, but feel more confident when there is a consensus or the market line is significantly different than both 99 and 100. If the posted total is 103, what is my EV?

Usually in my experience, the total is too close to my model for me to bet it (which means no play), or it's 2 pts off (which means a standard $x bet), or it's 3 or 4 pts off on occasion (which either means no play because i may not have all info, or it means a $1.5x or $2x bet). Whether the under 103 is -110 or -105 or -106 doesn't change my bet size much, it makes such a small difference on EG it isn't worth the hassle and it can just make you miss a bet at a juicy number.

again, i am NOT SAYING KELLY IS WRONG IN THEORY. just think some guys here are using it to make them feel like scientists who know everything about the right number and its distribution, etc., and are solving a pen and paper textbook problem. remember all we know is the BEST GUESS at the true number not the true number.

I don't disagree with you. But Kelly becomes more and more important the more variation there is in your estimated EV's. For example, this past year my baseball bets (based on Kelly) were anywhere from 1% to 15% of my bankroll. In that situation, I would potentially be leaving a lot of EG on the table if I only varied my bets between 1x and 2x.

[specialronnie29]

I don't disagree with you. But Kelly becomes more and more important the more variation there is in your estimated EV's. For example, this past year my baseball bets (based on Kelly) were anywhere from 1% to 15% of my bankroll. In that situation, I would potentially be leaving a lot of EG on the table if I only varied my bets between 1x and 2x.

Fair enough but I figured most were in the category where those bets were rare, ie. in bigger markets

in props and stuff by all means vary your bets. and also you have time to calculate it when you're staring at a prop board. right before game time or during 2h there isnt much time. unless it's SIA or bodog, juicy game lines wont last for very long.

[roasthawg]

I agree with the op in that it's more important to understand the concept of compounded returns rather than utilizing the exact kelly formula.

I'll revise this by adding that without the Kelly formula I probably wouldn't have come up with the 2% number which is the percentage of my bankroll I bet on every play. Edge is such a hard thing to quantify in sports gambling that I make the assumption that my edge is always small... even when my model tells me it's large.

[Thremp]

How do you know you have an edge if you can't estimate the size of the edge?

[MonkeyF0cker]

I like to know where my 'specific' number is coming from. Let's go with an NBA 2h total. It could be with 09-10 data, my number for an NBA total is 100 and with 07-08 and 08-09 data it is 99. I think the 09-10 data is most predictive, but feel more confident when there is a consensus or the market line is significantly different than both 99 and 100. If the posted total is 103, what is my EV?

Usually in my experience, the total is too close to my model for me to bet it (which means no play), or it's 2 pts off (which means a standard $x bet), or it's 3 or 4 pts off on occasion (which either means no play because i may not have all info, or it means a $1.5x or $2x bet). Whether the under 103 is -110 or -105 or -106 doesn't change my bet size much, it makes such a small difference on EG it isn't worth the hassle and it can just make you miss a bet at a juicy number.

again, i am NOT SAYING KELLY IS WRONG IN THEORY. just think some guys here are using it to make them feel like scientists who know everything about the right number and its distribution, etc., and are solving a pen and paper textbook problem. remember all we know is the BEST GUESS at the true number not the true number.

So less data = more predictive because you arbitrarily THINK so? Okay then. I think that sums up this thread pretty nicely.

[specialronnie29]

So less data = more predictive because you arbitrarily THINK so? Okay then. I think that sums up this thread pretty nicely.

No

it is my gut instinct that says the latest season's data is most predictive, but i do not act based on that belief. it was just an aside. You model nfl - you have probably at some time had to make a GUT choice about how far back in the data to go. Do you have 1982 data in your nfl models? if i thought it was the best predictor i would ignore my 07-08 and 08-09 data entirely.

if my 09-10 data says total is 98, the 07-08 and 08-09 says it's 99, and the posted total is 99 -105, i do not make a bet. i wait for a consensus so that when i make a bet i at least know that it's not a single year of data driving it.

monkey i admit i write like an immigrant but i think youre doing a bad job reading. the sentence that sums this thread up best is:

'because the edge is unknown, and the kelly stake requires an exact number for the edge, religious adherence to kelly to the penny is silly and often impractical'

[Thremp]

When math fails, gut instinct.

[specialronnie29]

How do you know you have an edge if you can't estimate the size of the edge?

sometimes you dont. if one model says total of 98 and another says a total of 99, and posted total is 98.5 i lay off.

if the predictions of the 3 models are 98, 98.5 and 99 respectively and the total is 102, i still dont know exactly what my edge is but i know i have one.

you really dont get it thremp. do you actually bet sports?

[Thremp]

I don't model. Clearly one (or both) of your models is inaccurate. Do you understand why this would be a failure of the constraints of KC? (IE: Must know your edge.) Or why this would fail in a practical application of knowing/estimating your edge.

We can all safely say that there is no point in discussing which staking method to use when you have no edge as the optimal wager size is 0. So we must assume that you have an edge, and then if we can assume you can determine the binary to some degree of accuracy, then we must be able to quantify this edge to some reasonable degree of accuracy.

Your posts are misguided and ill informed. This pervasive example of the "specialronnie Multiple Model Theorem" makes little sense to anyone I've spoken to who models, and little sense to me as a button clicking muppet.

[specialronnie29]

you can have multiple models (ie methods) in addition to multiple data periods

here is an example to show you why you're wrong, or why youre misunderstanding me

What is the value of buying a half-point off the 7 in mlb? Well most people take data going back 10 years and see what fraction of ALL games landed on the number 7. from that a cent-value is calculated. that is one model. here is another model. you take the same data going 10 years back except you calculate what fraction of games WITH A TOTAL OF 7 actually land on 7. You will get a different answer. And if instead you only go back 5 years in the data (whether because you think the game has changed since the steroid era or for some other reason), you will get another answer. So which model is better? And how far back in the data should you go? to year 2000? 1995? 1960? this is not much of a problem in MLB because there are so many games played, but it is a big problem for the NFL.

this is the same thing for NFL when calculating the value of buying off the 3. you can see how frequently ALL games were decided by a difference of 3, or only when the FAVORITE won my exactly 3, and you can even break it down further by different totals. buying off the 3 is worth less if the total is higher. so when faced with a line of -3 -100 and a line of -3.5 +1xx, which is the better value? if -3.5 has better value, what is your precise and UNIQUE edge over betting -3 -100? there is not a unique answer that trumps all the other ones and you're deluding yourself if you think otherwise.

sometimes the different answers you get are not significantly different. that's a nice situation. if they are vastly different then something is wrong. but often they are close but a bit different, and unless you have a good reason to choose one model over the other, you need a consensus among the models to know you have an edge.

here's another example if you still haven't got it. you have in front of you a coin that is not necessarily fair. at the moment only god knows the true probability of it landing on heads. you flip it 1000 times in your bedroom and 70.4% of the time it lands on heads. the next day you do the same thing on a street corner and it lands on heads 70.2% of the time. some guy comes along, sees your coin, and offers you a line of Heads -110 and Tails -110. you know you have an edge on Heads but what is your edge? Is it based on 70.4%, 70.2% or 70.3%? Maybe the surface of the street corner is slightly less conducive to heads. Some people would say 70.3% and others 70.2%. What is the optimal kelly stake now?

ronnie's pt:
you dont know, it doesnt matter as much as you think, and if you wait around too long humming and hawing the guy may change his mind and walk away

[Data]

you can have multiple models (ie methods)...

...What is the value of buying a half-point off the 7 in mlb? Well most people take data going back 10 years and see what fraction of ALL games landed on the number 7. from that a cent-value is calculated. that is one model. here is another model. you take the same data going 10 years back except you calculate what fraction of games WITH A TOTAL OF 7 actually land on 7. You will get a different answer.

From what you were previously saying, I had strong suspicions that what you call "a model" is not something what even resembles one. Here you just spelled it out. That confirms my suspicions.

Regardless, you do have a valid point about the limits though. Another valid, in a twisted way, of course, point is that with the "models (ie methods)" as described above, Kelly staking strategies are useless.

[roasthawg]

How do you know you have an edge if you can't estimate the size of the edge?

I can easily estimate the size of my edge... I choose to ignore size estimates though and simply focus on whether or not there is an edge. My models are not perfect... far from it. Ignoring edge size is one of the ways I attempt to overcome these imperfections. My figuring is that some of the games that look to have a 4% edge actually have a 2% edge and vice versa. Some of the games that look to have an edge will actually be coinflips. So when you ask "how do you know if you have an edge"... for me the answer is that I don't, at least not at the individual game level. But overall, after hundreds of games, my models do tend to pick winners at around a 53% clip. So for me it makes more sense to estimate my edge for all of my bets and set my bet amount at 2% of my bankroll.

[specialronnie29]

From what you were previously saying, I had strong suspicions that what you call "a model" is not something what even resembles one. Here you just spelled it out. That confirms my suspicions.

Regardless, you do have a valid point about the limits though. Another valid, in a twisted way, of course, point is that with the "models (ie methods)" as described above, Kelly staking strategies are useless.

i used very simple examples to get the pt across. what do you define to be a model? using as primitive data as you can to predict the game line? (ie. doing something like predicting who will win a game by estimating the number of expected points of each player and then adding up, as opposed to: predicting the 1q total as a function of the game total and game spread)?

then youre right, ive never had a model as unfortunately i dont know enough about any sport to do this confidently. and now that i think about it, youre right that the above examples are not models. i was mostly trying to illustrate the point that you never know with certainty which is the best guess at the true probability of something occuring(game landing on 7 or team winning by 3), which makes exact use of kelly impossible.

while I use data to answer the 1st and 2nd examples above, i wouldnt really call them models as it is just a basic application of method of moments. here is an nba example for you then and let me know if you disagree about this one.

suppose you want to predict the 2h spread and total. Here are some MODEL specifications that I might use given I do not know enough about basketball to estimate it from primitives:

2h spread = fcn(1h spread)
2h spread = fcn(game spread, 1h total)
2h spread = fcn(1h spread, 1h total, market game total, market game spread)
2h spread = fcn(1h spread, 1h total, market game total, market game spread, how tired Kobe looked, how discouraged losing 1h team looked)

which is the right one? Any stats guy is going to scoff at the last one. Not necessarily because they disagree that the last two variables are important but because any measure of them is going to be arbitrary and can bias the prediction. And even if you know which specification has all the relevant info, there are different ways to estimate it. there are 100s of textbooks on which estimation method should be used.

Now that I think I have a 'model' that might satisfy your definition, I can make all the same points as above. Which model is best? Despite what anyone says, there is no exact scientific method to choose among them. One might fit better but eats up degrees of freedom if your sample is small. How far back in the data should you go? Most would choose the model that fits their THEORY of what makes sense. like to predict the 2h spread you would probably be foolish to ignore the 1h spread as that will affect play in the 2h.

at the end of the day it seems you other guys are just picking one model and one period of data you think is best and going with it. if you are and you truly believe it is the best, then go ahead with kelly. If you are unsure and go with multiple specifications, estimation methods, and data periods, then you will have multiple competing models and unless you're willing to pick one of them as your go-to model, you cannot know your edge.

this talk of 'how can you know you have a +ev bet if you cant quantify your edge' is garbage. sure i can quantify my edge for each specification, and i can show you that for each one my edge is positive, but when i am making my bet i do not know my exact edge.

why do you guys think billy walters has 20 different guys independently giving him predictions on the same game instead of just one? Why arent they all predicting the exact same line every time, to 3 decimal places?

[peacebyinches]

I'll admit I didn't read this thread word for word, but I do understand the point ronnie is making about the theory behind the Kelly criterion actually serving a greater purpose than the actual application of "supposed" +EV bets using strictly defined mathematical parameters when your own estimation of any hypothetical edge is for all intents and purposes an incalculable variable.

When I consider the potential faults, biases, and improper attributions that undoubtedly exist in my models (and every single handicapper's model in the entire universe for that matter) I find that a calculated edge of 57% will not have any statistically significant predictive power over a calculated edge of 52% or even 62%.

[Dark Horse]

An old file that centers on maximum winning expectation (and sets it low), and minimum winning expectation (setting it high), as well as maximum % of BR. These things can all be personalized. This is just an example :

Kelly can produce astronomical bet sizes that many aren’t comfortable with. There is such a
thing as peace of mind...

Playing around with the numbers and my own comfort level I came up with these personalized
values. I would -conservatively- set 60% as the maximum attainable winning expectation for any
model (even though that is not true), and bet 1/3 Kelly. That would mean that, at -105 odds, any
winning expectation of 60% or higher would automatically result in a 6% bet size. If I never bet
anything with a lower winning expectation than 54%, and always bet at -105 odds, that would
give me seven bet sizes:

54% - 1.89%
55% - 2.58%
56% - 3.26%
57% - 3.94%
58% - 4.63%
59% - 5.31%
60% and above - 5.99%

The difference per percentage point is roughly 0.7% of bankroll:
6.0 - 5.3 - 4.6 - 3.9 - 3.2 - 2.5 - 1.8.

Err on the side of caution. If my expectation was between 58% and 57%, I take the lower
of the two.

For comparison, for full Kelly a 54% winning expectation would produce a 5.68% (!) bet size
(with -105 odds). That may work for some, but not for me.

======================================

If my maximum bet size was not 6% but 5% of bankroll, then I would simply use 2.0 - 2.5 - 3.0 -
3.5 - 4.0 - 4.5 - and 5.0 % of bankroll for winning expectations 54% through 60+%.

[That Foreign Guy]

Comfort level is a weird thing. The discussion of money vs bankroll is probably one that's better to split off to another thread, but it's taken me a while to become happy with the idea of betting £500 on a game even when it's what quarter Kelly demands. That's serious money, you know?

[MonkeyF0cker]

I can easily estimate the size of my edge... I choose to ignore size estimates though and simply focus on whether or not there is an edge. My models are not perfect... far from it. Ignoring edge size is one of the ways I attempt to overcome these imperfections. My figuring is that some of the games that look to have a 4% edge actually have a 2% edge and vice versa. Some of the games that look to have an edge will actually be coinflips. So when you ask "how do you know if you have an edge"... for me the answer is that I don't, at least not at the individual game level. But overall, after hundreds of games, my models do tend to pick winners at around a 53% clip. So for me it makes more sense to estimate my edge for all of my bets and set my bet amount at 2% of my bankroll.

I have a very hard time believing that if you can estimate your edge and know your model "pick(s) winners at around a 53% clip" that you are not leaving money on the table by flat betting. Fractional Kelly is one way of safeguarding against overstaking. If you have sufficiently backtested your model and are beating the market, why are you questioning the validity of your results? The only way it seems feasible that some form of Kelly is not best suited for your situation (where flat betting would be more ideal) is if your largest edges are underperforming. I'm assuming you've tracked/compared data of flat betting versus Kelly. So where are the outliers?

[sharpcat]

at the end of the day it seems you other guys are just picking one model and one period of data you think is best and going with it. if you are and you truly believe it is the best, then go ahead with kelly. If you are unsure and go with multiple specifications, estimation methods, and data periods, then you will have multiple competing models and unless you're willing to pick one of them as your go-to model, you cannot know your edge.

I believe your most accurate model would be decided by back testing. There is no reason to use 3 models and there is definately no reason to just randomly choose which of your 3 models you are going to use.

[specialronnie29]

I believe your most accurate model would be decided by back testing. There is no reason to use 3 models and there is definately no reason to just randomly choose which of your 3 models you are going to use.

nowhere do i say you randomly choose a model

back testing is the best way but it can take a lot of time accumulating the data if you do not have old LINE data. an example being 2h lines/totals. if youre deriving the 2h line or total from widely available info like game spread, game total, historical box score data then you can derive some sort of 2h line but you may not have historical data on the 2h lines themselves

[sharpcat]

nowhere do i say you randomly choose a model

back testing is the best way but it can take a lot of time accumulating the data if you do not have old LINE data. an example being 2h lines/totals. if youre deriving the 2h line or total from widely available info like game spread, game total, historical box score data then you can derive some sort of 2h line but you may not have historical data on the 2h lines themselves

Yes in this scenario it would complicated to quantify your edge but one could also argue that 2H lines are not near as efficient as full game lines and closing numbers would be less useful. There are other ways to back test your model you do not need to have 2H lines although it would be beneficial. 2H lines can still be considered pretty new on the seen but historic lines can still be found going back 3 seasons or more.

As MonkeyFocker said your likely leaving money on the table if your not using Kelly and you have a model that you are confident is pretty accurate, and smaller markets like 2H are much softer markets than full games. If your model is slightly less accurate than you believe it to be in a soft market I doubt it will hurt you much if you slightly overestimate your edge, especially considering that due to lower limits your bet sizes in these markets are likely to be exceeding max limits even using fractional Kelly.