Do I have enough data to start using the Kelly Formula?

[mathdotcom]

This thread is a biggest comedy in the think tank.

Here's a question: if the Cavs were supposed to be favored by 75, what is the ex-ante probability of the Cavs "only" winning by 12? Using a conventional value of what a point is worth in NBA, your answer is going to be one in a few million, at least.

I could go on and on. This is hilarious. Anyone who subscribes to the 'bet only your best pick per day' philosophy is an automatic moron. Imagine most days your edge is 1-2% and all of a sudden one day there are 5 games where you have edges of over 10%. Only a complete fukking idiot would pick just the best bet and pass on the other 4 games.


You guys are only hurting this guy by not being honest with him.

[Dylan]

Assuming you can accurately estimate your edge Kelly is going to help maximize the growth of your bankroll while minimizing your risk of ruin. One of the worst things you can do when using Kelly is overestimate your edge and over bet because it will greatly increase your risk of ruin and the destruction of your bankroll, the inability to accurately estimate your edge is one of the reasons why some choose to use fractional Kelly.

Considering that your example above suggests that your "model" gauges your "edge" depending on the point differential between the actual line and your prediction but does not take the value of each individual point into consideration, along with the huge differential in your prediction compared to an efficient market suggests that you have some serious flaws. Absolutely no way in hell your model is spotting inefficiencies in the market that are allowing you to have a 28% edge unless you are betting on the peewee football league with max wagers of $50.

My advice would be to chalk your 28-15 record up to luck and go back to the drawing board you are nowhere near being ready for Kelly.

I think you've misunderstood something. My edge wouldn't fluctuate based on my predicted point differential. It fluctuates based on previous bets' success.

Let's say after 10 bets I'm 7-3 for a 70% win percentage, and in those 10 bets the average odds were was -105.04 (1.952.)

The Kelly would be 38.4880% of my bank, which let's say was at $2000. 38.4880% of 2000 is $769.76. Let's say for bet number 11, the odds are +106 (2.06).

So I place my $769.76 bet, at +106. I win.

Now my bankroll is $2815.95, my average odds are -103.97 (1.961818182), and my record would be 8-3 for a 72.7272%. For bet number 12, my Kelly would be 44.3718%.

Hope that clears things up.

[Dylan]

I agree with Dylan. USC is -10 points today, but I have them at -86 in my model, which is the highest for today. The Kings are -68, but I'm not going to bet them because only one bet a day is +EV.

Thanks for the response, I was beginning to think I was insane for having outliers.

[Dylan]

This thread is a biggest comedy in the think tank. Here's a question: if the Cavs were supposed to be favored by 75, what is the ex-ante probability of the Cavs "only" winning by 12? Using a conventional value of what a point is worth in NBA, your answer is going to be one in a few million, at least. I could go on and on. This is hilarious. Anyone who subscribes to the 'bet only your best pick per day' philosophy is an automatic moron. Imagine most days your edge is 1-2% and all of a sudden one day there are 5 games where you have edges of over 10%. Only a complete fukking idiot would pick just the best bet and pass on the other 4 games. You guys are only hurting this guy by not being honest with him.

My edge doesn't fluctuate based on my predicted point differential. It fluctuates based on previous bets' success.

[FourLengthsClear]

Thanks for the response, I was beginning to think I was insane for having outliers.

AlwaysDrawing's response to you was 100% sarcasm.

Your model/method is badly flawed but if it is working for you just keep what your doing until you are in a position to correct those flaws. Utilising Kelly based on previous win/loss ratios is not a good idea.

[Dylan]

AlwaysDrawing's response to you was 100% sarcasm. Your model/method is badly flawed but if it is working for you just keep what your doing until you are in a position to correct those flaws. Utilising Kelly based on previous win/loss ratios is not a good idea.

Oh, I didn't realize. Sarcasm is hard to detect online.

Thanks for all the advice, FourLengthsClear. It's much appreciated.

[mathdotcom]

Thanks for the response, I was beginning to think I was insane for having outliers.

I am dying to know how you constructed your model.

[Dylan]

I am dying to know how you constructed your model.

I collect pregame stats, and after the game I record the score. After I had collected enough data I came up with a formula that had a high stat to score correlation.

That's about it.

[bztips]

I think you've misunderstood something. My edge wouldn't fluctuate based on my predicted point differential. It fluctuates based on previous bets' success.

Let's say after 10 bets I'm 7-3 for a 70% win percentage, and in those 10 bets the average odds were was -105.04 (1.952.)

The Kelly would be 38.4880% of my bank, which let's say was at $2000. 38.4880% of 2000 is $769.76. Let's say for bet number 11, the odds are +106 (2.06).

So I place my $769.76 bet, at +106. I win.

Now my bankroll is $2815.95, my average odds are -103.97 (1.961818182), and my record would be 8-3 for a 72.7272%. For bet number 12, my Kelly would be 44.3718%.

Hope that clears things up.

Ok ok, I completely retract my prior comment supporting your prior comment. This statement makes it very clear that you have absolutely no idea what is meant by "edge". Stay away from Kelly at all costs until you figure out how to estimate your edge.

[mathdotcom]

I collect pregame stats, and after the game I record the score. After I had collected enough data I came up with a formula that had a high stat to score correlation.

That's about it.

OK. But when that formula yields estimates that are so far off in markets with huge liquidity [NFL and less so NBA], you have to be skeptical. You can sometimes find huge edges in weird ignored markets, but if your NFL spread is more than a couple points off the market then your model is wrong.

And once you come up with a better model, you should read some sports betting books on money management. Betting once per revolution of the Earth is the dumbest sports betting misconception that has ever existed.

[AlwaysDrawing]

You guys will all be crying while I soak up that sweet sweet cash on USC tonight. Tomorrow is a Friday, so I usually play my worst play in the model, because I usually hit 30% on my regular plays on Fridays (3/10 the last 10 Fridays).

I'm thinking the Warriors, since everyone knows the Lake Show is usually money at home. My line is Lakers -63, so I won't be surprised to see the line at Warriors +63, but I'd take it as low as Warriors -11. Just buy 6 half points, one for each day of the year so far.

Best of luck to all.

[Dylan]

OK. But when that formula yields estimates that are so far off in markets with huge liquidity [NFL and less so NBA], you have to be skeptical. You can sometimes find huge edges in weird ignored markets, but if your NFL spread is more than a couple points off the market then your model is wrong. And once you come up with a better model, you should read some sports betting books on money management.

This happens very rarely. 99% of the time my estimates are within a few points of the spread.

[sixrolla]

Anybody who thinks like this regarding money management, and has a model that predicts 130-55 NBA scores is out to ******* lunch and you guys need to give him a dose of reality.

Or give him encouragement and let him keep pounding lines the wrong way lol.

[Dylan]

Ok ok, I completely retract my prior comment supporting your prior comment. This statement makes it very clear that you have absolutely no idea what is meant by "edge". Stay away from Kelly at all costs until you figure out how to estimate your edge.

You can use Expected Win Probability% in replace of edge. This is what I'm referring to.

[mathdotcom]

This happens very rarely. 99% of the time my estimates are within a few points of the spread.

If your model could ever weight a 'factor' so heavily that it yields Cavs -75 then you likely have a small sample issue somewhere. Since you didn't explain your model clearly I can't say where. But it's funny you mentioned OUTLIER because rather than the prediction being an outlier, your model is being strongly biased DUE to an outlier.

A simple example:

Suppose your model for NBA is you take the pts per game avg of each player to come up with a team total. And suppose one of the players for the Cavs in an upcoming game has only played one game before, and in fact during that game he only played one minute, but in that one minute he scored 10 pts. Then to estimate his pts/game you could multiple those 10 pts by 48 to come up with 480 (obviously a crazy estimate). Then you plug that into your model and the model thinks the Cavs are going to destroy the other team.

That is a crude example but I am confident that sort of concept is what is causing your crazy predictions -- that you had an extreme outlier in your sample that generated your out of sample prediction. That said, the fact that this could happen casts some doubt on your model.

Anyways there are tons of NBA games in the next few months so you have very little to lose by betting small and seeing if your win % stays high. But please, for fukk's sake, bet based on your edge and not based on the angle of the fukking planet earth.

[FourLengthsClear]

You can use Expected Win Probability% in replace of edge. This is what I'm referring to.

The win% shown in that calculator is your projected figure for the matchup you are about to bet on, not any historic level.

[Dylan]

If your model could ever weight a 'factor' so heavily that it yields Cavs -75 then your weight is wrong. Since you didn't explain your model clearly I can't say why. But it's funny you mentioned OUTLIER because clearly you're EXTRAPOLATING based upon a small sample, and an outlier is affecting your results. A simple example: Suppose your model for NBA is you take the pts per game avg of each player to come up with a team total. And suppose one of the players for the Cavs in an upcoming game has only played one game before, and in fact during that game he only played one minute, but in that one minute he scored 10 pts. Then to estimate his pts/game you could multiple those 10 pts by 48 to come up with 480 (obviously a crazy estimate). Then you plug that into your model and the model thinks the Cavs are going to destroy the other team. That is a crude example but I am confident that sort of concept is what is causing your crazy predictions -- that you had an extreme outlier in your sample that generated your out of sample prediction. That said, the fact that this could happen casts some doubt on your model. Anyways there are tons of NBA games in the next few months so you have very little to lose by betting small and seeing if your win % stays high. But please, for fukk's sake, bet based on your edge and not based on the angle of the fukking planet earth.

When the prediction for the Heat-Cavs game was made I had 7813 games in my sample. I won 4417 and lost 3396 of those game.


Anyways there are tons of NBA games in the next few months so you have very little to lose by betting small and seeing if your win % stays high.

Thanks for this advice.

[Dylan]

The win% shown in that calculator is your projected figure for the matchup you are about to bet on, not any historic level.

Thanks for the info.

[Dash2in1]

Of course betting on the edge is better than 'betting based on the angle of the ******* earth'. Having said that, using the site's kelly calculator for multiple independant simultaneous bets isn't especially useful, especially if the amount of bets mount. Well, that is, unless you are willing and it's possible to parlay every single possible combination of bets for the correct odds, but that's not going to happen. (Thx bztips ;-) )

[Dylan]

Of course betting on the edge is better than 'betting based on the angle of the ******* earth'. Having said that, using the site's kelly calculator for multiple independant simultaneous bets isn't especially useful, especially if the amount of bets mount. Well, that is, unless you are willing and it's possible to parlay every single possible combination of bets for the correct odds, but that's not going to happen.

I don't use the site's Kelly Calculator. I can calculate it myself. I was just saying that you can use Expected Win % instead of edge.

[Duff85]

I agree with Dylan. USC is -10 points today, but I have them at -86 in my model, which is the highest for today. The Kings are -68, but I'm not going to bet them because only one bet a day is +EV.

Based on your model Kelly is suggesting that I rob multiple convenience stores to get down the appropriate stake on USC-7. I'm tailing.

[mathdotcom]

When the prediction for the Heat-Cavs game was made I had 7813 games in my sample. I won 4417 and lost 3396 of those game.


Anyways there are tons of NBA games in the next few months so you have very little to lose by betting small and seeing if your win % stays high.

Thanks for this advice.

See, you're making no sense again. What the hell do you mean 'won'? At the start of the thread you said you were 42-15 or whatever it was.

Now I'm convinced you're back testing over the same sample used to produce that 4417-3396 figure.

Guys like you have come and gone in the Think Tank. They always come in saying they have a model, but then they can barely talk Stats 101 -- it always turned out they were using some ridiculous system or making some gross statistical error.

Your appropriate Kelly stake is zero.

[Dylan]

See, you're making no sense again. What the hell do you mean 'won'? At the start of the thread you said you were 42-15 or whatever it was. Now I'm convinced you're back testing over the same sample used to produce that 4417-3396 figure. Guys like you have come and gone in the Think Tank. They always come in saying they have a model, but then they can barely talk Stats 101 -- it always turned out they were using some ridiculous system or making some gross statistical error. Your appropriate Kelly stake is zero.

This should clear things up. As of today, my records are:

NCAAB games I've bet on: 30-14
All NCAAB games: 3412-3205

NBA games I've bet on: 40-34
All NBA Games: 6244-4842

[jgilmartin]

This should clear things up. As of today, my records are:

NCAAB games I've bet on: 30-14
All NCAAB games: 3412-3205

NBA games I've bet on: 40-34
All NBA Games: 6244-4842

Cool, but when you list "All NBA Games: 6244-4842" isn't that in-sample games? Makes it totally irrelevant.

[Dylan]

Cool, but when you list "All NBA Games: 6244-4842" isn't that in-sample games? Makes it totally irrelevant.

I'm not sure what you mean my "in-sample games." Basically, since I started recording NBA games against the spread, I've correctly predicted 6244 games, and been wrong 4842 times.

[jgilmartin]

I'm not sure what you mean my "in-sample games." Basically, since I started recording NBA games against the spread, I've correctly predicted 6244 games, and been wrong 4842 times.

I am asking if the 11,086 games (well, presumably more, due to pushes) that comprise that record were used in the creation of your formula.

[FourLengthsClear]

I'm not sure what you mean my "in-sample games." Basically, since I started recording NBA games against the spread, I've correctly predicted 6244 games, and been wrong 4842 times.

Your data sample covers a period of around 10 years, yes you have only made 44 bets.

Did you establish your model/method before you started recording these stats or is the method back-fitted?

[Dylan]

Yes, all games are included in the creation of my formula. The formula is adjusted each day to achieve the highest correlation between stats and scores.

[FourLengthsClear]

Yes, all games are included in the creation of my formula. The formula is adjusted each day to achieve the highest correlation between stats and scores.

Oh dear.

[Dylan]

Your data sample covers a period of around 10 years, yes you have only made 44 bets. Did you establish your model/method before you started recording these stats or is the method back-fitted?

I used data from previous years' to create my initial formula. One I had enough games, I back-tested subsequent games to get most of the 3412-3205 record, adjusting the formula after each day for the highest possible R^2. This year is my first year placing actual money on NCAAB, and my second for the NBA.

[mathdotcom]

Oh dear.

I called it.

Every few months we get one of these guys.

[rory borealis]

There are just about 4 or 5 of you guys posting on this thread who are just brilliant in math... I made it through 17 weeks of Calculus in college through "volumes of solids of revolution", "integration by parts" but then started "crashing and burning" on "partial differentiation and "vectors in space".

I took a business statistics class and was totally lost

I want to thank you all for the help you provide to us with less understanding... you people sure are smart

[rory borealis]

I called it.

Every few months we get one of these guys.

...and sometimes you get TWO within that very same time frame because here I come...



DIFFERENT SITUATION
I don't KNOW my Kelly Edge... it quite possibly could be a NEGATIVE expectation and probably falls between picking 47.6% winners and 52.4% winners against the 11/10.

...but what if it doesn't?

I'm not going to risk huge money since I'll make $0.50 bets at 5Dimes @ -105 and deposit the minimum $250.

So that is 500 units of bankroll that I would have which I'm guessing would mean I'd need a 0.2% positive expectation to be wagering properly on a single wager.

Now I've never really had more than a vague understanding of "independent" and "dependent" events so I'll just find an unoccupied corner of this thread to 'spill my guts' about what I've read in Ken Uston's, "Million Dollar Blackjack" and then whatever subset of the mathematically-capable who are so inclined can meander on over and perhaps provide some assistance.

In Blackjack, Ken Uston was talking about "early surrender" that once existed in Atlantic City... If I remember correctly,he made the parenthetical comment that with a 95% 5% element of ruin at a 0.1% basic strategy edge one would need 900 units of bankroll.

I don't know about "independent" vs "dependent" but he stated that if one was playing to a 1.5% edge against the Blackjack dealer that one could increase one's betting action by adding (i think) 11% to one's wager and dividing that new amount into two halves and play two hands against the same dealer's upcard.

So I don't care about losing this $250 at 5Dimes that I'm going to deposit but I'm going to FADE the LARGEST POINTSPREAD DIFFERENCE MARGINS of the Dunkel Index or maybe FADE the largest CONSENSUS FIGURES that SBR provides and HOPE that I enjoy an AVERAGE of 0.2% edge on each wager that I fade.

Since both DUNKEL and the CONSENSUS FIGURES would offer plays to FADE on each & every game getting ready to be played, I'm going to try to jam as many plays as I can WITHOUT OVERBETTING my assumed Kelly fading edge of 0.2%

(All of a sudden, I'm probably wrong, but I think I'm asking you guys a 'fair coin test' with a bookie's vigorish flavor to it)

So... on a hypothetical Sunday at 1 PM Eastern Standard Time... there are 8 NFL games going off with 16 5Dimes wagering opportunities @ -105 (8 ATS sides and 8 O/U totals bets). On that very same Sunday in the NBA at 1 PM EST there are 10 games with 20 wagering opportunities @ -105...

...so now we have 36 wagering opportunities all firing off at 1 PM EST.

...but WAIT!!! ...there's more!... @-105 5Dimes only allows -105 betting on ATS sides on NCAA BB and there are 50 games going off at 1 PM.

So that is 86 wagering opportunities available on Sunday @ 1 PM EST

So I'm going to fade the largest CONSENSUS DIFFERENCES or the LARGEST DUNKEL LINE/ POINTSPREAD LINE margins and assume that by fading them that I enjoy an AVERAGE of 0.2%.

1) How many of those 86 wagering opportunities can I avail myself of without OVERBETTING my (assumed) AVERAGE KELLY EDGE of 0.2%?

2) At that same assumed AVERAGE EDGE of 0.2% what size bankroll would I need if I wanted to FADE all 86 wagering opportunities that are all simultaneously firing off at Sunday 1 PM EST?

[rory borealis]

Should I have added a desired "confidence interval"?

Is "confidence interval" the same/opposite thing as "element of ruin"?.... because now I'm thinking Uston was probaly saying 5% "element of ruin"?

If "yes" then I want to be 97% confident that I'm not overbetting although I don't know why that would be a factor if I previously stated that I'm assuming a FIXED AVERAGE EDGE of 0.2% on every wager

[FourLengthsClear]

1) How many of those 86 wagering opportunities can I avail myself of without OVERBETTING my (assumed) AVERAGE KELLY EDGE of 0.2%?

2) At that same assumed AVERAGE EDGE of 0.2% what size bankroll would I need if I wanted to FADE all 86 wagering opportunities that are all simultaneously firing off at Sunday 1 PM EST?

1) Strictly speaking, you are overbetting if you have more than one concurrent wager.

2) You would need a bankroll which is 19.56% bigger in order to place wagers of the same size (in dollar terms) based on 86 concurrent events all at -105 and all with a 0.2% edge. Expressed another way your bet size per wager should be reduced to 0.1756%.

This figure is derived by reducing the recommended risk amount for a single wager (0.21%) by that same amount for every additional wager. So for two concurrent bets the risk amount would be 0.21% multiplied by (1-0.21%) or 0.2096%. For three concurrent wagers it would be 0.2091%. Repeating that calculation another 83 times gives you the figures above.