College betting model

[Bill the cop]

The following betting concept was posted by me over at Sportsforumworld at the beginning of college football (it's one of my "thinking outside the box" plays).

First of all I'm not a professional gambler, in fact I'd rather be fishing for tarpon off the coast of Costa Rica (which I just got back from) than trying to beat the sports books. With that said, there's entirely too much hand-wringing about not being able to beat the WA closing lines. I'm saying that they can be beat and it isn't even that hard to do. I do no handicapping whatsoever, all my bets are the result of data analysis. Also, I seldom make straight bets (exceptions below), the majority of my bets involve teasers or other forms of parlays.

Below is an actual example of one of the methodologies I use. This particular database reflects the closing lines depicted by Don Best for college football 2000-2009 (6338 games). The subset is for full game lines that are less than 2 to 1 LTR, e.g., USC-28 with a total of 55 or a LTR of 1.96. But the bets themselves are all first half bets, the USC 1h line would be about -14.5 with a total of 27.5. And the sample size for this subset is 461 or 7.3% of the database. Just to be perfectly clear, the best bets by far would be correlated parlays, but none of the respectable books will allow them with an LTR of 2 or less, so we're going to skin this cat another way! First, the results of the 461 games for the first half as they actually turned out from a correlated parlay perspective:

Fav/O 163 for 35.4%

Fav/U 43 for 9.3%

Dog/O 89 for 19.3%

Dog/U 153 for 33.2%

F/P 5 for 1.1%
D/P 5 for 1.1%
P/O 1 for .2%
P/U 2 for .4%

Obviously the correlated parlays of Fav/O or (and) Dog/U would be nice, but of course, they won't let you bet these. So what I do is bet two STRAIGHT bets on the Dog and the Over. At first glance these bets appear to be anti-correlated, but let's do the math.

If you were to bet $110 on the Dog and $110 on the Over for all of the 461 games you would have $101420 invested.

You would collect 89x$420=$37380 for when the Dog and Over came in and 316x$210=$66360 when the Fav/O or Dog/U came in (win one, lose one, lose $10 juice). $37380+$66360=$103740-$101420=$2320 profit. This appears to be only a 2.3% +EV ROI. But if you look at it from a ROR perspective all you are really risking is the $10 juice on the 316 times the F/O or D/U came in ($3160) and the $220 lost when the F/U came in (43x$220=$9460). So the way I figure it the real risk is $9460+$3160=$12620 or +EV of 18.4%. Simply stated, 70% of the time you're going to lose $10, 10% of the time you're going to lose $220 and 20% of the time you're going to win $200.


Also, I know all about the "in sample" and "out of sample" paradigm as it relates to data analysis, But honestly, I'm not inclined to sit on what I consider a profitable subset for 3 or 4 years to see if it withstands the test of time. Some may say that a sample size of 461 is not enough to extrapolate from when making these types of bets, but it's enough for me!

My first full year betting this subset (2009), SS 46, 1 F/U, 9 D/O and 36 F/O D/U.

So far this year, SS 27, 1 F/U, 7.5 D/O and 18 F/O D/U

[tomcowley]

Cliffs: On games with big spreads, dogs hit 54% and overs hit 56% in my sample, so I'm just going to bet them blindly assuming that they'll continue to, even after I tell everybody about it. And present it as some random dog/o system because I can't even analyze the results into the simplest components.

[Bill the cop]

Cliffs: On games with big spreads, dogs hit 54% and overs hit 56% in my sample, so I'm just going to bet them blindly assuming that they'll continue to, even after I tell everybody about it. And present it as some random dog/o system because I can't even analyze the results into the simplest components.

Of course you're missing the value of the low LTR which is where the money is.

If the 1h line is San Jose St +23 and the over is 27, the chances of losing both bets (F/U) is about 5%, but the chances of cashing both Dog and Over is about 20%. If you don't have the LTR data you're just betting blindly.

[CFA]

lol^

[Flight]

Thanks for the information Bill.

Have you tried testing this on out of sample data? For example, play the D/O on every game from 1992-1995?

[roasthawg]

Run the numbers with the Fave and the Under...

163 * -10 = -$1,630
43 * 200 = $8,600
89 * -220 = -$19,580
153 * -10 = -$1,530

Total profit/loss = -$14,140 over those same 461 games wagering $110 per bet. You will get crushed if you're on the wrong side of this... you better have a reason as to why the dog/over is so much more likely to hit than the fav/under.

[donjuan]

Roasthawg,

BTC doesn't deal with reasons. He just data mines and expects future performance to be the same, while ignoring both variance and changing market forces (i.e. Pinnacle/Matchbook).

[RickySteve]

At first glance these bets appear to be anti-correlated, but let's do the math.

This was my favorite part.

[Wrecktangle]

I see BTC posting here now.

Perhaps he is tired of combing the malware out of his machine that you can easily get posting at SFW.

[wrongturn]

Interesting findings. Two straight bets on anti-correlated events look like semi-hedge bets.

[Bill the cop]

Run the numbers with the Fave and the Under...

163 * -10 = -$1,630
43 * 200 = $8,600
89 * -220 = -$19,580
153 * -10 = -$1,530

Total profit/loss = -$14,140 over those same 461 games wagering $110 per bet. You will get crushed if you're on the wrong side of this... you better have a reason as to why the dog/over is so much more likely to hit than the fav/under.

You've reversed the numbers. The 43 is the times the F/U hit, the 89 is the D/O data.

163*-10=$1630
153*-10=$1530
43*-220=$9460

89*200=$17800

[Bill the cop]

Thanks for the information Bill.

Have you tried testing this on out of sample data? For example, play the D/O on every game from 1992-1995?

What must be understood is there are 4 possible outcomes when making 2 bets on the same game (side and total), D/U, F/O, D/O, F/U. In games with LTR (that's line to total ratio) that are less than 2, the distribution of those 4 outcomes are different than games with higher LTRs. About 70% of the time the F/O and D/U will come in, these numbers are consistent with correlated parlay outcomes (which is why they won't let you bet them), the remaining 30% is divided between D/O and F/U. The lower the LTR is, the higher percentage of D/O bets will hit and the fewer F/U.

Let's take an extreme example to make my point. A 1h line is San Jose +24.5 and the Over is 26, which is a LTR of 1 to 1.06. Of the 4 outcomes I cited above, the possibility of the F/U coming in is almost nill (the Fav. would have to win by exactly 25-0 for that to happen). So the 4 possible outcomes is essentially reduced to 3. The F/O and D/U will probably come in about 80% of the time, the D/O and the unlikely F/U would share the rest.

The key is to have accurate LTR data.

[Bill the cop]

Interesting findings. Two straight bets on anti-correlated events look like semi-hedge bets.

The following is the actual breakdown of results by LTR of 2 or less for the 461 game sample size. It's not a hedge at all, it's a paradigm I spent a lot of time analyzing before I started betting it last year.

1. to 1.4, F/O and D/U 86%, D/O (winner) 11%, F/U (loser) 3%

1.4 to 1.6, F/O and D/U 70%, D/O (winner) 24%, F/U (loser) 6%

1.6 to 1.8, F/O and D/U 70%, D/O (winner) 20%, F/U (loser) 10%

1.8 to 2, F/O and D/U 65%, D/O (winner) 22%, F/U (loser) 13%

[Peregrine Stoop]

BTC is making quite a bold claim here. He's claiming the spread and total markets on 1h are not efficient.
He might be right.
He might not be.

Time would be much better spent finding an out that still lets one bet the true correlated parlays of fave/over and dog/under.

[Peregrine Stoop]

BTW, Bet Phoenix shades these ATS lines quite heavily compared to most of the major books.
I wonder why that is

[wrongturn]

BTC, my calculation says if DO% minus FU% is larger than 4.762%, the bets are +EV. (abbreviations are appropriate i guess )

So maybe there are still somewhat good subset above LTR 2?

[Peregrine Stoop]

I had some smart people explain this to me about why this 'system' only works if the 1h markets are inefficient.
Create a 2x2 square

|////|Dog|Fave|
|Over| a | b |
Under| c | d |

now, a, b, c, and d are the proportions of outcomes
we know that a + b + c + d = 1
It should also be that a + b = .5, c + d = .5, a + c = .5, b + d = .5
otherwise, the market is inefficient.
BTC's datamining has found an area in the past where the market has mispriced.
The question is how good do we feel about that mispricing to continue in to the future.
With the advent of more and more computer power with smarter guys digging into 1h markets, many propose that this market inefficiency will disappear like the NFL game home dogs.
If you back BTC and his picks here, you are saying that you expect the inefficiency to continue, at least in the near term.

I know which side I am on.

[roasthawg]

You've reversed the numbers. The 43 is the times the F/U hit, the 89 is the D/O data.

163*-10=$1630
153*-10=$1530
43*-220=$9460

89*200=$17800

Yeah, I reversed the numbers on purpose. What's the reasoning as to why the dog/over hits so much more often than the fav/under? If there's no solid explanation as to why then it's just as likely as not that you get crushed here.

[Bill the cop]

BTC, my calculation says if DO% minus FU% is larger than 4.762%, the bets are +EV. (abbreviations are appropriate i guess )

So maybe there are still somewhat good subset above LTR 2?

See post #13, at LTRs higher than 2, diminshing returns.

[Justin7]

For the record, this is not a model. It's more of a system.

BTC has show that these lines were inefficient in the past. These ideas are only valid if you assume that these markets will remain inefficient going forward. We'll see if that happens.

[Bill the cop]

For the record, this is not a model. It's more of a system.

BTC has show that these lines were inefficient in the past. These ideas are only valid if you assume that these markets will remain inefficient going forward. We'll see if that happens.

I don't see these markets as being inefficient at all. The 1h line is established like any other, they want to generate two way action. For 1h lines the process is pretty simple, they take right at 60% of the game line and 50% of the whole game total. If more money comes in on one side (or total), they adjust the line.

All I'm doing is exploiting the value of the low LTR games. Take #191 B.St/San Jose St. The game line is 40.5 and total is 56, so the 1h lines are +24.5 and total 28.5. So I bet the +24.5 and Over. If San Jose scores even 3 points in the 1h, I can't lose both bets (but I sure can win both bets). If San Jose dosen't score 1h, the only way I can lose both bets is if B.St wins by EXACTLY 25,26, 27, or 28 (no more , no less).

[tomcowley]

I don't see these markets as being inefficient at all.

You mean the part where dogs hit 54% and overs hit 56% indicates an efficient market? ROFL. Do you even have the first idea what the term means?

[Bill the cop]

You mean the part where dogs hit 54% and overs hit 56% indicates an efficient market? ROFL. Do you even have the first idea what the term means?

Well, let's look at college FB over the last 10 years and see how efficient they are.

All lines, full game, SS 6629
Dogs, 3318-3204-107 for 50.1%
Overs, 3222-3318-89 for 48.6%

1h
Dogs, 3288-3239-89 for 48.6%
Overs, 3199-3311-119 for 48.3%

Now let's look at games with low LTR (1-3) SS 1478, full game
Dogs, 727-730-21 for 49.2%
Overs, 722-735-21 for 48.8%

1h
Dogs, 739-724-15 for 50%
Overs, 772-677-29 for 52.2%

Looks like a pretty efficient market to me, if anything a small bias in favor of the F/U.

So when you get up off the floor from laughing, lets see the data from your hypothetical database that shows that Dogs at 54% and Overs at 56%.

I'm beginning to think you guys don't even have a database and are just pulling numbers out of your asses.

[donjuan]

Well, let's look at college FB over the last 10 years and see how efficient they are.

All lines, full game, SS 6629
Dogs, 3318-3204-107 for 50.1%
Overs, 3222-3318-89 for 48.6%

1h
Dogs, 3288-3239-89 for 48.6%
Overs, 3199-3311-119 for 48.3%

Now let's look at games with low LTR (1-3) SS 1478, full game
Dogs, 727-730-21 for 49.2%
Overs, 722-735-21 for 48.8%

1h
Dogs, 739-724-15 for 50%
Overs, 772-677-29 for 52.2%

Looks like a pretty efficient market to me, if anything a small bias in favor of the F/U.

So when you get up off the floor from laughing, lets see the data from your hypothetical database that shows that Dogs at 54% and Overs at 56%.

I'm beginning to think you guys don't even have a database and are just pulling numbers out of your asses.

I'm beginning to think you have no clue what an efficient market is. Actually, I know you don't have a clue what one is after reading this post. This is beginner level stuff: http://en.wikipedia.org/wiki/Efficie...ket_hypothesis

[Bill the cop]

Yeah, nice dance move, now show me the data that has Dogs at 54% and Overs at 56%. Forget it.

[donjuan]

Yeah, nice dance move, now show me the data that has Dogs at 54% and Overs at 56%. Forget it.

Dance move? Your post had literally NOTHING to do with efficiency and you throw around terms like "efficiency" and "model" without having the slightest clue what they mean.

[tomcowley]

Yeah, nice dance move, now show me the data that has Dogs at 54% and Overs at 56%. Forget it.

I should have every right to call you out after this post, since they're the numbers from your very first post in this thread.

Fav/O 163 for 35.4%

Fav/U 43 for 9.3%

Dog/O 89 for 19.3%

Dog/U 153 for 33.2%

F/P 5 for 1.1%
D/P 5 for 1.1%
P/O 1 for .2%
P/U 2 for .4%

Favs 163+43+5 = 211. Dogs 89+153+5= 247. Dogs cover 53.9 (54)%.

Overs 163+89+1=253. Unders 43+153+2=198. Overs cover 56.1 (56)%.

Or you can just read what Justin7 said

BTC has show that these lines were inefficient in the past. These ideas are only valid if you assume that these markets will remain inefficient going forward.

You really had no idea that you were winning because of an inefficient subset, did you? Did you really think there was something magical about your dog/o betting system that went beyond betting dogs when they were covering at 54% and betting overs when they were covering at 56%? LOL.

[durito]

But do you have a database TC?

[sideloaded]

Wasn't his point that he found an inefficient subset? Isn't the point of gambling finding market inefficiencies at making money off them? At least short term?

[xyz]

If you were able to bet the dog at +100 and over at +100 on BTC's first half plays, instead of -110, would it be blindly profitable?

[Justin7]

If you were able to bet the dog at +100 and over at +100 on BTC's first half plays, instead of -110, would it be blindly profitable?

In the past, obviously yes. In the future? If the market is incompetent, yes. If the market adapts and becomes more accurate (as it has in the past), the advantage will go down, disappear, or become -EV. But who knows what the market will do.

[Data]

You mean the part where dogs hit 54% and overs hit 56% indicates an efficient market?

How do these numbers, respectively 1 STD and 1.3 STD away from expected, contradict the null hypothesis of the given market efficiency?

I would attack BTC's position from variance perspective. Why you all are talking about market inefficiency? I think your reasons are right but the means are faulty.

[donjuan]

How do these numbers, respectively 1 STD and 1.3 STD away from expected, contradict the null hypothesis of the given market efficiency?

I would attack BTC's position from variance perspective. Why you all are talking about market inefficiency? I think your reasons are right but the means are faulty.

Because the efficiency of the market right now is important going forward, obviously.

[Data]

Because the efficiency of the market right now is important going forward, obviously.

No, that was not it, you all said the market was inefficient in the past. I see no evidence of this.

[donjuan]

Feel free to provide a quote showing I said the market was inefficient (although it may very well have been).