PSA: Large soccer underdogs = fake edges

[ChinMusic]

Here's an exercise. Go to the articles section of Pinnacle, and read the article called "Explaining the Favourite-Longshot bias". They make no secret of the fact that there is more vig on the side of the underdog, which I think is the same point that slips is trying to make. It's possible that PLP is already taking this into account, but I'm not sure about that.

[ChinMusic]

Game #94 (Koln vs Arsenal) is another example of an inflated longshot edge.

#s as of 10:09am update are:

(pinnacle odds = win freq)
V +685 = 12.4%
T +383 = 20.1%
H -228 = 67.5%

I can see how PLP derives these numbers quite easily by applying the same margin to each outcome.

1 / 7.85 / 1.02955 = .124
1 / 4.83 / 1.02955 = .201
1 / (1+1/2.28) / 1.02955 = .675

The trouble with this is that it does not represent how Pinnacle actually distributes its margin. Here's a quote from the Pinnacle article I mentioned earlier.

"But what about contests with clear favourites and underdogs, for example with betting odds of 1.20 and 6.00? An even distribution of the margin would see odds shorten to 1.17 and 5.85 respectively. This, however, is commonly not what happens.

Instead, we are more likely to see odds that look like 1.19 and 5.41. The odds for the underdog have been shortened far more than the odds for the favourite. In terms of margin percentage, the underdog has a margin of 11%, whilst the favourite has a margin of just 1%."

Suppose ProLine gave a 6.00 to the underdog in this game. PLP's method would yield an edge of +8% when in fact it should be +0%. This is not a trivial difference.

I agree with MattyKGB that we should be looking for solutions rather than problems. However, I do not know how to solve this one! And my stale math skills are having trouble following Matty's explanation.

[Slips]

bettors love big faves, think they are locks and almost all the VIG is built into the fave. I know this goes directly against the longshot bias theory but I think that's wrong and dated. I've been profitably cashing super longshot soccer tixs for over a decade. Maybe I'm just really lucky and the benefiter of fake edges, who knows?!

[leoj]

Probabilities on CFL/NFL football totals seem off too, too much of a shift in probability for each half point off from pinnacle.

[ProlinePlayer]

Probabilities on CFL/NFL football totals seem off too, too much of a shift in probability for each half point off from pinnacle.

example?

I am satisfied that the football totals are accurate.
There may of course be differences due to how the probabilities are calculated.
I can think of several methods -
1/ use alternate line sets as available
2/ calculate probabilities using poisson
3/ reference to a dataset of past results

The program by default uses option 3, excepting cases were the sample size drops too low and then switches to option 2- It is possible that 1 would be stronger but is difficult to implement and I do not believe that the differences would be significant.

PLP

[ProlinePlayer]

I agree with MattyKGB that we should be looking for solutions rather than problems. However, I do not know how to solve this one! And my stale math skills are having trouble following Matty's explanation.

As I said earlier none of this is really new.

You may recall that several years back ( 4, 5 ?) you and I and Dilbert had an email discussion over this matter. In the end no other method of calculating the probabilities was considered better.

PLP

[leoj]

Probabilities on CFL/NFL football totals seem off too, too much of a shift in probability for each half point off from pinnacle.

example?

I am satisfied that the football totals are accurate.
There may of course be differences due to how the probabilities are calculated.
I can think of several methods -
1/ use alternate line sets as available
2/ calculate probabilities using poisson
3/ reference to a dataset of past results

The program by default uses option 3, excepting cases were the sample size drops too low and then switches to option 2- It is possible that 1 would be stronger but is difficult to implement and I do not believe that the differences would be significant.

PLP

Look at 14O on Saturdays list, a 2 point differential between OLG and pinnacle, you're indicating about 10 cents per half point, about double what pinnacle charges.

[MattyKGB]

Probabilities on CFL/NFL football totals seem off too, too much of a shift in probability for each half point off from pinnacle.

example?

I am satisfied that the football totals are accurate.
There may of course be differences due to how the probabilities are calculated.
I can think of several methods -
1/ use alternate line sets as available
2/ calculate probabilities using poisson
3/ reference to a dataset of past results

The program by default uses option 3, excepting cases were the sample size drops too low and then switches to option 2- It is possible that 1 would be stronger but is difficult to implement and I do not believe that the differences would be significant.

PLP

Are you using raw Poisson or applying some kind of transformation? If it's raw Poisson, it will be off.

A property of Poisson is that its variance is equal to its mean. So let's consider a football game with a total of 47. We can look at the total score as 7X+3Y where X,Y are the number of touchdowns and field goals. X and Y are what should be Poisson distributed (or close enough). Let's suppose the mean of X is 5 touchdowns and the mean of Y is 4 field goals. Then, Var(7X+3Y) = 7^2 * 5 + 3^2 * 4 = 281, which is much larger than the variance of 47 that you'd use if you apply raw Poisson to the total of 47.

You would do a lot better by using a normal distribution instead of poisson. Use something like sqrt(6*total) for the standard deviation, or sqrt(6*47)=16.79 for the standard deviation and 47 for the mean in this example. It's not really correct either, but should be much better than poisson.

[Slips]

In the end no other method of calculating the probabilities was considered better. PLP

Your chosen method shouldn't result in probabilties that show profitable opportunities at many competing sportsbooks. This is not 2010, the market has become more efficient than that.

[ChinMusic]

I agree with MattyKGB that we should be looking for solutions rather than problems. However, I do not know how to solve this one! And my stale math skills are having trouble following Matty's explanation.

As I said earlier none of this is really new.

You may recall that several years back ( 4, 5 ?) you and I and Dilbert had an email discussion over this matter. In the end no other method of calculating the probabilities was considered better.

PLP

I remember. I tried to come up with something better, and failed miserably. But the problem is still there, waiting for a good solution. The edges on soccer underdogs are way overstated. Matty is our best hope here.