I got a comment the other day asking about whether it is possible to get the expected number of goals scored from over/under odds, similar to how you can do this for odds for win, draw or lose outcomes. The over/under odds refer to the odds for the total score (the sum of the score for two opponents) being over or under a certain value, usually 2.5 in soccer.
It is possible, and rather easy even, to get the expected total score from the over/under odds, at least if you assume that the number of goals scored by the two teams follows a Poisson distribution. This is the same assumption that makes the method for extracting the expected goals from HDW odds possible. The Poisson distribution is really convenient and reasonable realistic probability model for different scorelines. It is controlled by a single parameter, called lambda, that is also the expected value (and the expected goals in this case). One convenient property of the Poisson is that the sum of two Poisson distributed variables with parameters lambda1 and lambda2 is also Poisson distributed, with the lambda being the sum of the two lambdas, i.e. lambdasum = lambda1 + lambda2.
So how can you find the expected total number of goals based on the over/under odds? First you need to convert the odds for the under outcome to a proper probability. How you do this depends on the format your odds come in, but in R you can use the odds.converter package to convert them to decimal format, and then use my own package called implied to convert them to proper probabilities.
After you have the probabilities for the under probability, you can use the Poisson formula to find the value of the parameter lambda that gives the probability for the under outcome that matched the probability from the odds. In R you can use the built-in ppois function to compute the probabilities for there being scored less than 2.5 goals when the expected total goals is 3.1 like this:
under <- 2.5 ppois(floor(under), lambda=3.1)
This will give us that the probability is 40.1% of two or less goals being scored in total, when the expected total is 3.1. Now you can try to manually adjust the lambda parameter until the output matches your probability from the odds. Another way is to automate this search using the built-in uniroot function. The uniroot function takes as input another function, and searches for the input value that gives the result 0. We therefore have to write a function that takes as input the expected goals, the probability implied by the odds, and the over/under limit, and returns the difference between the probability from the Poisson model and the odds probability. Here is one such function:
obj <- function(expg, under_prob, under){ (ppois(floor(under), lambda=expg) - under_prob) }
Next we feed this to the uniroot function, and gives a realistic search interval for the expected goals, between 0.01 and 10 in this case, and the rest of the parameters. For this example I used 62% chance of there being scored less than 2.5 goals.
uniroot(f = obj, interval = c(0.01, 10), under_prob = 0.62, under = 2.5)
From this I get that the expected total goals is 2.21.
You might wonder if it is possible to get the separate expected goals for the two teams from over/under odds using this method. This is unfortunately not possible. The only thing you can hope for is to get a range of possible values for the two expected goals that sums to the total expected goals. In our example with the expected total goals being 2.21, the range of possible values for the two expected goals can be plotted as a line like this:
If course, you can judge some pairs of expected goals being more likely than others, but there is no information about this in over/under odds alone. It might be possible, I am not 100% sure, that other non-Poisson models, which would involve more assumptions, could exploit the over/under odds to get expected goals for both teams.
Great post. Question: Is there any function in Excel that would allow you to calculate the implied expected goals from either a set of 1×2 odds and/or O/U 2.5 odds?
I don’t have access to excel right now, so I can’t try it, but there is a built-in function for Poisson probabilities and solvers and opstimizers, so it should be possible to implement the procedures in Excel.
Any chance you can help out with how the excel poisson prob work? I can’t figure it out. Assuming the input is that the
the over/under is 2.5 and the probability for the under is 0.45?
Use the Poisson function as this: POISSON.DIST(2,mean,TRUE), then you need to select the mean so that the results matches the probability of under 2.5 goals. You use Excel’s built-in solver to find where the difference between the result and your probability is 0.
Dear Opisthokonta,
Could You please describe how would You bet on the draw? How would you choose the ideal draw bet?
Thanks!
I guess it’s the same as bets on any other outcome.
Would there be a way to calculate the expected goals of each individual team by using the bookmaker’s implied odds to win along with the over/under odds?
You can try to plot the line of the sum of the two expected goals, and then plot the point of the expected goals from the win-draw-lose odds. They will usually not coincide exactly, but the point will usually be pretty close to the line. One possible way to combine the two is to select the midpoint between the point and the nearest point on the line, or something.
How about we assume the implied probability from market odds is Dixon Coles model?
Say we have 3 different odds on hand, under-over-1.5/under-over-2.5/home-draw-away and then solve the system of equations using DC(E)=implied prob(E), where E=under-1.5/under-2.5/draw. In this way we should get lambda, mu, rho solved.
What is your opinion?
Btw, do you have any thought that related to expected goal, xG?
It should be possible, take a look at the discussion I had with octonion on twitter a couple of months ago: https://twitter.com/octonion/status/1421459078367358985
In the expg_from_probabilities function in the goalmodel package you can set rho youself, but at the moment it cant find rho from a set of odds from a single match. You can however try to tweak rho using odds from several matches.
https://github.com/opisthokonta/goalmodel
Thanks for reply.
I am also wondering if dixon coles model can apply to half time result? Should we have different tau function as the goal distribution is quite different from that for full time result?
The DC model should work for half time scores, but I havent looked into it myself. It might very well be that the rho parameter will be different if you look at first half only. There are of course fewer goals scored in just half the time, which should give more probability for low scoring goals, which is where the DC adjustment is applied.
Opisthokonta: I want to download your package, but keep getting this issue:
WARNING: Rtools is required to build R packages, but is not currently installed.
Please download and install Rtools 4.0 from https://cran.r-project.org/bin/windows/Rtools/.
Downloading GitHub repo opisthokonta/goalmodel@HEAD
WARNING: Rtools is required to build R packages, but is not currently installed.
Please download and install Rtools 4.0 from https://cran.r-project.org/bin/windows/Rtools/.
Error: Failed to install ‘goalmodel’ from GitHub:
Could not find tools necessary to compile a package
Call `pkgbuild::check_build_tools(debug = TRUE)` to diagnose the problem.
This happens after I get an error about not having additional build tools.
I think you need Rtools installed to build the package. The error message gives you the link to where it can be downloaded.
Great post! Question — if you added a variable on asian handicap, would it be possible to separate the EV of total goals into EV of home goals and away goals? How would you do that in the optimization part?
I am not familiar with asian handicaps, so I am not sure how to incorporate those types of odds to the analysis.
Hi, thanks for this. I wonder if you have any more thoughts on making the parameters derived from the 1×2 market match up with the total implied by the o/u market? What extra information exists in the 1×2? Thanks.
No, I haven’t any more ideas other than those discussed in some of the other comments.