Firstly a % and a probability are essentially the same thing.
You would need to define your random variables better. Let $A$ denote the random variable for team $A$ and can take values in $\{-1,0,1\}$. This way $P(A = 1)$ is the win chance, $P(A=-1)$ is for loosing, and $P(A=0)$ is if a draw occurs.
Normally you would have a set of head-to-head scores. Something like $\{(A=1,B=-1),(A=1,B=-1),(A=0,B=0),(A=-1,B=1)....\}$.
Assuming this sequence is not correlated (generally poor assumption, as location, mentality, physical fitness change with time). You can answer questions like $P(A=1,B=-1)$, by look at the head-to-head history of the teams.
A more fine tuned approach would be to model things such as the existance of a home game, and the relative observed strength between the teams as continuous random variables (a Gaussian distribution).
An interesting example is provided here:
https://arxiv.org/pdf/1501.05831
where the authors consider modeling a Gaussian based on $\mu = \Delta s + h$ where $\Delta s = s_1 - s_2$ is the relative strength between the teams, and $h$ is a home playing bias term. Therefore if you play at home you have more bias to win. Using this approach it is then possible to place distributions over certain variables, and make a fully Bayesian model, which also includes randomness on the day.
Basically you need to define a model, and use optimisation to learn the relevant features based on observed data - a more advanced version of the simple joint distribution discussed before. Definitely you should be fully comfortable with Bayesian inference before proceeding in this direction.
The approach here is even simpler than the other link. Look at this one first:
https://publisher.uthm.edu.my/ojs/index.php/ijie/article/view/2780
