Indeed, it would be interesting to add more features. However, for international games, I saw that it was quite hard to improve accuracy beyond what I already have. One other approach would be to create alternative ranking systems (modified Elo...) and use them as features.
Elo ratings (such as those kept at eloratings.net) actually work quite well. Ultimately, they work like a weighted moving average.
The thing that I find hard to predict and build into my models is style of play. By style I mean: spatially-intensive, high-time-inpossession-the-ball-time (e.g. Spain with tiki-taka and Germany to some extent) versus time-intensive, opportunity-seeking/opportunity-creating (such as Brazil, Argentina, etc.)
Why? Because passing-intensive teams seem to display more of an own effect -- they fall or rise on the strength of their team, since it's an intricate, very technical and collaborative style. The results of opportunity-seeking teams are much more dependent on the strength of the adversary -- i.e. much more Elo-like.
Ideally, I'd be able to infer from the data a (exponentially biased to recent games) own-team/spatial play dependence factor as opposed to a strength-of-opponent/opportunity-seeking factor. In principle if all victories were explainable by a combination of those two variables the Elo residual/surprise would be a measure of this, but hey, teams get better/worse at opportunity-seeking too, even teams specialized in tiki-taka.
I tend to agree with you, these factors would not improve the already good accuracy, however due to KO phase the games are all-or-nothing ( unlike other types of international games ), so some little factor ( extreme heat or a missing player ) could incline to a victory or not.
In the case of missing a player, take Uruguay for example: they just lost their best player after he bit (yes, à la Mike Tyson) the shoulder of a player on his last game. Fifa banned him for some 8 games, so he'll be absent for the rest of the cup. Uruguay just took a big blow.
They are quite different indeed. It would be interesting to compare, at the end of the tournament, the likelihoods of both models on all matches. This would give a pretty good idea on how good were the models.
Indeed you are right, the figure showing the "chance to reach, chance to knock-out" was accidentally an outdated one (from a test simulation with a lower number of trials). The real figure will be updated soon.
They are quite close though, good job noticing that!
--
it has been updated now
