Euros prediction (2)
This is a follow up on my previous post. The group stage is now finished and I have updated the model as more games were being played. In particular, I have slightly changed the set up by adding a smooth time trend, modelled as a RW(2), based on the date in which each game in my dataset was played (considered as a unique identifier). This basically accounts for the “historical performance” by each team and has slightly improved the model performance, I think.
I think the resulting trends make sense — for instance, Italy were rather poor at the beginning of the observation period (and going worse as time progresses towards successive failures in 2014-2018), but massively on the up in recent times (since Mancini took over — not to brag, but basically the Italian national team is run by the Sampdoria golden generation, so this was to be expected…). Many of these trends are kind of obvious, but I found a couple of them interesting — for example, the rise in Denmark historical performances may explain why the model perhaps tend to overestimate their current outcomes (though in the end, it didn’t get it too wrong, as they managed to go through).
The results for round 3 of the group stage are shown in the graph below (in red, I’ve highlighted the actual result of the game).
I think I’m relatively happy with the performance of the model, considering that round 3 is when a lot of the tactical thinking may happen (particularly with teams that may know they go through with a draw — it was that bad this time, I thought, but I guess some of the games were a bit “meh”…). In quite a few cases, the outcome most likely according to the model was in fact the realised one. In others, while not bang on the money, the actual result was among the “dark tiles” (those that the model deemed relatively, if not most likely). Some predictions were quite off — for example, the model didn’t give much credit to Sweden beating Poland 3-2 (in fact, Poland were favourite to win the game, according to the model) and Germany were supposed to do much better than they actually did against Hungary (then again, most pundits probably would as well…).
I’ll try and make up the time to update the model for the knockout stage.