search

alan-turing-institute / AIrsenal

Machine learning Fantasy Premier League team
MIT License
295 stars 87 forks source link

Weight recent matches more heavily in player and team models #141

Open jack89roberts opened 4 years ago

jack89roberts commented 4 years ago

At the moment we use the last 3 seasons of data with all matches being weighted equally, i.e. matches 3 seasons ago are treated as important as matches this season when determining a team's (or player's) strength. Form can vary a lot over 3 seasons, and also within seasons, so we should add time/form dependencies to the models.

anguswilliams91 commented 4 years ago

Just a quick comment on this one. I looked at allowing the team abilities to vary over time in #83 and didn't see a lot of variability. I don't necessarily believe this result, but the approach was quite a natural way to try this, so perhaps a good starting point on this issue would be to try to understand this model better (the one in sandbox/dynamic_model.stan)

saba83rish commented 4 years ago

Form is an important concept personally for picking a player, do we incorporate that from FPL API? or are there other ideas you are thinking here?

nbarlowATI commented 4 years ago

Hi @saba83rish yep, for sure it would be good to consider form in the player model, but we haven't really looked at that yet. I think we were planning on incorporating something into the team level as a first step (which is what is in the dynamic_model.stan ) but it's true that player-level form is also important. Ideas on how to do it would of course be very welcome! :)

saba83rish commented 4 years ago

@nbarlowATI the FPL shows a form value in it, probably its available via FPL API.