I'm running an machine learning based algorithm team for this year's FPL season and will be (hopefully) documenting its progress here throughout the season. The idea is to have multiple iterations of the algorithm before the season starts and stick to one model for the season.
The first draft is based on expected goals and assists from the previous 4 gameweeks modeled using a SVM classifier that predicts whether a player will score or assist in the next 4 gameweeks or not. The higher the confidence score, the more likely the player is to get picked.
This version is based on attacking potential only and currently does not take into account opponent difficulty, injuries or potential clean sheet points. The current structure is to give out transfer suggestions in the following format (this is for last season)
Using this version of the model, the following were the top 10 suggestions made for GW 1 to 4 of the new season using GW 35+ to 38+ expected data from the previous season.
Using this information and the confidence scores of the model, the following was the first draft made for the algo team. I've currently filled in the defenders manually. As I add more features to it, the model will become more sophisticated and I'll post the updates
And here's the current shape of the raw data taken from Fantasy Football Scout for the 2019-20 season. Currently, only the expected data from the previous four gameweeks form the feature set.
No comments:
Post a Comment