search

kunzaatko / pm_mode

1 stars 0 forks source link

Poisson regression as predictive model #1

Closed Many98 closed 3 years ago

Many98 commented 3 years ago

Working on poisson time independent and dependent regression predictive model. This approach is based on the assumption that number of scored goals by team fits poisson distribution.

Many98 commented 3 years ago

First tests with time independent Poisson regression shows that it gives poor results due to too many new teams in increments and updating model on small data containing previously unseen teams can result in infinite weights which raises errors.

Many98 commented 3 years ago

Model gives predictions only for matches in DataFrame which was seen while fitting. For unseen data it gives row in Dataframe [0., 0., 0.]. Simple tests (model fitted on first 5 seasons and tested on next seasons) show that model prefers predict win of home team or away team rather than draw. Accuracies calculated on previously seen data are around 40-50 %

Many98 commented 3 years ago

Further improvements can be achieved by using time dependent model or use bivariate poisson model which treats scores in single match as dependent instead of independent like basic poisson model.