The PoDi prediction model, February to April 2018

I have done some work on the PoDi model. Indeed, so much work that I don’t have time to give a full report on it. So I’ll sum up everything in seven “tweets” (old style tweets: 140 characters!).

I started working on PoDi 4, but it takes time, so I created a quick-to-complete “unsophisticated” version starting from PoDi 1.

1) Evolution = 10th CF at each age. 2) Best dummy = 8th in the world. 3) GK RB 2xCD LB 2xD/CM 10 pts, RW AM/SS LW CF 15, 2nd best gets half.

Then I designed a 2nd version adding uncertainty. 1) Chance to fail: <17 year old 91%, 17 58%, 18 42%, 19 37%, 20 33%, 21 23%, >21 0.

2) Dummies randomly picked from a log curve. 3) Give points to the top 4 players: give 17.5% to each, 30% proportionally to market value.

Then we checked how the models fare. Test set: all non-English players picked before 2011. Training set: everything else.

Mean absolute error: Daniele2017 9.3, PoDi1 9.5, Daniele2010 9.9, Unso2 10.2, Unso1 10.5. But differences are not statistically significant!

Even this gets 11.0 and no significant difference. Yearly points: if 0 points the year before 95% 0, 5% 0.5; else: if age<29 +1, else -1.