Demo mode - connect Supabase (.env.local) to show live predictions.
Performance
Every week we measure the model's log-loss on played matches - the metric that rewards honest probabilities, not lucky picks - and compare it to the bookmaker, the toughest benchmark there is. Lower = better. No cherry-picking, no hidden weeks.
1.010
model log-loss (avg.)
0.967
bookmaker log-loss (avg.)
+0.043
gap to bookmaker
bookmaker still ahead
272
matches scored · 6 wks
Weekly log-loss Signal (model) Bookmaker
25 May+0.051
01 Jun+0.039
08 Jun+0.044
15 Jun+0.026
22 Jun+0.071
29 Jun+0.029
29 Jun47 matchs · accuracy 51%
Signal 0.994Bookmaker 0.965
Back to normal
22 Jun44 matchs · accuracy 48%
Signal 1.043Bookmaker 0.972
Tough week - many draws
15 Jun48 matchs · accuracy 52%
Signal 0.987Bookmaker 0.961
Well calibrated
08 Jun46 matchs · accuracy 50%
Signal 1.012Bookmaker 0.968
Decent week
01 Jun42 matchs · accuracy 50%
Signal 1.005Bookmaker 0.966
25 May45 matchs · accuracy 49%
Signal 1.021Bookmaker 0.970
End of season - squad rotation
Why publish this? Because sites claiming "92% win rate" never show their full history. Beating the bookmaker on log-loss is extremely hard - its odds aggregate the information of millions of bettors. Our goal is to close the gap week after week, and to document the method that gets us there. Average accuracy: model 50% · bookmaker 54% (state of the art sits between 50 and 58%).