Honest numbers.
Scored on 297 matches it never saw in training. Lower log-loss is better.
Six seasons, each predicted only from earlier seasons. The blend beats XGBoost alone every season; Bet365 stays ahead by about 0.02.
23
pre-match signals
XGBoost + logistic
blended, then calibrated
Strict time order
never trained on the future
Beats naive
trails Bet365 by ~0.02 log-loss
Log-loss, lower is better
Naive guess
always the training average
1.0674
XGBoost alone
23 fair-play signals
0.9613
Matchday Fate
XGBoost + logistic blend, calibrated
0.9596
Bookmaker
Bet365, the benchmark
0.9698
On this small window the model edges the bookmaker, but that gap is inside the noise. Over full seasons the bookmaker is still ahead by about 0.02 — expected, since it prices in team news this model never sees. Every change to the model has to beat the old one in all six past seasons before it ships.
What goes in
- Elo strength (plain + margin-adjusted)
- Form over last 5 and 10
- Goals, shots, shots on target, corners
- Rest days
- Head-to-head streak
What stays out
- Betting odds of any kind
- Anything known only after kickoff
- Lineups and injuries (no free source yet)
Research demo, not betting advice.