MATCHWEEK —
ENSEMBLE MODEL · LIGHTGBM + POISSON
recent accuracy
—
PAST PREDICTIONS
HOW THE MODEL PERFORMED ON COMPLETED MATCHWEEKS
MODEL DETAILS
ARCHITECTURE · FEATURES · CROSS-VALIDATION PERFORMANCE
CROSS-VALIDATION PERFORMANCE
Walk-forward CV — train on seasons 1..k, validate on k+1. 9 folds, 2015-16 through 2025-26.
Winner accuracy
55.8%
naive: 44.1%
Goals MAE
0.895
naive: 1.006
Exact score %
11.4%
naive: 8.3%
Log-loss
0.941
vs 1.315 baseline
ENSEMBLE ARCHITECTURE
Winner prediction
10% LightGBM + 90% Poisson
Poisson dominates — better calibrated for draws and away wins
Goals prediction
50% LightGBM + 50% Poisson
Even blend — LightGBM captures non-linear interactions, Poisson gives count-appropriate rates
FEATURES 34 TOTAL
Season-level stats
- xG For/90, xGA/90, xGD/90
- xA/90 (assists per 90)
- PPDA (pressing intensity)
- OPPDA (opponent pressing)
- Shot on Target %, Save %
Rolling form
- Last 5 match points
- Last 10 match points
- Last 5 goal difference
- xG conversion rate (38-game)
Contextual
- Days rest (capped 30d)
- H2H avg pts (last 5 meetings)
- Matchday number
- Elo rating + differential
- Closing market odds (B365 + Pinnacle)
DATA SOURCES
FBref
400+ squad/opponent stats via ScraperFC
Understat
Match xG, season table, player stats via Selenium
SofaScore
Team and player league stats via ScraperFC API
Football-Data.co.uk
Closing odds — B365 + Pinnacle, all 11 seasons