## What Are Features?
Features are the input variables that a predictive model uses to generate predictions. The quality of your feature engineering often matters more than the choice of model. Good features from a simple model beat poor features from a complex model.
## The Core Features for a Football Model
**Attack strength:**
Rolling average xG scored per match (home and away separately), opponent-adjusted.
**Defence strength:**
Rolling average xGA per match (home and away separately), opponent-adjusted.
**Home advantage:**
Binary flag (1 = home, 0 = away). Plus optionally: estimated home advantage multiplier from historical win rate for this specific team at home.
**Rest advantage:**
Days since last competitive match for each team. Difference between teams' rest days.
**Form indicator:**
Points earned / expected points earned in last 5–10 matches. Captures recent trajectory.
**Motivation:**
Distance from a threshold (top 4, relegation, title) — teams fighting for something may be motivated differently. Operationalised as points gap to nearest relevant threshold.
## Feature Engineering Pitfalls
**Lookahead bias:** Using future information to calculate historical features. If you calculate a team's "season-average xG" using all matches including those after the prediction date, you are cheating — your model will appear to work better than it does on genuinely future data.
**Overfitting through feature count:** More features allow the model to fit the training data better but generalise worse. Prefer fewer, theoretically motivated features.
**Collinearity:** Features that are highly correlated (home xG and home points per match are correlated) can destabilise model estimation. Test for correlation and drop redundant features.
Create a free account to track your progress and save bookmarks.
