Causal Inference: Understanding What Actually Drives Outcomes

## Correlation vs Causation in Sports Data Sports data is full of correlations that are not causal. A bettor who bets on correlations without understanding causality will find that exploitable patterns disappear when the underlying causal structure changes. **Example correlation:** Teams that have won their last 3 matches win their next match at a higher rate than teams that have lost their last 3. Is this causal (winning builds confidence and momentum) or spurious (it reflects underlying quality differences — better teams tend to win consecutively)? Answer: primarily the latter. Control for team quality, and the "recent wins" effect largely disappears. ## The Instrumental Variable Approach When the causal effect of a variable cannot be identified from observational data (because confounding is present), instrumental variables (IVs) provide a path to causal identification. **Sports example:** Does playing in the Europa League midweek cause worse league performance at the weekend? Teams that play in Europe are better teams — they perform better in the league regardless. A simple correlation confounds quality with fixture congestion. IV approach: find a variable that determines European competition entry but is otherwise unrelated to ongoing league performance (e.g. a points threshold rule that randomly assigns teams around the cutoff to European competition). Use this threshold as an instrument. ## Regression Discontinuity in Sports When a threshold determines treatment (e.g. teams finishing 4th qualify for Champions League; 5th do not), comparing teams just above and just below the threshold provides a near-causal estimate of the treatment effect. This approach has been applied to estimate the causal effect of Champions League participation on subsequent league performance — useful for outright betting models. ## The Practical Implication Before including any feature in a betting model, ask: "Is there a plausible causal mechanism linking this feature to match outcomes?" Features without causal grounding are more likely to be overfitting patterns that disappear out of sample.