Correlation Is Necessary, Not Sufficient

Every football offseason produces a genre of article: teams that do X win Y% of the time. Teams that run the ball 30+ times. Teams that win the turnover battle. Teams that lead at halftime. The stat correlates with winning, the correlation is real, and the article treats that as the finish line.

It's the starting line. We learned this building a prediction model, the expensive way, more than once — and it left us with a maxim we now apply to every stat that crosses our desk: correlation is necessary, but nowhere near sufficient. A number can genuinely relate to winning and still be completely useless for predicting it. Here are the three experiments that taught us why.

Experiment 1: we doubled a stat's correlation and nothing happened

One of our model's inputs compares two teams' season-long win rates. Early in a season it was nearly worthless — in Week 1, every team is 0-0, so the feature was flat. We rebuilt it: instead of a crude "who has more wins," a continuous win-rate margin, each rate regularized toward .500 with a small prior so early-season noise didn't swing it wildly.

It worked, at the feature level, beautifully. The stat's correlation with outcomes doubled (its r² went from 0.043 to 0.072), vaulting it into the second-strongest signal in our entire pool, behind only one thing (we'll get to that).

Then we retrained the model. Accuracy went from 70.25% to 70.5% — a rounding error, well inside statistical noise. Worse, when the model's search picked its features freely, it usually didn't even select the improved stat. We had made a signal measurably stronger and the model shrugged. All that predictive power was already accounted for.

The one thing every stat loses to

The feature that dominates our pool — the reason the improved win-rate stat was only #2 — is the market's own number: the point spread and its cousins, correlation roughly double the best conventional stat we can compute.

That's the mechanism behind the shrug. The point spread is a forecast that has already absorbed team quality, form, injuries, and situation. So when your favorite stat correlates with winning, of course it does — and so does the spread, more so, having eaten that same information Tuesday morning. Adding the stat on top contributes almost nothing new. It's not that the stat is fake. It's that it's redundant.

Experiment 2: 360 stats, all of them near-zero

We pointed the whole machine at predicting over/unders — 360 computed features, two decades of games. The single most-correlated feature explained about two-tenths of one percent of the variance in outcomes (r² ≈ 0.002). Not the average feature — the best one. That's the statistical texture of a coin flip. The totals market had priced essentially everything we could measure. (Full post-mortem in our totals article.)

Experiment 3: the #1 stat that made the model worse

Here's the sharpest version. When we cleaned up our wind data, wind speed became the single most correlated feature we had for totals — #1 out of 360. By the logic of the offseason-stat genre, it should have been our best input.

We added it to the model. Accuracy went down — about two points worse on the hold-out. The most correlated feature we owned actively degraded the forecast.

Why? Wind matters in maybe 12% of games — the windy ones. As a model input applied to every game, it's noise 88% of the time, and that noise cost more than the signal earned. Wind is real. It's a genuine pattern (15+ mph games lean under, reliably). It's just not a model factor — it's a conditional filter for a specific slice of the schedule. Real signal, wrong tool.

What "sufficient" actually requires

So the next time you read that some stat predicts winning, run it through the three questions that survived our experiments:

Is it out of sample? A pattern found in the same data it's being praised with proves nothing — we've watched our own search "discover" 57%+ patterns that were pure overfitting, collapsing to coin flips the moment they met new data.
Is it already in the number? If the point spread has seen the same information — and it has seen almost all of it — your stat is describing the forecast, not beating it.
Does it survive next to the obvious predictors? A stat that correlates with winning on its own but adds nothing once team quality and the market number are already in the model isn't an insight. It's an echo.

Almost nothing clears all three. That's not cynicism — it's the actual difficulty of the problem, and it's why we're skeptical of confident single-stat claims, including the tempting ones our own research produces. A correlation is an invitation to investigate. It was never the answer.

This maxim came out of our prediction-model research. How the point spread earns its place at the top of that pool is its own piece: In Defense of the Point Spread.