Variance and Sample Size: Why Short-Term Results Lie
Ten games is not a trend. Neither is a five-game streak. Understanding variance protects you from the most common analytical mistakes in sport — and from poor decisions with money.
A simple simulation
Imagine a forecaster who is genuinely right 55% of the time. Over 20 predictions, the chance they go 50% or worse is roughly one in three. Over 100 predictions it is about one in six. Only over several hundred does skill reliably separate from noise. The same maths applies to teams: a true 0.550 baseball team can easily post a .450 month.
Regression to the mean
Extreme results tend to be followed by less extreme ones, because extremes usually involve some luck. A striker converting 30% of shots will usually fall back toward the league norm; a goalkeeper saving everything usually cools down. Metrics like xG, shot quality and point differential are "stickier" than wins and goals, which is why analysts lean on them.
How to judge forecasts
- Calibration: do 70% predictions come true about 70% of the time?
- Brier score / log loss: scoring rules that reward confident correct forecasts and punish confident wrong ones.
- Sample size: demand hundreds of forecasts before believing anyone's record — including ours.
The human side
Our brains are pattern detectors and hate randomness. We remember the dramatic win and forget the quiet losses. This is why chasing losses and "due" thinking are dangerous: past independent results do not make the next one more likely. If you ever feel the urge to recover losses, step away — see our responsible gaming resources.
Keep learning
Next: Reading Advanced Stats: xG, EPA, Net Rating and Corsi
Get one 3-minute lesson every Sunday in the free Edge Report.
Get the Edge Report