Supervised learning for an amount rather than a result, such as goals over a season. Judged on how far it misses on seasons it never saw.
Linear regression
Football question: How many shots is a goal worth?
How it works: Draws the straight line through the data that makes the squared misses smallest; its slope is what one more of something is worth.
Good at: Every weight reads as a number. One goal every 3.1 shots on target, and several features at once, each with the others held the same.
Watch out for: The world can change under it. Shots alone fell from 65% explained to 30% on the test seasons after recorded shots jumped.
Don't use it when: The relationship bends, or what you measure is recorded differently from the seasons it learned on.
Time series and ARIMA
Football question: How many points will a club take next season?
How it works: Uses the club's past seasons in order. Exponential smoothing keeps a level and moves it 70% of the way towards each new season.
Good at: Smoothing over every past season beat last season alone, usually within 9 points. No momentum from match to match, which is worth knowing too.
Watch out for: Club histories are short, and much of the long memory is just Celtic and Rangers being far ahead every season.
Don't use it when: Something big has changed, such as a new owner or a promotion, or you need full ARIMA without a long series to fit it.
Survival analysis
Football question: How long does a manager last, when some are still in the job?
How it works: Kaplan-Meier works through the departures in order, counting each tenure for as long as it was watched, and draws the share still going at each point.
Good at: Uses the unfinished stories. Half of top-flight managers are gone within 465 days; averaging only the finished jobs says 366.
Watch out for: Stopping the clock has to be neutral. Relegation isn't, since a relegated manager is often sacked soon after.
Don't use it when: The question is how many, not how long, or almost every story has already ended. Then a plain average does the job.
Change-point detection (CUSUM)
Football question: Is it a bad run or a bad manager?
How it works: Keeps a running tally of points dropped below the odds, forgives a little each game, and sounds an alarm when the tally passes a level set in advance.
Good at: Catches a real change as early as the noise allows, and the trade-off between missing one and a false alarm is set openly.
Watch out for: Football's changes are small next to luck. The alarm that caught 71% of sacked managers also went off on 54% of par ones on the same fixtures.
Don't use it when: The change you're hunting is tiny next to the game-to-game wobble, and you have only a few dozen games to see it.
Cohort analysis
Football question: Do midfield partnerships get better the longer they play together?
How it works: Groups partnerships by when they first started together and follows each group game by game, measured like for like against what the two sides' strength predicted.
Good at: Shows how groups change with time, and the table shows how fast they thin out. 1,859 partnerships in one season.
Watch out for: Survivorship. Partnerships seem to settle after a poor first game, but the same 130 followed for 15 games never improved; the poor ones were broken up.
Don't use it when: Comparing rows as they stand when members leave for a reason. Keep the cohort fixed and follow the same members.
Hypothesis test (shuffle test)
Football question: Are TV games duller, or is the gap luck?
How it works: Assumes no difference, deals the group labels out again at random thousands of times, and counts how often luck alone matches the real gap. That share is the p-value.
Good at: Needs no formula and works for any measure. Moved games' extra 0.42 goals a game turned up in 2 of 20,000 shuffles.
Watch out for: A small p-value says the groups differ, not why. Without Celtic and Rangers the gap fell to +0.17, with p = 0.20.
Don't use it when: You're trying split after split until one comes up. Of 1,000 coin-toss splits, 49 looked significant.