Glossary
What the words on this site mean.
Each term the pages use, said in plain words first. Nothing here needs another entry to be read.
36
Terms, each in plain words
18
Betting words
14
Scoring a forecast
4
Keeping a test honest
Betting words
What a price, a line and a bet are called.
- moneyline
- the price on which team wins the game outright.
- spread
- the number of points the market expects the favoured team to win by.
- total
- the combined points of both teams that the market posts as its line.
- cover
- win by more than the market expects, or lose by fewer.
- push
- a final margin exactly on the line, so the bet is neither won nor lost.
- parlay
- one bet on several results at once, which pays only if every one of them happens.
- leg
- one of the results a combined bet needs.
- vig
- the bookmaker's cut, built into the price of both sides.
- hold
- the share of the money staked that the bookmaker keeps on average.
- de-vig
- with the bookmaker's cut taken out of the price.
- break-even
- the share of calls that must clear just to not lose money at the price paid.
- closing line
- the last price posted before the game started.
- CLV
- how much better the price taken was than the last price before kickoff.
- exchange mid
- the price halfway between the exchange's best offer to buy and best offer to sell.
- lean
- a side we published before the game.
- edge
- a real, repeatable advantage over the betting price.
- settled
- graded once the result was official.
- void
- cancelled with the stake returned, usually because the player did not take part.
Scoring a forecast
How a stated chance is checked against what happened.
- calibration
- how often our stated chances match what happened.
- calibration error
- the average gap between a stated chance and how often it came true.
- reliability curve
- each band of stated chances drawn against how often those outcomes happened.
- Brier score
- the average squared miss of a forecast, where lower is better.
- log loss
- a score that punishes a confident miss hardest, where lower is better.
- AUC
- how often a forecast ranks an outcome that happened above one that did not, where 0.5 is a coin toss.
- interval
- the range the true figure plausibly sits in, given how much luck could move it.
- Wilson interval
- a range for a rate that stays sensible near 0% and 100%.
- bootstrap
- redrawing the sample many times to see how far a figure moves by luck.
- pp
- percentage points.
- naive baseline
- a forecast that simply repeats each player's rate from the season before.
- correlation
- how much two results tend to happen together.
- residual
- what is left of a figure once the expected part is taken away.
- shrinkage
- pulling each player's figure toward the average of similar players, more so when there is little data.
Keeping a test honest
What stops a result from being chosen after the fact.
- pre-registered
- written down and committed before the result was known.
- out of sample
- predicted without the result being used to build the model.
- walk-forward
- each season predicted using only the seasons before it.
- hash chain
- each row carries a fingerprint of the row before it, so editing any row breaks every row after.
Method states how a number gets onto a page, and About says what the site is for.