Ask ten fans who the best team in the NBA is and you will hear ten answers. Ask a ratings model and you get a number: how many points better than an average team, per game. That single idea, measured carefully, explains a surprising amount of basketball.
Two ratings per team: scoring and defending
In ProbaPredict’s basketball model, every NBA and EuroLeague team has two ratings in points. One says how many more points it scores than an average team. The other says how many more it allows. They are learned from every game since 2015 and move a little after each result: more early in a season, when we know least, and less once the picture is clear.
A game’s projected score starts from the league’s average points, then adjusts for both teams’ ratings, home court and rest. In the NBA, the second night of a back-to-back costs a team about 1.5 points. Preseason games count for rest, but never for ratings.
Why basketball is more predictable than football
Basketball has far more scoring events than football, so luck evens out faster. In a back-test on six seasons the model never saw while it was being tuned, it picked 65.2% of NBA winners. Simply backing the home team picked 55.3%. In the EuroLeague the gap was narrower: 64.4% against 62.1% for the home side.
That is useful, but it is not a crystal ball. A typical NBA result lands about 13 points away from the projected margin. A forecast of Denver by 8.5 is a forecast, not a promise, and a 13-point swing is ordinary.
What it looks like on a real night
On the NBA predictions page for Tuesday 27 October, Oklahoma City were 77% to beat Phoenix and won 125-114. Denver were 73% to beat Golden State and lost 114-134. Neither result broke the model. A 73% favourite should lose roughly one game in four, and across a full NBA slate several of those happen every week. Win chances on the page never go above 90% or below 10%, because the back-test showed the most lopsided games were slightly less certain than the raw numbers suggested.
Honest limits
Small samples mislead. At the time of writing the site’s live record for NBA games is five correct picks from nine graded, which tells you almost nothing yet. Over a season, the record becomes meaningful; after a week, it is noise. Injuries and late rest decisions also move games, and a rating built on past results cannot see news that has not happened yet.
The takeaway
If you want to judge a team, look at points scored and allowed against the league average, not the win-loss column alone. If you want to judge a forecast, ask for the track record, including the misses. The same applies to any prediction you read this season.
Mike Jeremy. is an editor at ProbaPredict, which publishes sports model probabilities and grades them in public. For information only.
