As with many things in sports analytics, it started with Bill James.
James was looking for a way to translate a baseball team’s runs scored and runs allowed into wins and losses. He noticed that the relationship could be expressed as follows:
The above could then be used to express a team’s winning percentage as a function of runs scored and runs allowed:
Because of the squared terms in the equation above, James chose to dub this the Pythagorean formula. James found that when he used this formula to predict a team’s win total given its runs scored and runs allowed, he would usually get within plus or minus four wins of the team’s actual total.
As it turns out, James’ formula is actually a logit model in disguise:
Solving the expression above for winning percentage yields:
Where, in the case of baseball, the coefficient beta has a value of approximately 2.
Of course, James’ finding eventually bled over into sports*, including basketball.
* The person often credited with starting this movement into other sports is former STATS researcher and current Philadelphia 76ers president of basketball operations Daryl Morey, although that narrative may be apocryphal. Morey definitely did work in this area, but whether or not he was the first to do so is up for debate.
One of the earliest online references I could find for using James’ structure to predict team wins in the NBA was Dean Oliver’s old website Journal of Basketball Studies. Oliver recommended an exponent of 16.5 to convert points scored and points allowed into wins:
In the first edition of Pro Basketball Prospectus (2002), John Hollinger used a simple linear model to predict teams wins using average point differential (D):
But by the 2005 edition of his NBA annual — which was now titled Pro Basketball Forecast — Hollinger settled on a Pythagorean model with an exponent of 16.5, the same as Oliver.
When I launched Basketball-Reference.com back in 2004, I found that an exponent of 14 produced the lowest root-mean-square error (RMSE) across all team seasons in the NBA:
The table below presents the RMSEs (per 82 games) for these three models by decade, as well as across all seasons (listed Oliver, Hollinger, and Kubatko):
1940s — 6.58 / 7.29 / 4.95
1950s — 4.81 / 4.56 / 3.78
1960s — 3.38 / 3.29 / 3.64
1970s — 3.59 / 3.51 / 3.26
1980s — 3.07 / 3.00 / 3.09
1990s — 3.68 / 4.07 / 3.22
2000s — 3.41 / 2.81 / 2.81
2010s — 3.42 / 3.28 / 2.99
2020s — 4.11 / 4.97 / 3.56
Overall — 3.66 / 3.66 / 3.22
A few notes about these results:
An exponent of 14 (Kubatko) is superior in six decades (1940s, 1950s, 1970s, 1990s, 2010s, and 2020s) as well as overall.
The simple linear model (Hollinger) is superior in two decades (1960s and 1980s).
The simple linear model and an exponent of 14 produce essentially the same results in the 2000s.
An exponent of 16.5 (Oliver) does not produce the lowest RMSE in any decade.
From the above, it’s clear that although an exponent of 14 does the best overall, it does not do as well in extreme eras. Here’s a list of average points per game by decade:
1940s — 73.8
1950s — 90.1
1960s — 115.0
1970s — 108.3
1980s — 109.3
1990s — 101.3
2000s — 96.9
2010s — 102.2
2020s — 111.5
The shot clock was not introduced until the 1954-55 season, so scoring levels were way down for the 1940s and first half of the 1950s. On the other hand, scoring levels were never higher than they were in the 1960s, as the average team scored 115.0 PPG. Not coincidentally, these three decades have the highest RMSEs when using an exponent of 14.
To try to fix this problem, I decided to blend the best of the simple linear model and the Pythagorean method by incorporating average point differential into a logit model. This resulted in the following formula to predict team wins:
For example, the 2021-22 Denver Nuggets had an average point differential of 2.30 PPG. The predicted win total for the Nuggets using the model above is:
In this case the Nuggets actually won 48 games, so the formula was almost right on.
Let’s compare the RMSEs for the logit model to the RMSEs for the Pythagorean model with an exponent of 14 (Pythagorean listed first):
1940s — 4.95 / 3.45
1950s — 3.78 / 3.39
1960s — 3.64 / 3.32
1970s — 3.26 / 3.37
1980s — 3.09 / 2.94
1990s — 3.22 / 3.24
2000s — 2.81 / 2.77
2010s — 2.99 / 2.98
2020s — 3.56 / 3.80
Overall — 3.22 / 3.14
As you can see, the RMSEs in the three extreme decades — the 1940s, 1950s, and 1960s — have been reduced with the logit model, in some cases dramatically so. It’s interesting that the logit model is thus far performing worse in the 2020s, but that could be a fluke because we only have three full seasons to work with.
I dug into the issue with the 2020s a bit and found that three teams missed their prediction by eight or more games in 2021-22 using the logit model (error is observed minus predicted)
Boston Celtics (–9.3 wins)
Portland Trail Blazers (8.6 wins)
Utah Jazz (–8.4 wins)
Only one team missed the mark by that much using the Pythagorean model:
Celtics (–8.0 wins)
The Trail Blazers were an extreme team last season, losing 15 games by 30 or more points, five more such losses than any other team in NBA history. This skewed their average point differential, which is used in the logit model.
The Celtics were on the opposite end, tying an NBA record by recording three 40-point wins. Once again this skewed their average point differential, causing the logit model to be off by almost 10 wins.
For now I’m chalking this up as a weird one-season fluke, but it’s worth keeping an eye on.



Awesome! Thanks for sharing!