Dr. Green and White's Basketball Analysis: Kenpom versus Torvik

The analysis of sports statistics can take many forms, from individual player performance to full season results. In the world of college basketball one of the most useful categories of metrics are tempo free efficiency stats. 

These stats measure things such as offensive and defensive performance on a points-per-possession basis. Combining offensive and defensive values yields an effective efficiency margin which can be used to rank teams. For NCAA Tournament selection and seeding, this class of metrics are referred to as predictive metrics as they can be used to estimate victory margins for any arbitrary matchups.

There are several different systems out there that measure and track college basketball efficiency. Two of the most notable system are the ones created by Ken Pomeroy ("Kenpom") and that created by Bart Torvik. Both systems are attempting to do the same things, but they use slightly different methods and get slightly different results.

Calculating raw efficiency is very straightforward. But the two systems differ in their subtle (and not transparent) adjustments for variables such as opponent strength, garbage time possessions, etc. Both individuals have their own website that slice and dice the data in different ways.

Most people have a personal preference for one system over the other. I prefer to use Kenpom, as it is possible (with a subscription) to download the data to comma-separated values (csv) file and I am generally more familiar with the layout of his website. Your mileage may vary.

But over the years I have wondered if one system is better than the other. In an attempt to answer that questions, I ran an experiment on the 2024-25 Big Ten season where I tracked the predictions made by each system, the point spread, and the actual margin of victory/defeat for all 180 Big Ten regular season games.  

Here is what I found.

What it Means to Be "Better"

One of the challenges in this type of analysis is that the concept of "better" can be difficult to define. Figures 1 and 2 below show the raw comparison between the actual margin of victory or defeat in all 180 Big Ten games and the values predicted by Kenpom and Torvik, respectively.

Figure 1: Actual game results plotted versus the margin of victory/defeat predicted using Kenpom efficiency margins for all 180 regular season Big Ten games in the 2024-25 season.

Figure 2: Actual game results plotted versus the margin of victory/defeat predicted using Torvik efficiency margins for all 180 regular season Big Ten games in the 2024-25 season.

If either system were a perfect predictor of game outcomes, all of the blue data points would fall onto the solid black diagonal line. Instead, both plots look like the results of a shotgun blast. Upon first glance, deciding which system is "better" is like explaining which of two random ink blots looks the most like your mother.

Fortunately, there are a few common statistical metrics that can be helpful. The dotted line and equation in each figure is the linear regression fit for each data set. The "r-squared" metric is slightly better for Torvik (0.168 versus 0.161 where closer to 1.000 is better), as is the intercept (0.212 versus 0.389 where closest to zero is better). 

However, the Kenpom data has a slightly better slope (0.895 versus 0.855 where closer to 1.00 is better) and as a result the trend line visually looks closer to the black diagonal. 

I also looked at the difference between each data point and black diagonal. In terms of standard deviation, Torvik (12.27 points) holds a slight edge over Kenpom (12.30 points). However, is terms of the sum of the absolute values of the differences, Kenpom (1715.9 points) holds a slight edge over Torvik (1729.0 points). But this difference is less than one percent over 180 games.

In terms of raw "wins" and "losses", Kenpom also has an edge by the slimmest of margins. The Kenpom predicted point differential was closer than Torvik to the actual margin in 91 of the 180 games. If just one game were to have switched, it would be a dead heat.

Despite the relatively poor correlation between the two systems and the actual results, it should be noted that the point spreads generated by the Las Vegas casinos do not fair much better. Figure 3 below shows the correlation between the posted Vegas spreads about 24 hours before tipoff and the actual game results.

Figure 3: Actual game results plotted versus the Vegas spreads for all 180 regular season Big Ten games in the 2024-25 season.

The overall look of Figure 3 is very similar to Figures 1 and 2. Based on the linear correlation parameters, the Vegas line outperforms both Kenpom and Torvik, but only slightly. The Vegas data set has a higher value for r-squared (0.170), a better slope (0.964), a lower standard deviation (12.22 points) and a smaller total absolute deviation (1699 points).

The performance of the spread should not come as a major surprise. A more detailed study of the spread versus college basketball outcomes over several decades shows that the spread can, on average, predict the final margin of games very accurately. See Figure 4 below.

Figure 4: Average margin of victory versus the final Vegas spread for the set of college basketball games from 2004 to 2025, n = 83,000.

While the performance of the Vegas spread is very good on average, the variance is very high. Over two decades, the standard deviation works out to be just over 10 points (10.34 to be exact as of 2025). 

Based on this fact, it is possible to calculate the odds that any team will win a game based solely on the Vegas line and the known standard deviation by assuming a Normal/Gaussian distribution. That correlation is shown below in Figure 5.

Figure 5: Odds that the favored team wins versus the final Vegas line for the set of college basketball results from 2004 to 2025, n = 83,000. The solid line is the fit based on a Normal Distribution with a standard deviation of 10.34 points.

In my opinion the Vegas spread is the gold standard of predictive metrics. If a system out there could consistently beat the spread, its creator would be wise to keep it secret. The experts in Vegas would then very likely modify their formulas to compensate.

If you can't beat 'em, forecast 'em!

Based on this analysis, it is clear that Kenpom, Torvik, and Vegas all gives results that are similarly correlated to the actual results of college basketball games. If we accept the idea that the Vegas spread is the best overall predictor, the value of systems like Kenpom and Torvik is not their ability to predict game outcomes. It is in their ability to forecast the Vegas line of future games.

From this point of view, the next relevant topic to explore is which system, Kenpom or Torvik, does a better job at forecasting the Vegas line? Figures 6 and 7 below show these correlations for the 2024-25 Big Ten regular season.