What are the chances that it is going to rain tomorrow? Who will win the next election? What about the Super Bowl?
Statistical models can give a “best-guess” estimate, helping to update the confidence level that a certain event will occur. Look no further than the local weatherman, though, to see the limitations of statistics. No model can truly predict the future, and the more complicated the scenario, the tougher it gets.
With that disclaimer out of the way, what do forecasting models say about the upcoming Duke-North Carolina game? Will head coach Jon Scheyer and the Blue Devils continue their hot streak against their in-state rivals, or will an experienced Tar Heel squad reclaim the rivalry in Chapel Hill?
Let’s say a model gives North Carolina a 77.67% chance to win. That number sounds great for the Tar Heel faithful, but before jumping to any conclusions and storming Franklin Street, it’s necessary to first examine the context of where that prediction came from.
To start, consider each team’s body of work. Although Duke had a recent hiccup at home against Pittsburgh, the two rivals sit squarely atop the ACC table by nearly every statistical measure. They are the only two teams in the conference ranked inside the AP Poll top 25 and both sit more than 10 points ahead of the average ACC team in net rating — a measure of a team’s offensive and defensive efficiency.
While the formula for success is somewhat consistent for all college basketball programs — solid shooting, effective defense — each team beats its opponents in unique ways. North Carolina excels at getting to the free-throw line and limits opponents’ second-chance and fast-break points at top-five rates nationwide. In fact, the Tar Heels’ 7.0 second-chance points allowed per game is the top mark of any team so far this year. On the other side, the Blue Devils are one of the best teams in the country at taking care of the ball: Their 9.3 turnovers per game and assist-to-turnover ratio of 1.73 are both eighth-best.
Figure 1 shows some of these rates compared to the league average. Note that a lower turnover percentage is desirable, and FTA rate represents a team’s attempted free throws as a percentage of their field-goal attempts. Both sides are better than the average ACC team in each category, and in net ratings, significantly so.
With some basic knowledge about each side, the next step is to find a model that can consistently predict both teams’ successes. This article will use the cbbdata library, available publicly on GitHub and developed by former Duke men’s basketball manager Andrew Weatherman. Within that library, there are functions that can predict the outcome of a game or even an entire season for a given team.
These predictions are calculated from an algorithm created by Bart Torvik, who tracks and records college basketball data similar to what Ken Pomeroy does for kenpom.com. Essentially, the algorithm estimates how many points each team will score per possession and how many total possessions each team will get. Then, it multiplies these two values to obtain a predicted final score and consequently calculates each team’s chance of winning.
Figure 2 shows Duke and North Carolina’s projected win percentage compared to the actual outcomes for all ACC games from 2020-2023. The dotted vertical line represents a 50% win probability. Any point to the left of the line is a projected loss, and any point to the right is a projected win. The solid horizontal line represents a zero-point differential, meaning that any point below it is an actual loss for that team and any point above it is an actual win. Put simply, the points in the top-right and bottom-left quadrants represent correct predictions, whereas points in the top-left and bottom-right quadrants are places where the model got it wrong.
As we can see, the model mostly got it correct. When the model predicted a win, the actual result was a win 72.03% of the time. The same is true for predicted losses, with the model correctly predicting a loss to 69.44% accuracy. For something as complicated as a basketball game, these are respectable figures.
Surely this model can predict a Duke-North Carolina game, right?
Figure 3, which contains all regular-season games between the two teams from 2015-2023, shows that this is not entirely the case. The x and y axes are the same in this graph as those in Figure 2, but now the shape of each point corresponds to home and away games for the Blue Devils.
The model correctly predicted wins just 58.33% of the time and correctly predicted a loss 50.0% of the time. That means that given a projected Duke loss, the chances of the Blue Devils actually losing that game are no better than a coin flip.
Recall that the game prediction for Saturday’s matchup gives the Tar Heels a 77.67% chance to win. North Carolina comes into the game with a stronger resume and looks like the better team on paper, but that does not always paint the full picture. Take the Tar Heels’ historic win in former head coach Mike Kzryzewski’s final game at Cameron Indoor Stadium, or the Blue Devils’ 2016 victory against then-No. 3 North Carolina at the Dean E. Smith Center. Ask any fan from either side, and they will tell you that this game is just different.
The story of the Duke-North Carolina rivalry is written by the moments that no one sees coming. Statistics inform a lot of things, but when it comes to predicting these games, numbers don’t seem to mean much at all.
All data were sourced Sunday, Jan. 28. All of the calculations and code used in this article can be found in the GitHub repository linked here: https://github.com/D-Fenoglio/duke-unc-rivalry-forecasting.
Editor’s note: This piece is one of many in The Chronicle’s 2023-24 Duke men’s basketball rivalry edition. To read more, click here.
Get The Chronicle straight to your inbox
Signup for our weekly newsletter. Cancel at any time.
Dom Fenoglio is a Trinity sophomore and an assistant Blue Zone editor of The Chronicle's 119th volume.