You want to make college basketball predictions and you’re smarter than the average fan and want to look to advanced metrics. Maybe it is March and you’re looking to get a leg up in your bracket, maybe you’re looking to place a few bets as gambling rapidly increases in legality, or maybe you’re just a stat nerd. It does not matter why you are interested, but college basketball metrics are crucial to being able to fully understand the sport.
While the NBA has experienced an analytical revolution in recent years, that has largely featured player individual stats. Stats such as RAPTOR, PIPM, or ESPN’s RPM have burst onto the scene as ways to judge players behind the box score. College basketball has very few outside of your standard sports reference box score advanced stats. The only stats that incorporate stats outside of the box score are a college basketball version of PIPM and PROPGATU by BartTorviik which stands for points over replacement per adjusted game at that usage. Instead, most college basketball metrics focus on team stats.
In the NBA there is SRS or simple rating system. SRS is found using the SOS stat on basketball reference along with average point differential. SRS does the job wonderfully in the NBA, and sports reference has a version of it for college as well, but it is not nearly as accurate.
This is because of how scheduling works in college basketball. In the NBA every team plays each other at least once, there are 30 teams in an 82 game schedule. The difference between the Lakers and Timberwolves was big, but it comes nowhere close to the chasm that is the difference between top college teams and the worst college teams.
Most teams in college basketball play about 31 games in the regular season, there are 357 teams in division one. So a simple SOS stat like the one on basketball reference is not sufficient. So many analysts set out to make more complex metrics.
If you pay attention to college basketball you will most likely hear people talk about things such as kenpom, T-Rank, quadrant wins, and a boatload of other metrics. Let’s get into some of the most important metrics
Different Types of Advanced Metrics
Resume Based Metrics
While these types of metrics may not be relevant in predicting who will win a matchup they are still invaluable. Resume metrics are mainly used to determine who should be seeded where come March Madness, using the quality of a team’s wins and losses to determine what they have done. They can also be useful if you believe who a team has beaten is the best way to judge who is a better team.
Wins Above Bubble (WAB)
WAB aims to calculate how many wins an average bubble team should get against each schedule and then compares it to what they actually accomplished. If you do not know what the “bubble” is, it is essentially teams on the edge of making March Madness. These can be classified by the last eight at-large bids and the first eight teams out of the NCAA tournament. These teams are usually ranked in the 35-50 range.
For example, Syracuse, a team expected to be on the bubble, plays at Rutgers on December 8th. The average bubble team has a 31% chance to win that game, this equates to about 0.31 wins. If Syracuse would win this game their WAB score will go up 0.69, if they lose their score will go down 0.31. If you sum up every game on a team’s schedule you will get their final WAB score. The number represents how many more or less wins the average bubble team would get against the team in question’s schedule. If their WAB score is well above zero you can expect them to be a quality team.
In theory, it is an excellent idea and is a great normalizer for teams that play different schedules. But while it is still very useful there are some flaws. For starters each bubble team has different strengths and weaknesses, with how matchup driven basketball can be at times, different bubble teams will have a different chance of winning any given game. Overall it is a good stat to compare teams’ schedules and how they fared against them.
The Quadrant System
This is by far the simplest metric I’ll be talking about but also one of the most used. The quadrant system is interconnected with the NET rankings (more on them later). The quadrant system aims to divide a team’s games into four quadrants while placing emphasis on where the game was played. Here is a breakdown of all of Duke’s games from last season broken down into the four quadrants, along with the quadrants’ parameters.
Quadrants one and two are considered quality games, if someone is listing a win as a “quad one win” it means they earned a quality win. If a team would lose a quad three or four-game it is usually considered a bad loss. There is some controversy over whether games against quadrant three can be considered as bad losses. For example, Duke’s loss to SFA is a quad three loss, but they are still a quality team. As a general rule of thumb if a team has more than three quad three losses it is safe to say the team has some bad losses.
Where the quadrant system falls short is that it is completely arbitrary. A road win against a team ranked 75th is a quad one win, but a road win against a team ranked 76th is considered a quad two win. In reality, the difference between the two wins is minuscule. Another flaw is that the metric used to determine where a team ranks for the quad system is NET. While relying on just one metric is already not a good idea, the NCAA still has not made the formula public and NET has multiple flaws. The quadrant system is great for placing a broad stroke for a team’s wins but has little nuance to it and you should always look more into it.
Predictive Metrics
These metrics are used to predict who the best teams are going forward, and also attempt to show who the best teams have been so far. Predictive metrics have three main components, offensive rating, defensive rating, and strength of schedule. Oftentimes the offensive and defensive ratings are weighted for strength of schedule. Different ratings add different components as well.
NET
NET is the official metric used by the NCAA and determines the quadrant system. Unfortunately, the NCAA has been frustratingly vague on how NET is calculated, but this is what we do know.
There are two factors incorporated into NET. The first is adjusted net efficiency which includes efficiency, the strength of opponents, and location of games. The second factor is team value index. Team value index is a strength of record stat that is based on opponent, location, and winner.
Overall, NET has shown to be not nearly as accurate as the next two metrics I’ll be going over. Nor can you do as much with it. But it is worth going over because of how it is intertwined with the quadrant system.
T-Rank and Kenpom
Kenpom is the industry standard of college basketball rankings. Created by Ken Pomeray over 15 years ago there has been a significant correlation between teams that are overseeded or underseeded according to kenpom and how they do in the tournament. The two seeded Duke team that fell to Leigh was all the way down at 18th in kenpom. The eight seeded Wisconsin to upset Villanova in 2017 was in the top 15 of kenpom. The only downside of kenpom is outside of his base rankings things such as game odds and archives of rankings throughout the years are locked beyond a paywall. That is where T-Rank created by Bart Torviik comes in.
While not as accurate and a slightly different formula than kenpom, T-Rank uses a similar way to rank teams, and data beyond just the rankings are easily accessible.
The core of these rankings is offensive and defensive efficiency points scored and points allowed per possession (“PPP” = points per possession, often rendered as points per 100 possessions). Both rankings separate offensive PPP and defensive PPP, then they adjust for the opponent and location. Finding adjusted efficiency for a given game is simple.
Adj. OE = PPPo / (Opponent’s Adj. DE / Average PPP)
Adj. DE = PPPd / (Opponent’s Adj. OE / Average PPP)
You average up each game to get their adjusted offense and defense. The difference between the two is the overall adjusted efficiency which you will find on kenpom. On Barttorviik.com (T-Rank), Torviik uses Bill James Pythagorean expectation formula to calculate barthag, which is a team’s expected winning percentage against an average team. The formula does get a little bit complex with some added factors for accuracy. While kenpom is a bit more secretive about what exactly goes on we do know this about T-Rank.
– There is a recency bias, games in the last 40 days count for 100%, then remove 1% for each day until a game is 80 days old, 80+ day’s old game’s count for 60%.
– An adjustment that discounts blowouts in mismatches. Remember when I said the difference is between the top and bottom of college basketball is massive? This helps account for it. If the margin of victory (MOV) is more than 10 points and the difference in Barthags is above a threshold, the game starts getting discounted. If the MOV is 20 points or higher, the discount is (Higher Barthag – Lower Barthag – .5) * 2. So if a team with a Barthag of .8000 is playing a team with a Barthag of .2000, and it wins by 20 points, the game value will be 1 – (.8 – .2 -.5) * 2, or 80%.
– Both Kenpom and T-Rank include a preseason aspect for the first half of the season, this is to not have skewed results early in the season due to sample size.
All of the following information along with a couple of other minor things create the rankings you can find on kenpom.com and barttorviik.com, these two are considered the cream of the crop when it comes to metrics. There are some other notable metrics, Sargin, BPI, KPI. But if you want the most accurate and easy to use metrics, these two are what you want to work with.
How To Create Score Predictions With Kenpom
Say 2020 March Madness actually happened, and Baylor vs Gonzaga was the national championship, if you were to make a prediction on the score what would the best data-backed answer be? Using kenpom’s adjusted offense, defense, and tempo we can make a realistic guess. For starters let’s look at the profiles of the two teams.
Baylor
Adj Off: 113.5
Adj Def: 88.1
Adj Tempo: 66.2
Gonzaga
Adj Off: 121.3
Adj Def: 94.4
Adj Tempo: 71.9
The basis of this is the deviation of a team’s rating from the average. Against an average team, one would expect Baylor to score 113.5 points per 100 possessions, but Gonzaga’s defense is better than average, 5.6 points better than average. The average team would score just 94.4 points per 100 possessions. To find out how many points Baylor should score against Gonzaga you find the deviation. 13.5 subtracted by 5.6 is 7.9. So on average Baylor should score 107.9 points per 100 possessions against Gonzaga.
The same can be used to find how many points Gonzaga would put up on Baylor. Gonzaga’s offense was a mind-boggling 21.3 points above average, easily the best in the country. Baylor’s defense was elite as well, second in the nation good at 11.9 points above average. Just like we did earlier we find the deviation. 21.3 subtracted by 11.9 is 9.4, so Gonzaga should score 109.4 points per 100 possessions against Baylor.
The efficiency numbers for each team are great, but basketball is not played in 100 possessions. Predicting how many possessions in a game is tricky, but the best way is to just average the two adjusted tempi of the team’s which would get you 69.1 possessions in the game, so 69 possessions in a game between Gonzaga and Baylor. Sometimes the better team will control the pace and make the opposition play as they do, but with two evenly matched teams, the average tempo is fine enough.
107.9/100 (dot) 69=74.45
109.4/100 (dot) 69= 75.49
So a Baylor vs Gonzaga matchup on average would be 75.49 to 74.45, extremely close. Now, college basketball is weird so the game could end up being something wacky like 65-50, but the above score is a good indication of what the average results would be if teams played 10 times.
[B] Using Kenpom to Predict Upsets In Your March Madness Bracket
March Madness brackets are the most popular aspect of college basketball. Along with making a bracket picking upsets is a common practice. Every March you will hear analysts talk about teams with “red flags” and have a chance of being upset. You’ll hear the same analysts talk about lower seeds that have “everything going for them” in terms of upsetting another team. Now, most of the time they are just talking out of their ass, but there are trends you can look for. If you see these trends proceed with cau#
2 vs #15
Two seeds with an AdjO or AdjD outside of the top 50
Picking a 15 over two seeded is bold and ill-advised. But if you really want to pick one, or are just looking at teams to not put in your final four this is one of them. Since 2010, seven two seeds have fit these criteria, they are just 3-3. All other two seeds are 33-1, with the lone exception being MTSU’s upset over Michigan State in 2016. The three two seeds that did manage to win? Two of them (2010 Villanova and 2011 Notre Dame) fell in the second round. Only Michigan, who reached the elite eight in 2014 managed to avoid this curse. Even two seeds near the bottom part of the top 50 are liable. 2017 Duke and 2018 Cincinnati were 47th and 49th respectively in offense or defense fell early. The Blue Devils got hammered by South Carolina while the Bearcats blew a 20 point lead, both in the second round.
If the 2020 tournament happened, I would have looked out for Creighton as a possible first-weekend upset. The Bluejays were a projected two seed but had the 78th ranked defense on kenpom.
#3 vs #14
Not really a trend here, maybe look at 14 seeds that are top 100 in Kenpom
Like the two 15 seed upsets, it’s not worth picking a three vs 14 seed upset as the reward is not nearly worth the risk. But teams in the top 100 have a 15.7 winning percentage against three seeds since 2010, which is terrible but teams outside of the top 100 have an even worse 8.0 winning percentage. Overall just stay away from this upset.
#4 vs #13
If the four seed has an AdjO or AdjD outside the top 40 and the 13 seed has one of the two inside the top 50 there’s gonna be an upset.
This is where picking an upset is actually a decent pick. In the last 12 years when the mentioned combination occurs four seeds are 1-6, when it doesn’t they are 38-3. The only outliers are Morehead State over Louisville, Marshall over Wichita State, and UC Irvine over Kansas State. Even then all three-four seeds had either an offense outside the top 40 or a defense outside the top 40.
In 2020 two projected four seeds had a defense outside the top 40, Kentucky and Oregon. Of the projected 13 seeds, only North Texas had an offense or defense in the top 50. If we saw a matchup between Kentucky/Oregon against North Texas there is a good chance we’d have seen an upset pulled.
#5 vs #12
5 seeds that fall outside of the top 60 in either offense or defense. Also, pay attention to 12 seeds in the top 60 overall.
The five vs 12 upset is the most common upset picked, as this is where the top teams in a one-bid league usually fall in the seeding line. Since the tournament expanded in 1985, there have only been five times a 12 seed failed to upset a five. In that same time frame, five seeds are 50-90 against 12 seeds, much higher than the seeding differential would make you think.
Five seeds that fall under that 60 cut line are a miserable 3-12 in the past eleven years. Top five seeds that aren’t are 22-7. 12 seeds in the top 60 are 13-11, the rest are 6-14.
Ideally, you would look at upsets that fit under both categories, since 2009 six upsets have fit under this umbrella. Weak 12 seeds, as well as weak five seeds, have five upsets since 2009, strong five and 12 seeds have four upsets since 2009. The only 12-5 upset to not fit under any categories was Little Rock over Purdue in 2016.
In 2020, two projected 12 seeds were in the top 60, Yale at 58 and Cincinnati at 44. Both Liberty and Stephen F Austin were below 60. Of the projected five seeds none were outside of the top 60 in either defense and offense but two were really close. BYU sat at 60th in defensive efficiency, while Auburn was at 55th.
#6 vs #11
Six vs eleven is extremely complicated because oftentimes the two teams are close in kenpom. With that being said these matchups can be divided into three groups.
True Upsets: Where the teams have at least a 25 spot gap in kenpom rankings
Non-Upsets: Where the 11 seed is higher than a six seed
Neither: self-explanatory
These true upset games are 11-10 in favor of the six seed in the last eleven years, so you should generally pick one true upset. West Virginia vs any of the projected eleven seeds would be a true upset, as would Michigan against NC St and ETSU.
Non-upsets are 9-5 in favor of the 11 seed, there were zero possible non-upsets that could have happened in 2020. As for games that fall under neither, just use your own discretion as there is not a trend either way.
Advanced metrics are an integral part of college basketball. Whether it be to calculate resumes or make predictions there are a plethora of metrics to choose from. I just scratched the surface of what you could do with them. Visiting warrennolan.com or barttorviik.com and just messing around is a good way to lose an hour. Whether you want to have a better bracket or just better understand the college basketball landscape, these metrics are a great way to do it.
On The Hardwood 