GT Hoops General Topics

[MtnWasp]

If you have any evidence of the NET being driven by initial rankings I would like to see it.

While the NET is fairly opaque, it uses mostly the same criteria as a typical basketball predictive model. It has a couple of political tweaks (caps on margin of victory, for example) that make it less accurate than one without political tweaks (like kenpom), but they are small factors in the results.

NET Rankings:

• Purpose:
  The NET rankings, primarily used for sorting Division I men's basketball teams, serve as a primary sorting tool for the NCAA tournament selection process.
• Calculation:
  The NET rankings consider a team's overall record, quality of wins, and strength of schedule.
• Factors:
  
  • Opponent Strength: The NET system evaluates the strength of a team's opponents, giving more weight to wins against stronger teams.
  • Home/Away Advantage: The NET system adjusts for the home-court advantage, as teams tend to perform better at home.

"
So then, how are the initial values for strength of schedule and "quality" of wins derived and how do they effect the final rankings?

 

[Root4GT]

NET Rankings:

• Purpose:
  The NET rankings, primarily used for sorting Division I men's basketball teams, serve as a primary sorting tool for the NCAA tournament selection process.
• Calculation:
  The NET rankings consider a team's overall record, quality of wins, and strength of schedule.
• Factors:
  
  • Opponent Strength: The NET system evaluates the strength of a team's opponents, giving more weight to wins against stronger teams.
  • Home/Away Advantage: The NET system adjusts for the home-court advantage, as teams tend to perform better at home.

"
So then, how are the initial values for strength of schedule and "quality" of wins derived and how do they effect the final rankings?

The first NET ranking comes out in early December. Teams typically have 6-8 games played before the first NET ranking comes out.

 

[stinger78]

If you have any evidence of the NET being driven by initial rankings I would like to see it.

While the NET is fairly opaque, it uses mostly the same criteria as a typical basketball predictive model. It has a couple of political tweaks (caps on margin of victory, for example) that make it less accurate than one without political tweaks (like kenpom), but they are small factors in the results.

NET uses the RPI SOS numbers, which used early rankings. Not sure if that includes weighted preseason SOS. That is easy to adjust for, though, and I am confident they do.

 

[slugboy]

NET Rankings:

• Purpose:
  The NET rankings, primarily used for sorting Division I men's basketball teams, serve as a primary sorting tool for the NCAA tournament selection process.
• Calculation:
  The NET rankings consider a team's overall record, quality of wins, and strength of schedule.
• Factors:
  
  • Opponent Strength: The NET system evaluates the strength of a team's opponents, giving more weight to wins against stronger teams.
  • Home/Away Advantage: The NET system adjusts for the home-court advantage, as teams tend to perform better at home.

"
So then, how are the initial values for strength of schedule and "quality" of wins derived and how do they effect the final rankings?

According to the NCAA, NET ratings to not use any preseason ranking or data prior to the season. That is why they wait several weeks to publish anything—the model is still figuring out the strength of schedule and other factors

How do NET rankings work in NCAA tournament selection? - NCAA.org

The NCAA Evaluation Tool — the NET — rankings help determine which teams make the NCAA Division I Men's and Women's Basketball Championships by evaluating game results,

www.ncaa.org

The NET is designed to be optimal in March, not in early December. It is first released in December each year due to demand from the NCAA membership, media and fans. There are going to be more outliers early in the season because of the limited data available, more so than what you might see in other metrics that use preseason data as a starting point for those rankings. The NET does not use any data from prior seasons or any preseason rankings or inputs.

 

[ESPNjacket]

NET uses the RPI SOS numbers, which used early rankings. Not sure if that includes weighted preseason SOS. That is easy to adjust for, though, and I am confident they do.

If the NET uses RPI SOS numbers it should be immediately abolished.

 

[stinger78]

If the NET uses RPI SOS numbers it should be immediately abolished.

It’s a formula *derived* from the RPI, not the RPI SOS itself. My bad. Still, the two are related.

NCAA College Basketball RPI Rankings (updated today)

NCAA college basketball RPI rankings and ratings for 2026, updated daily for all teams and available historically for every day of the 2026 season.

www.teamrankings.com

 

[ESPNjacket]

It’s a formula *derived* from the RPI, not the RPI SOS itself. My bad. Still, the two are related.

NCAA College Basketball RPI Rankings (updated today)

NCAA college basketball RPI rankings and ratings for 2026, updated daily for all teams and available historically for every day of the 2026 season.

www.teamrankings.com

That's not what it says. It says it displays a SOS derived from the old RPI on the team sheets, which are distributed to the selection committee.

The SOS in most predictive models is calculated from offensive and defensive efficiency numbers. That is quite different than RPI's very simple win/loss only method.

 

[slugboy]

That's not what it says. It says it displays a SOS derived from the old RPI on the team sheets, which are distributed to the selection committee.

The SOS in most predictive models is calculated from offensive and defensive efficiency numbers. That is quite different than RPI's very simple win/loss only method.

I didn’t read about how NET calculates SOS, but it uses offensive and defensive efficiency, and caps the impact of blowouts

The complaint I’ve seen is that November games count as much as February games. That’s a matter of taste

 

[spdrama]

Data Scientist explains NET in blog, concluding “Overall, I think the NET rankings are good for the sport of basketball and the tournament. They can help reduce bias in selecting and seeding teams during March Madness and improve the quality of the sport during the regular season”. The blog in its entirety is linked here

College Basketball's NET Rankings Explained | Towards Data Science

How data science influences March Madness

towardsdatascience.com

 

[Peacone36]

I do not agree. They ballooned in the SOS *after* they began to play each other post-NYD. Classic positive feedback loop.

Not their fault the ACC sucks.

Keep naming coaches in waiting and hiring coaches because old good coaches liked them. Working out real well.

 

[Peacone36]

Kenpom starts the season with initial rankings based on how hard the program makes Ken and it impacts the entire season.

Haslammetrics starts from zero and doesn’t normalize until late December.

 

[slugboy]

Kenpom starts the season with initial rankings based on how hard the program makes Ken and it impacts the entire season.

Haslammetrics starts from zero and doesn’t normalize until late December.

NET is more like Haslametrics, since there’s no initial data, and it also takes a while to shake out

 

[ESPNjacket]

NET is more like Haslametrics, since there’s no initial data, and it also takes a while to shake out

Ahh. Making stuff up again. It seems to be a bad habit.

From Ken Haslam's website:

"Prior to 2020-21, the algorithm I utilize knew nothing about each team's history and, therefore, treated all teams as absolute equals on Day 1 of the season. That resulted in some unfamiliar names appearing near the top of the rankings over the first month or two of the season as the algorithm continued to build a larger and larger sample set. Starting with the 2020-21 season, I began to implement preseason baselines, which were Day 1 ratings estimates based on team prestige, last season's ratings, returning players, transfers, recruiting scores, and coaching changes. This allowed me the ability to provide educated estimates of ratings and projections on the very first day of the season. Regardless of the above strategies, be it with preseason baselines or without them, the ratings should become increasingly accurate over time as more and more actual results roll in from the present season each year."

 

[Northeast Stinger]

Ahh. Making stuff up again. It seems to be a bad habit.

From Ken Haslam's website:

"Prior to 2020-21, the algorithm I utilize knew nothing about each team's history and, therefore, treated all teams as absolute equals on Day 1 of the season. That resulted in some unfamiliar names appearing near the top of the rankings over the first month or two of the season as the algorithm continued to build a larger and larger sample set. Starting with the 2020-21 season, I began to implement preseason baselines, which were Day 1 ratings estimates based on team prestige, last season's ratings, returning players, transfers, recruiting scores, and coaching changes. This allowed me the ability to provide educated estimates of ratings and projections on the very first day of the season. Regardless of the above strategies, be it with preseason baselines or without them, the ratings should become increasingly accurate over time as more and more actual results roll in from the present season each year."

The age old conundrum. If the formula spits out data that doesn’t make sense to me I change the formula.

 

[stinger78]

So what’s the upshot of all this? Do we love NET SOS now?
If indeed it does “adjust” as the season progresses, that’s a good thing, and I would expect that. However, I have seen nothing that tells me there is a feature that moderates the positive feedback loop of constant in-conference play.

 

[Connell62]

Ahh. Making stuff up again. It seems to be a bad habit.

From Ken Haslam's website:

"Prior to 2020-21, the algorithm I utilize knew nothing about each team's history and, therefore, treated all teams as absolute equals on Day 1 of the season. That resulted in some unfamiliar names appearing near the top of the rankings over the first month or two of the season as the algorithm continued to build a larger and larger sample set. Starting with the 2020-21 season, I began to implement preseason baselines, which were Day 1 ratings estimates based on team prestige, last season's ratings, returning players, transfers, recruiting scores, and coaching changes. This allowed me the ability to provide educated estimates of ratings and projections on the very first day of the season. Regardless of the above strategies, be it with preseason baselines or without them, the ratings should become increasingly accurate over time as more and more actual results roll in from the present season each year."

All this **** makes my head hurt, but that is why gamblers follow Haslam closely. In essence, he is building a set of power ratings at the beginning of the year just like any real handicapper.

That said, it has become increasingly more difficult to be accurate with all of the movement in the transfer portal.

 

[slugboy]

Ahh. Making stuff up again. It seems to be a bad habit.

From Ken Haslam's website:

"Prior to 2020-21, the algorithm I utilize knew nothing about each team's history and, therefore, treated all teams as absolute equals on Day 1 of the season. That resulted in some unfamiliar names appearing near the top of the rankings over the first month or two of the season as the algorithm continued to build a larger and larger sample set. Starting with the 2020-21 season, I began to implement preseason baselines, which were Day 1 ratings estimates based on team prestige, last season's ratings, returning players, transfers, recruiting scores, and coaching changes. This allowed me the ability to provide educated estimates of ratings and projections on the very first day of the season. Regardless of the above strategies, be it with preseason baselines or without them, the ratings should become increasingly accurate over time as more and more actual results roll in from the present season each year."

I was responding to @Peacone36 ’s comment about Haslametrics, and took on faith that what he said about Haslametrics starting from zero was correct.

The NCAA does say that NET does not use data from before the season.

 

[ESPNjacket]

All this **** makes my head hurt, but that is why gamblers follow Haslam closely. In essence, he is building a set of power ratings at the beginning of the year just like any real handicapper.

That said, it has become increasingly more difficult to be accurate with all of the movement in the transfer portal.

Here is all you need to know to be ahead of 95% of the college basketball fans:

- Kenpom's model is very similar to what Vegas books use to set lines. (Dean Oliver's approach that is well documented for anyone who cares to learn about it.)

- Ken Haslam took a different approach (referred to in the quote above) that is also statistically valid. It is a good alternative for someone looking to wager that is looking for an edge over the house. The results end up about the same most of the time though.

-NET is the same approach as kenpom/Vegas books, with some of the predictive variables dumbed down by politics. It is also more opaque so you don't really see the important part, which is the strength of team rating that drives predictions.

I like kenpom for my use because he exposes the most important part, the strength of the team that he shows in a useful format: the Netrtg is the efficiency margin (off/def) adjusted to the strength of each opponents' def/off. As I mentioned earlier, this type of model is how lines get set. They are then adjusted for home court (about 3.5 points) and other non-mathematical factors (injuries/illness/intangibles).

The reason I bash the quad concept is that it takes away much of the value of statistical models. As expected, the distributions of the teams follows a bell curve so the difference between the top 10 teams is greater than the difference between the next 15, yet a 25 team quad treats teams 1 and 24 as equals. Why have a statistical model at all if you aren't going to use the most valuable part?

If it isn't obvious, I have a lot of practical experience with predictive models in real world applications. I am happy to discuss this stuff with people who are willing to deal with facts about statistical modeling and what we know about specific models. There is a lot more to it than this format will allow but I am fairly experienced at explaining these types of concepts in terms that are easy to understand.

 

[Peacone36]

So what’s the upshot of all this? Do we love NET SOS now?
If indeed it does “adjust” as the season progresses, that’s a good thing, and I would expect that. However, I have seen nothing that tells me there is a feature that moderates the positive feedback loop of constant in-conference play.

Nah, you don’t have to love the picture on the puzzle in order to appreciate the pieces that make it up. Do I love the NET? No. Do I think SOS has a part to play in determining the field? Absolutely. Are all metrics inherently flawed? Absolutely. Kansas/Gonzaga/UNC this season. I wanna say Michigan State the prior year, all Kenpom darlings. That’s why I have always been an eye test advocate (as a piece). I also really love Kenpom, I use it all the time, coupled with other metrics (Paper Tiger, SOS, and yes, even quads).

 

[Peacone36]

I was responding to @Peacone36 ’s comment about Haslametrics, and took on faith that what he said about Haslametrics starting from zero was correct.

The NCAA does say that NET does not use data from before the season.

I do weekly radio spots with Erik during CBB season and he has always claimed it to be the case. I’m inclined to take him at his word.