NIT Bracketology: Monday, February 26th

Our college basketball model is updated through this weekend’s games. Here are its latest probabilities. Here’s its latest NIT Bracketology. Here’s where the NIT picture stands:

Last Four In, First Four Out, Etc.

On the low side:

• Last Four In (from the cut line going upwards): UNLV, Yale, St. Bonaventure, Minnesota
• First Four Out (from the cut line going downwards): Florida State, Loyola, UCF, George Mason
• Others in the mix (>10% NIT chance, below NIT field, not in our field or in the First Four Out): LSU, Georgia, UMass, Boston College, Rutgers, Miami

On the high side:

• Last Four In (from the cut line going downwards): Providence, Seton Hall, Utah, Texas A&M
• First Four Out (from the cut line going upwards): Colorado, Virginia, Villanova, Texas

Conference Tournament Favorites

We aren’t projecting any conference tournament favorites to land in the NIT right now, with our NCAAT cut line at 45^th, our NIT cut line at 82^nd and the right combination of automatic bids and at-large bids in between those numbers to make that work. In the event any of these teams below were to lose their conference tournament, though, we’d have them in. Even McNeese and Samford are each more than 90% likely to make the NIT in simulations in which they don’t receive an NCAAT automatic bid.

• Indiana State
• Princeton
• James Madison
• Grand Canyon
• McNeese
• Samford

Where We’re an Outlier

On Friday, we compared our brackets to Bracket Matrix’s. Today, we’re looking at our probabilities against Bart Torvik’s TourneyCast™, a well-respected model doing the same thing we’re trying to do—put numbers on teams’ tournament chances. Five notable discrepancies, measured by NCAA Tournament probability:

Mississippi: Torvik 2%; TBC 24%

Our models agree that Ole Miss is unlikely to make the NCAA Tournament. They disagree on how unlikely. I don’t know exactly how Torvik’s model works, but I know Mississippi isn’t getting a whole lot of adjustments in ours in the way of the little things: They don’t have a bad Q1 win percentage, they’re unlikely to reach five Q1 wins, and despite lacking top-tier opponents their nonconference schedule doesn’t come with an unsightly number. If you amp up the weights given to Wins Above Bubble and Résumé on Torvik’s résumé comparison tool, you see a few more historic teams like Mississippi that did make the field. Given our predictions for Mississippi’s game results aren’t very different, my guess is that this is a disagreement over how indicative KPI and SOR are of tournament inclusion. It’s possibly a big disagreement, which makes it one to monitor going forward.

Texas A&M: Torvik 10%; TBC 39%

Another SEC team, another team we’re higher on. Texas A&M already has five Q1 wins, and we see them as rather likely to end with six or more, something which historically gets teams into the tournament. Again, my guess is that Torvik’s model is looking at teams’ accomplishments differently than we are, and possibly weighting that side of the coin a little less, proportionally to how much it’s weighting how good teams are.

Pitt: Torvik 33%; TBC 10%

This time, we’re on the low side. One interesting thing with Pitt is that their kenpom ranking is 54^th but Torvik’s power rating for them sits at 45^th, which isn’t a massive difference but falls on a high-stakes portion of the curve, with 45^th our median cut line. T-Rank itself, Torvik’s rating formula, isn’t as high on the Panthers as his power rating is, and I would assume T-Rank is what’s being used to simulate remaining games. So, while this again points to a difference in weighting between what teams have done/how good teams are, it also potentially points to a difference in what tools we think the committee uses to gauge a team’s ability.

St. John’s: Torvik 65%; TBC 20%

I’ll want to look at this one again tomorrow, given what we’re about to say regarding our model not reacting quickly enough to significant results. St. John’s wasn’t a very big underdog yesterday, if they were an underdog at all. Our model relies on kenpom, which called the Red Storm 40% likely to win. Still, it’s possible KPI and SOR reacted more dramatically to the result than we expected, and that we’ll see St. John’s jump in tomorrow’s probabilities. Until we’re sure we disagree, we aren’t going to dig much further into the disagreement.

Gonzaga: Torvik 94%; TBC 71%

We’re much more bearish than Torvik on the Zags. We still think they’ll get in, but we aren’t as confident as he is, and that’s despite having a comparable view of their WCC Tournament chances. Again, we might be underestimating how much the committee cares about a team’s ability, or we might be failing to account for what thresholds of NET and/or kenpom get teams into the tournament regardless of résumé.

Model Talk

We’ve talked before about the uncertainty which accompanies our KPI and SOR proxies. They’re our top concern, because in addition to an inaccuracy, they can produce an inconsistency: Our proxies handle the result the day after it happens, and then it gets handed off to the updated real rankings on the second day (the rankings update after we run the simulations every morning). So, if a noteworthy result happens on Saturday, our model takes a stab at gauging its impact on Sunday, but more movement can occur into Monday. The effect of a surprising result can ripple for two days.

We got a little bit of a stress test these last two days from Air Force upsetting New Mexico on Saturday. We’re getting another one with Ohio State’s upset yesterday of Michigan State.

Our model didn’t react dramatically enough to New Mexico’s loss. Yesterday morning, it still had the Lobos 92.8% likely to make the NCAA Tournament. Today, that number’s down to 81.6%. Some of this might come from a good day for other bubble teams yesterday—St. John’s got a huge win over Creighton, Nebraska solidified its case, and Ohio State and Maryland each improved their chances—but a lot is that our model underestimated the impact of that loss on New Mexico’s tournament chances. We’d like our model to err on the side of underreacting (we’d rather walk things forward than walk things back), but we don’t want teams’ relevant probabilities moving more than ten percentage points on days they don’t play.

With the Michigan State loss, our model’s prediction of the KPI change was nearly spot on, which points us towards SOR as the problem. Sure enough, our model underestimated Michigan State’s SOR change as well. So, it’s SOR where we’re increasing the magnitude of the changes, though we’re still being cautious: Our new SOR proxy still slightly underrates the impact of the Michigan State loss on Michigan State’s SOR.

We’re not going to have perfect proxies this year. We’re using elo-adjacent systems to stand in for non-elo systems. It’s going to be imperfect. Hopefully, we can iron the error down to a negligible amount. Plenty more to come.

Programming Notes

We’ll be updating our probabilities page almost every day, but we aren’t planning to update our bracketology again until Friday, after every team’s weeknight game is completed. We’ll get to daily NIT Bracketology soon (the current thought is to start that on Sunday), but don’t expect updated brackets tomorrow, Wednesday, Thursday, or Saturday.

For the Archives

Where the probabilities stand today (for archival purposes):