The Models

Minnow

Predictive model

Minnow operates much the same way that advanced predictive models, such as KenPom and Torvik, do. It calculates a team's offensive and defensive efficiency (points scored and allowed per 100 possessions) and its tempo (possessions per game), then adjusts all three for the quality of the opponents, giving us adjusted offensive efficiency, adjusted defensive efficiency, and adjusted tempo.

Subtracting a team's adjusted defensive efficiency from its adjusted offensive efficiency gives us its net rating, which can be understood as "how many points per 100 possessions is this team better than the average D-I team?" That relationship is linear, so if Michigan's net rating is 39 and Ohio State's is 25, Michigan is expected to be 14 points per 100 possessions better than Ohio State.

To turn that into a predicted margin, all we need is the number of possessions in the game, which is where adjusted tempo comes in. Then I convert the predicted margin into a win probability by assuming a standard deviation of 10 points and calculating the z-score of that margin. For the non-statisticians among us, that translates to "how likely is a team favored by this many points to win this matchup?"

Minnow doesn't use any preseason priors, so it needs a few weeks of games before its ratings mean much. That's why the weekly rankings start the Monday after Feast Week.

Resumetric

Resume model

Resumetric is a "true" resume metric. It evaluates teams based solely on who they've beaten, where they beat them, and who those teams have beaten. It doesn't account for margin of victory, points scored, or any other stats.

The first thing I thought of when I set out to build it was an Elo system. The problem with Elo when you only have one season's worth of data is that every team starts from the same baseline, so it would value Arizona's season-opening win over Florida just as much as Eastern Michigan's season-opening win over Georgia State. What I wanted was something that works sort of like Elo but can go back and give teams the proper amount of credit for their wins in the context of the entire season. For the math nerds, that's a Bradley-Terry model, with a parameter for home, neutral, and away games.

This is what separates it from other resume metrics like strength of record or wins above bubble. Those still judge the quality of each win using advanced efficiency metrics, which typically account for margin of victory. Not this one.

Resumetric also tends to be the most confident of my models in the favorites. That can win it a lot of Expected Tourney Points, and cost it in Log Loss.

Branchy Brackets

Branchy Tree model

Branchy Brackets is a tree algorithm trained on a lot of factors, including each team's season-long tempo-free stats (provided by Torvik) and the adjusted efficiency metrics produced by Minnow, in the hopes of discovering relationships between specific data points to arrive at advanced insights.

What separates a tree algorithm from a linear model is that it can learn how variables interact. For example, it might see that team A is a good three-point shooting team and team B gives up a lot of threes, and give team A a boost for that combination. A linear model can't learn that on its own.

The downside is that it can be hard to tell why Branchy thinks what it thinks. Ever the weirdo of our model family, it tends to like favorites, and every now and then it takes a big swing that either pays off spectacularly or really, REALLY, doesn't.

The Composite

Hybrid

The Composite is just the average of the odds from my other three models (Minnow, Resumetric, and Branchy) for a given matchup. Its founding principle is a rather unscientific "let's mash the models together," and it's there to see if there's any benefit to a hybrid approach.

Torvik

T-Rank Control

Torvik is a well-respected advanced analytics model built by Bart Torvik and available for free at barttorvik.com. It's one of the two controls in the MLBR, and it's the measuring stick for whether my models can hang with the best of them. His site is also the source of the game data that every one of my models runs on.

To score Torvik fairly, I use his ratings as they stood on Selection Sunday, from his archive (the T-Rank Time Machine, which goes back to the 2009–10 season). His ratings include a Barthag: a team's chance of beating an average D-I team on a neutral court. Combining two teams' Barthags with a formula called log5 gives Torvik's odds for any matchup between them, even ones that never happened.

Chalk

Control

Chalk always picks the higher seed. It's typically a very strong bracket strategy, so it's a good baseline to tell whether my models are just outsmarting themselves. For evaluation purposes, Chalk picks the higher seed with 99.99% certainty, which is a good way to score well in Accuracy and a terrible way to score well in Log Loss.

MNPI

Retired

My very first model, retired after the 2025 tournament. It was a linear regression on teams' raw per-game stats and their simple rating system (SRS) score from Basketball Reference. It didn't account for tempo, margin of victory, or home court, and its data source started charging for access, so I had to part ways with it. You'll still see it in the 2025 MLBR post.

Evaluation Metrics

Accuracy

For what percentage of tournament games did the model's favorite win? It's the simplest way to evaluate a model and a good way to get a basic feel for its performance, but there's a reason it's not the gold standard. It only cares who the model picked, not how confident it was.

Log Loss

The industry-standard metric for probabilistic predictions, and the one I care about most. Log loss rewards not only correctness but also accurate probabilities. It's a penalty function, so lower is better.

Say a model gives the favorite a 75% chance in two games. If the favorite wins, the model takes a penalty of 0.29. If the underdog wins, the penalty is 1.39. Now say the model was 99% sure instead: it takes a tiny penalty of 0.01 if it's right, but a whopping 4.61 if it's wrong. That's more than three times the penalty for being only 24% more confident. In other words, it severely penalizes confidently wrong answers, encouraging models to not only predict correctly, but also predict with the right amount of confidence.

As a point of reference, a model that flips a coin for every game would have a log loss of about 0.693, so anything below that means we're doing better than randomly guessing.

Tourney Points

Tournament Points

If the models all filled out a bracket before the tournament started, how many points would they earn using the traditional bracket scoring system? That's 10 points for a correct pick in the first round, 20 in the second round, 40 in the Sweet 16, 80 in the Elite Eight, and so on, up to 320 for the champion. Essentially, if all of the models were in a single bracket pool, who would win?

It's the least scientific of the metrics and mostly just for fun. Because the late rounds are worth so much, it swings wildly on whether a model gets the champion and the Final Four right.

Expected Tourney Points

Expected Tournament Points

Exactly like Tourney Points, but with a twist. Instead of awarding points for a correct pick, we multiply the points by the chance the model gave that team of getting that far before the tournament started. If the model gave Team A a 70% chance of winning its first-round game and a 42% chance of reaching the Sweet 16, it earns 10 × 0.70 = 7 points when Team A wins in the first round and 20 × 0.42 = 8.4 points when it wins in the second.

Think of it as the average number of bracket points a simulation generated by this model would earn. Models that are confident in favorites tend to rack these up, since favorites win more often than not.

Site Terms

MLBR

Machine Learning Battle Royale

The Machine Learning Battle Royale: every year, my models make predictions about every possible matchup in the NCAA Tournament using only data from before Selection Sunday, and I grade them, along with Torvik and Chalk, using the four evaluation metrics.

The Battle Royale standings run the same evaluation on every tournament since 2008, and I usually include a writeup each year as part of my tournament post-mortem.

PGWE

Postgame Win Expectancy

How often a team would have won a game it already played if everyone on the floor had shot what we'd expect from three and the free-throw line.

Those are the two shots defenses have the least control over. Nobody guards the free-throw line, and while defenses have a lot of say in how many threes a team takes, they have very little say in how many go in. That leaves a lot of room for variance: a good shooter can go 1-for-8 one night and 6-for-8 the next, no matter who's guarding them.

None of that means a hot shooting night doesn't count. Making shots is the whole point, and the team that made them earned the win. But a team that wins because its shooters made far more than usual is leaning on something that's hard to sustain, and that we wouldn't expect to repeat if the two teams played again.

To measure it, I estimate what each player should be expected to shoot from three and the line, based on their track record, how much they shoot and play, home court, and, to a small extent, the opposing defense. Then, I replay the game 10,000 times with everyone shooting at those expected rates, keeping everything else the same. The share of those replays a team wins is its PGWE.

A team that won with a 20% PGWE needed its shooters to have a big night (or its opponent's shooters to have a really bad one), and got it. A team that lost with an 80% PGWE went cold at the worst time, or ran into a buzzsaw. Shots at the rim and in the midrange are left as they happened, since defenses have a much more reliable effect on how often those go in. You can find every game, broken down player by player, on the PGWE page.

Luck

The Luck column in the Rankings is how many more games a team has won than its PGWE says it should have. A team that's 10–2 with 8.5 expected wins has a Luck of +1.5. It's recalculated each week as players' expected shooting updates, so earlier weeks can shift slightly.

Net Rating

Net

Minnow's measure of team strength: how many points per 100 possessions a team is better than the average D-I team, on a neutral court. It's adjusted offense minus adjusted defense.

Adj O / Adj D

Adjusted efficiency

Points scored (Adj O) and allowed (Adj D) per 100 possessions against an average D-I team, adjusted for the quality of the opponents and where the games were played. Higher is better for Adj O; lower is better for Adj D.

Adj Tempo

Adjusted tempo

How many possessions a team would be expected to have in a game against an average D-I team. It's adjusted for the pace of the opponents a team has played, so it's not the same as its raw average tempo: a fast team that's played a lot of slow, grind-it-out opponents will have a slower raw tempo than it really prefers.

To get the expected tempo for a game between two teams, I take how much faster (or slower) than average each team plays and add both of those to the average:

Expected tempo = team 1 tempo + team 2 tempo − average D-I tempo

If Michigan's adjusted tempo is 72 and Alabama's is 74, you wouldn't expect that game to be slower than 74, since that would mean Alabama would have fewer possessions against Michigan (a team that also likes to play fast) than it would against an average team. That's why it isn't just the average of the two.

Feast Week

Thanksgiving week, packed with early-season tournaments. Since my models don't use preseason priors, the first weekly rankings, the Matchup Explorer, and bracketology all go live the Monday after it, once there are enough games to go on. From then on, everything updates every Monday.