overview

Last year, we debuted our first basketball model under the name ScoreSight. While other models use power ratings, our model utilized a different approach. We looked at underlying team data to project in-game team stats, creating lines and projections for games that did not follow the transitive property that power ratings models do. For example, on a neutral court, Team A would be favored vs Team B. Team B would be favored vs Team C. In power ratings models, Team A would also be favored over team C, but in our example, Team C could be favored over Team A, depending on how they match up against each other. In the previous month or so, we’ve revisited this model and fine-tuned it, filtering out noise and re-balancing weights. We’re confident that our new model is highly-calibrated.

Alongside each game projection is a series of game ratings, Quality, Importance, Excitement, and Overall. Quality measures how good two teams are, using our hypothetical round-robin results (where each team is matched up against each other team). Importance measures how the game’s results will impact NCAA Tournament selection status, average seed, and conference final standings. Excitement simply measures how close a game is to being 50/50. The Overall rating is a combination of Quality and Importance.

We use these game projections to simulate the remaining regular-season 20,000 times to create the best estimates possible, using Massey Ratings‘ data collection to gather both completed games and scheduled games. At the end of each simulation, we seed and simulate conference tournaments before creating an estimate of the NCAA selection committee’s 68-team “S-curve.” We do this with a special formula, based on the behavior of the selection committee since 2018, when NET was adopted. This formula considers a team’s record, schedule, margin of victory, relative power, and quadrant-based performance. We add in some randomness to account for changes in how the committee interprets data each year, and we think that this makes the simulated selection process a little more accurate. Specifically, after removing automatic qualifiers and the next top 25 teams, any teams with at least 90% of the resume score of the 37th team could be selected. We use a weighted-draw, heavily biased towards the best-resume teams. This way, teams with an outside chance of making the dance could be selected, just like on actual Selection Sunday.

Based on this seeded list of teams, we set the 68-team bracket and simulate the tournament – from the First Four in Dayton, to the Final Four in Indianapolis. Each team’s probability of reaching each round of the tournament and cutting down the nets on April 6 are then calculated. Like in real-life, teams can get “hot” in our simulations, performing better than expected in consecutive games, so we account for this by giving those teams very slight bumps in their win probabilities.

In addition to tournament probabilities, for all teams with at least a 5% chance of making the dance, we’ve calculated their average projected record, the probability that they win their conference regular season, separate probabilities for receiving an automatic and at-large bid, and their average seed in the tournament.

references

Massey Ratings // NCAA

version history

2.1.5

Model Public Launch

February 20, 2026

credits

Landon McIntosh

Model Architecture and Implementation

results archive

Since Model Public Launch: 8.1 mean absolute error (MAE) over 334 games