- Joined
- May 8, 2002
- Posts
- 1,171,235
- Reaction score
- 59
You must be registered for see images attach
FILE - A basketball with a March Madness logo is seen going through a net prior to a second round of the NCAA college basketball tournament between Notre Dame and Michigan, March 23, 2025, in South Bend, Ind. (AP Photo/John Mersits, File)
Copyright 2025 The Associated Press. All rights reserved
Selection Sunday brought the release of the 68-team bracket as the NCAA tournament field was finalized. The selection committee has said it relies on a mix of predictive and resume-based metrics when choosing and seeding teams, blending measures of team strength with results on the court. That process can be viewed as an algorithm. Metrics such as efficiency ratings, strength of record and schedule strength serve as inputs that are combined to produce a final ranking. While the exact formula is not public, the outcomes are. Using machine learning, that process can be reverse engineered. By modeling the committee’s rankings using the same metrics it references, it becomes possible to quantify which factors mattered most in shaping this year’s March Madness bracket.
Methodology Behind Reverse Engineering March Madness
You must be registered for see images attach
Texas guard Tramon Mark (12), center, scores a layup as North Carolina State forward Darrion Williams (1), right, defends during the first half in a First Four college basketball game in the NCAA Tournament, Tuesday, March 17, 2026, in Dayton, Ohio. (AP Photo/Kareem Elgazzar)
Copyright 2026 The Associated Press. All rights reserved.
To understand how the NCAA selection committee seeded this year’s tournament, the first step was to recreate the inputs behind those decisions. The dataset is comprised of the committee’s full 68-team ranking along with six metrics the NCAA has publicly emphasized: ESPN’s Basketball Power Index (BPI), KenPom efficiency ratings, NET rankings, Strength of Record (SOR), Bart Torvik’s T-Rank, and Wins Above Bubble (WAB). Each of these metrics captures a slightly different view of team quality. Some are predictive, focusing on how strong a team is based on efficiency and underlying performance. Others are résumé-based, rewarding teams for what they accomplished, including wins, losses and strength of schedule. By bringing them together into a single dataset, it becomes possible to compare how each aligns with the committee’s final rankings.
The goal of the reverse engineering process was to approximate the committee’s decision-making. The selection and seeding process can be thought of as a system that takes in multiple inputs and produces a final ranking. To replicate that system, a machine learning model was trained to take the six metrics as inputs and predict the committee’s ranking for each team. In simple terms, the model learns patterns in the data. It identifies how changes in each metric are associated with changes in a team’s position in the bracket. Some metrics may consistently align closely with the committee’s rankings, while others may have less influence. By learning these relationships across all 68 teams, the model builds a data-driven approximation of how the committee weighs different factors.
You must be registered for see images attach
UMBC forward Caden Diggs (11), left, drives to the basket as UMBC guard Devin Ceaser (13), center, defends during the first half in a First Four college basketball game in the NCAA Tournament, Tuesday, March 17, 2026, in Dayton, Ohio. (AP Photo/Kareem Elgazzar)
Copyright 2026 The Associated Press. All rights reserved.
Once the model is trained, it can be used to measure the relative importance of each metric. This provides a way to quantify which inputs mattered most in shaping the final bracket. To make this easier to interpret, the metrics were also grouped into two categories: predictive measures (KenPom, T-Rank and BPI) and résumé-based measures (NET, SOR and WAB). This allows for a direct comparison between how much the committee valued underlying team strength versus on-court results.
The Metrics That Mattered Most In March Madness Seeding
The model provides a clear picture of how closely each metric aligns with the committee’s final rankings. Among the six inputs, Wins Above Bubble (WAB) stands out as the dominant factor, accounting for more than half of the model’s explanatory power.
You must be registered for see images attach
Feature importance from a machine learning model shows Wins Above Bubble (WAB) as the dominant factor in NCAA tournament seeding, with résumé-based metrics outweighing predictive efficiency ratings.
Image created by author.
At a high level, the results show that the committee’s rankings are most closely associated with résumé-based performance. WAB alone explains a majority of the variation captured by the model, with NET and Strength of Record also playing significant roles. In contrast, predictive metrics such as KenPom, T-Rank and BPI carry substantially less weight.
Implications For Your March Madness Picks
The results suggest that résumé-based performance, particularly Wins Above Bubble (WAB), plays a far larger role in seeding than predictive efficiency metrics. In practical terms, teams that accumulated strong wins against quality opponents were consistently rewarded in the bracket, even if their underlying efficiency numbers were less dominant. Metrics like NET and Strength of Record reinforce this trend, further indicating that what teams accomplished carried more weight than how strong they might project going forward.
You must be registered for see images attach
UConn guard Malachi Smith (0) shoots over St. John's forward Rubén Prey (17) during the second half of an NCAA college basketball game in the championship of the Big East tournament, Saturday, March 14, 2026, in New York. (AP Photo/Yuki Iwamura)
Copyright 2026 The Associated Press. All rights reserved.
For bracket building, this creates a clear takeaway. Teams with high efficiency ratings but weaker résumés may be seeded lower than their underlying strength suggests, making them potential value picks for advancing further than expected. Conversely, teams that rank highly in résumé metrics may be seeded more favorably, but that does not necessarily mean they are as strong in predictive terms.
That said, these insights are designed to replicate this year’s committee decisions, not predict future performance. Its high accuracy reflects how well the selected metrics explain the final rankings, but it does not guarantee those rankings are optimal indicators of tournament success. Additionally, many of these metrics overlap in what they measure, and the committee’s process still includes subjective judgment that cannot be fully captured in data.
You must be registered for see images attach
Arizona's Koa Peat (10) heads to the basket as Houston's Joseph Tugler (11) defends during the second half of an NCAA college basketball game in the championship of the Big 12 Conference tournament Saturday, March 14, 2026, in Kansas City, Mo. (AP Photo/Charlie Riedel)
Copyright 2026 The Associated Press. All rights reserved.
It is not that one metric determines seeding, but that certain metrics align more closely with how teams were evaluated. Understanding that distinction can provide an edge in identifying where the bracket reflects résumé strength versus underlying team quality.
Continue reading...
