The NCAA model improved after I split the problem into seed zones

The first NCAA attempts were too blunt. The model got better when I stopped asking one estimator to handle the whole bracket and split the problem into seed bands.

What I built

I worked on a seed prediction setup that routes teams through top, mid, and low seed bands, then assigns final seeds with an optimization step.

Why I built it

I wanted a model that reflects how bracket slots are distributed instead of forcing one global pattern to fit every team.

The build

• tiered ensembles by seed range
• routing by NET rank and bid type
• feature set across ranking, records, conference, and external team stats
• assignment step using Hungarian optimization

What broke and what I learned

The fragile part was not just the model. It was data and process consistency across seasons and source files. Stable scripts and clean outputs mattered as much as feature choices.

Current status and next step

Current status is a pushed codebase update and a clearer modeling narrative. Next step is tighter validation across holdout seasons and better assumption docs per feature block.

Links:

• GitHub: https://github.com/pbathuri/Kory_The_Cat-NCAA
• Core readme source: Work/Kory_The_Cat-NCAA-main/Zerve/README.md