BracketGenius

A test of whether careful prompt design and published analytics frameworks can beat naive prediction on a hard forecasting problem.

An NCAA tournament prediction system built on published basketball analytics frameworks, run live through the full 2026 tournament with bracket tracking.

73% game accuracy across 63 games, against a roughly 75% benchmark from established models. One tournament is a small sample and the approach carries real overfitting risk. The case study explains both rather than treating the number as settled.

• Problem definition
• Prompt engineering
• Evaluation & testing
• Deployment

1. 01Team inputs are drawn from published basketball analytics frameworks rather than from a hand-picked feature set.
2. 02A structured prompt produces a matchup prediction with a stated rationale, so a wrong call can be read and not just counted.
3. 03Predictions were locked before tip-off for each round, which is what makes the accuracy figure meaningful.
4. 04Results were tracked against the real bracket through the full tournament.

Built on Lovable with the Claude API doing matchup evaluation.

Prompts encode the analytics framework explicitly, including how to weigh efficiency margins and matchup context.

Each prediction is stored with its inputs and rationale so the record is auditable after the fact.

• One tournament, 63 games. That is a small sample and does not establish an edge.
• 73% sits below the roughly 75% benchmark from established models.
• Prompt design was iterated with knowledge of past tournaments, which carries real overfitting risk.
• Nothing here is a betting tool, and I would not present it as one.