3
4

Advanced Testing Methods

Optionally enable advanced training strategies. These stack with your model and parameter choices.

Select which tournament year to predict. The model trains on all other years.

👥
Ensemble Method
Train a separate model on each year's data. Each year's best model gets a "vote" for the final prediction. You can also view how each year's model predicted that bracket.
2011
2012
2013
2014
2015
2016
2017
2018
2019
2021
2022
2023
2024

When weighted, years with higher bracket scores have more influence on the final prediction. Unweighted uses simple majority voting.

Each selected year trains its own model. View individual year predictions on the results page.

📊
Round Splitting
Split training by tournament rounds. Train separate models for early rounds vs later rounds, where matchup dynamics differ significantly.
2
R64 (32 games) Finals (1 game)
Upset Watch
Train a specialized model on games with large seed disparities (e.g., 1 vs 16, 2 vs 15). This model is applied only to matchups with significant seed differences.
5+

Matchups where seed difference >= this threshold will use the upset-trained model. Example: A 12-seed vs 5-seed (diff=7) would qualify with threshold=5.

🎲
Randomness Factor
Add controlled chaos to your bracket. The predicted favorite gets upset a certain percentage of the time, simulating the unpredictability of March Madness.
10%

In each game, there's this % chance the model's predicted winner gets flipped. Adds excitement and accounts for Cinderella runs!

Don't want any advanced methods? Just click Start Training with nothing enabled.

?