Ensemble Method: Instead of training on all years combined, each year gets its own model. The final bracket uses majority voting across all year-models. Best for robust predictions.
Round Splitting: Early-round games (lots of blowouts) have different dynamics than late-round games (close matchups). Training separate models for each stage can improve accuracy.
Upset Watch: A special model trained only on games with large seed gaps. When Cinderella stories happen, this model is better tuned to detect them.
Randomness Factor: Real brackets have chaos. This setting randomly flips some predictions, simulating the unpredictability that makes March Madness special.
All methods are optional and can be combined. If you're unsure, try the default (no methods enabled) first!