How The Model Works

Raw efficiency ratings alone can overfit to regular-season performance. We blend the model's win probability with historical seed upset rates using a Bayesian prior with 30% weight.

This ensures our model respects empirical tournament patterns. For example, 12-seeds beat 5-seeds 35% of the time historically — even when ratings disagree, the prior pulls the probability closer to that base rate.

Our primary tournament model is a bracket-path probability engine, not a raw simulation count. Once we compute every pairwise matchup probability, we propagate those probabilities through the bracket tree to estimate each team's chances of reaching every round.

That means championship odds and round advancement rates come from the bracket structure itself, not just noisy sampling. Path difficulty is still captured naturally: a strong team in a brutal region gets penalized because their route to the title is objectively harder.

We keep team win probabilities objective, then build bracket picks on top of them. That recommendation layer weighs projected round value, bracket-path strength, and round-by-round ownership pressure.

In other words: the probability model estimates who is most likely to advance. The recommendation engine decides who is smartest to pick for your pool.

Monte Carlo is still useful, but it's no longer the headline claim. We use large simulation batches to validate that the bracket-path model and stochastic tournament outcomes agree, and to stress-test pool-level variance where "what your opponents pick" matters.

In plain English: simulation is great for scenario analysis. It's weaker as the sole proof that the underlying team probabilities are well estimated, so we use it as a check, not as the entire methodology.

That means our recommendations are driven by the probability model and recommendation layer first. Simulation helps us understand how those recommendations behave under realistic tournament variance.