Onions Basketball Index

The Onions Basketball Index is a probabilistic model for predicting NCAA tournament outcomes. It combines KenPom efficiency ratings with historical upset patterns to generate a complete 63-game bracket prediction, including win probabilities for every matchup.

The name comes from the classic basketball phrase “that guy has onions” — meaning a player has guts, nerve, the willingness to take and make the big shot. The model tries to capture that same energy: which teams actually perform when it matters?

Methodology

The model has three versions, each building on the last:

• v1Logistic Efficiency Model — Uses KenPom adjusted offensive and defensive efficiency ratings as inputs to a logistic regression. The difference in team efficiencies produces a win probability for each matchup, which is then simulated forward through the bracket.
• v2Efficiency Splits— Breaks efficiency into offensive and defensive components rather than net rating. This captures stylistic mismatches — a great defense vs. a great offense plays out differently than two average teams.
• v3Upset Detection— Adds historical seed-matchup upset rates as a prior. If 12-seeds beat 5-seeds 35% of the time historically, the model incorporates that signal even when pure efficiency says otherwise. This is the version shown in the bracket below.

def predict_matchup(team_a, team_b, kenpom):
    """Logistic model: P(A wins) from efficiency gap."""
    eff_a = kenpom[team_a]["adj_eff"]
    eff_b = kenpom[team_b]["adj_eff"]
    diff = eff_a - eff_b

    log_odds = 0.1462 * diff  # fitted coefficient
    prob_a = 1 / (1 + math.exp(-log_odds))
    return prob_a

def simulate_bracket(bracket, kenpom, n=10000):
    """Monte Carlo: simulate full tournament n times."""
    champion_counts = defaultdict(int)
    for _ in range(n):
        winners = {}
        for round_num in range(6):
            matchups = get_matchups(bracket, round_num, winners)
            for a, b in matchups:
                p = predict_matchup(a, b, kenpom)
                winner = a if random.random() < p else b
                winners[(round_num, a, b)] = winner
        champion_counts[winners["final"]] += 1
    return champion_counts

Key Predictions (v3)

• Champion:Duke (1) over Arizona (1) — 55.6% win probability in the final.
• Final Four:Duke, Florida, Arizona, Michigan — all four 1-seeds make it in v3 as well, but with more upset variance along the way.
• Closest game:Arizona (50.3%) over Michigan in the semifinal — essentially a coin flip.
• Notable upset picks: Texas (11) over BYU (6), Saint Louis (9) over Georgia (8), Tennessee (6) over Virginia (3), and Texas Tech (5) over Alabama (4).

Championship Probabilities

Top teams by probability of winning it all (v3 model, 10k simulations).

The v3 upset detection model picks more upsets along the way — like Texas over BYU and Tennessee over Virginia — but still lands on an all-1-seed Final Four. The historical seed priors add variance in early rounds without dramatically reshaping the deep bracket, where efficiency advantages compound.