● METHODOLOGY

How the AI Model Works

Every pick on Bookie Bullies comes from a documented, reproducible AI model. No black box, no tout hype. Here's exactly what goes in, what comes out, and how picks are graded.

Updated 2026-09-08

MLB + college football coverage

Positive-EV filter

The Model in One Paragraph

For every game on the slate, the model ingests sport-specific inputs, simulates the matchup, and produces a win probability for each side. That probability is compared to the market moneyline. If the model's probability is higher than the market implies, the pick has positive expected value. Only positive-EV picks make the card. Every side also gets a confidence rating and a Kelly-fraction staking suggestion.

MLB Inputs

Starting pitchers: ERA, xFIP, stuff+, pitch mix, recent form, batter-handedness splits.
Bullpen: reliever strength, rest days, high-leverage usage, recent workload.
Offense: team wOBA splits vs. RHP/LHP, recent-form adjustment, lineup order.
Park factors: run-scoring environment, HR-factor, foul-territory adjustments.
Weather: wind direction and speed, temperature, precipitation risk.
Umpire: strike zone tendencies (K% delta vs. league average).
Lineups: live lineup card once posted, fallback to projected lineup.
Travel + rest: back-to-back travel, time zone crossings.

College Football Inputs

College football is where we publish leans, not picks, and the distinction is the whole point. A lean is what the model prefers. A pick is a lean that has proven it beats the closing line. Ours has not, so we label it honestly and grade every one in public at the college football board.

Opponent-adjusted power ratings: ridge-regularized least squares solving team strength and a home-field constant from final scores alone. Ratings for any game are fit only on games played before it.
Scoring-tendency ratings: a second solve for how high or low scoring each team's games run, used for the over/under lean.
FBS-only filter: teams with fewer than 12 appearances are treated as FCS guarantee-game opponents and excluded. Leaving them in inflated our home-field constant to 6.4 points against a real 3.7.
Minimum sample gate: a team needs 8 games in the fit before its rating is usable. Without it, unrated teams sit at league average and manufacture fake 25-point disagreements.
Market line from the detail string: we parse the posted line text, never the numeric spread field, which carries a post-hoc in-game number on completed games.

What the College Football Model Has Not Done

Six approaches have been tested against real closing lines. None beat the market:

Situational angles: 16 rules (big favorites, home dogs, ranked matchups, bowls, neutral sites). None held in both halves of the data.
Power ratings against the spread: 49.2% across 968 walk-forward games, with the market's number beating ours by 2.73 points of mean absolute error per game.
Scoring-tendency ratings against the total: about 51% across 429 games, market better by 0.75 points.
Prior-season efficiency ratings: 48 to 52%, indistinguishable from noise.
Current-season efficiency ratings: untestable. The only available history is end-of-season data that already knows how the games finished, so any backtest is measuring hindsight.

Our leans are ranked in a deliberately counter-intuitive order: smaller disagreements outrank larger ones, and leans on favorites outrank leans on underdogs. Both rules come from results rather than taste. Games where we differ from the book by 14 or more points went 2-7 on the first graded slate, and the ridge shrinkage that produces our ratings structurally over-likes underdogs, so a favorite lean is one of the few reads not generated by that bias.

Why Each Sport Has a Different Algorithm

Sports betting models that apply the same equation to every sport leave money on the table because the underlying score distributions differ:

MLB uses the Skellam distribution for run margins (integer-run differences) and Negative Binomial for run totals (overdispersed Poisson). Pitcher quality (FIP/xFIP) drives most of the margin signal. First-inning markets use a Poisson model per half-inning, which is the one product here that has beaten its number.
College football uses ridge-regularized least squares on final scores, with margins treated as roughly Gaussian around the fitted number. Scatter around a college spread runs about 14.7 points, which is wide enough that most single-game disagreements are noise.

How Probability Becomes a Pick

Raw model output is a margin or total estimate. That gets converted to a win probability via a normal CDF with sport-specific standard deviations. The raw model probability is then blended into a final probability using 50% model output, 30% market-implied probability, 10% Statcast ensemble, and 10% closing-line implied probability. The blend shrinks overconfident predictions and avoids over-betting noisy inputs. The final blended probability is compared to the market moneyline. Expected value is calculated as (prob × decimal odds) minus 1. Only picks with positive EV make the card.

Confidence Tiers

Confidence is independent of expected value. It measures signal strength, not edge size.

Lock: Inputs strongly agree, low variance, multiple corroborating signals.
Strong: Clear lean with one or two light counter-signals.
Lean: Model edge exists but inputs disagree or sample is thin.
Pass: No edge, no pick issued. These don't appear on the card.

Staking Suggestion

Every pick carries a unit-size suggestion based on a fractional-Kelly calculation (quarter-Kelly by default) against the blended probability. The suggestion is informational. Users set their own bankroll rules. Bookie Bullies does not provide money management advice and is not a licensed financial adviser.

Public Grading

Every pick gets logged the moment it's published. Graded outcomes (W / L / P for push) are posted by 8 AM Pacific the next morning. The running record is visible on the archive page. Nothing is hidden, edited, or retroactively removed.

Caveats + Limits

Models are probabilistic. Even positive-EV picks lose, often in streaks. Variance is real and the sample size required to confirm edge is in the thousands of picks, not dozens. Models also miss real-world context: last-minute injury news, umpire substitutions, weather that shifts late. Treat the output as a starting point, not a command. Wager only what you can afford to lose. Gambling problem? Call 1-800-GAMBLER.