How our models actually work

For each game, we compute a probability distribution over outcomes (moneyline, spread, total) using team and player inputs. We compare that probability to the sportsbook's implied probability. The difference is the edge. Edges above our per-market threshold become recommendations.

Edges below the threshold become "no bet." We'd rather skip a game than force a marginal pick.

The NBA engine is a 2,000+ line model that incorporates:

• Team efficiency (offensive and defensive rating, pace)
• Rolling efficiency over last 10 / last 5 games (weighted recency)
• Player impact — top-8 rotation, adjusted for availability
• Schedule context — rest days, back-to-back, travel miles
• Injury adjustments from official injury reports
• Line movement signal (where sharp money is moving the line)
• Home/road splits and pace-adjusted matchups

Output is a sigmoid-shaped probability with a calibration factor (K = 0.25 currently) tuned against historical closing lines. We cap maximum edge at 12% to avoid over-confident picks from thin samples.

The MLB engine is a 1,000+ line model focused on pitcher matchups and composite scoring:

• Starting pitcher ERA, xFIP, K/BB, adjusted for opponent lineup handedness
• Bayesian prior blending for early-season uncertainty (20-start weight)
• Lineup composition against the starter's arm side
• Bullpen strength and recent usage (leveraged innings remaining)
• Park factors for run environment
• Weather and wind (for totals and parks like Wrigley)
• Home plate umpire strike zone tendency
• Line movement signal

We publish NRFI probabilities separately because first-inning dynamics differ enough from full-game dynamics to deserve their own model.

These sports launch later in the 2026 roadmap. The base architecture mirrors NBA and MLB: team/player inputs, schedule context, injury adjustments, edge thresholds, public recalibration. Sport-specific wrinkles include recruiting-to-performance gap analysis for college football, pace and tempo adjustments for college basketball, and rest/travel situational edges for NFL.

When each sport launches, this page will expand with a dedicated methodology section. Until then, see the dashboards at /stats for the analytics backbone those models will sit on.

Not every predicted edge is a bet. Our decision tree:

Thresholds are sport-specific and tuned against historical CLV (closing line value). We err toward PASS — a missed bet is recoverable; a bad bet isn't.

Every Monday, the recalibration job:

• Pulls last 7 days of finalized games
• Compares model predictions to closing line value (CLV) and actual outcomes
• Runs a grid search over key weights (efficiency, recency, injury impact)
• If a reweighting reduces Brier score without overfitting, updates config
• Logs the change to our public methodology changelog

Live recalibration results: /transparency →

• NBA Stats API (stats.nba.com) — team, player, schedule, injuries
• MLB Stats API + Baseball Savant — pitcher splits, Statcast
• CollegeFootballData.com, CollegeBasketballData.com — NCAA data
• nflverse — NFL play-by-play and stats
• The Odds API — live odds across 40+ sportsbooks
• OpenWeatherMap — weather for outdoor sports
• Public injury reports and team announcements

Full data-source list with update cadence: /data-sources →

Equally important:

• We don't sell individual picks — all recommendations are model-derived and public
• We don't offer "guaranteed winners" — nobody should
• We don't retroactively edit historical picks
• We don't rank sportsbooks by commission rates
• We don't use unit sizing that amplifies losses ("10-unit locks")
• We don't model anything where we lack sufficient historical data

Our models are tools, not oracles:

• Early-season samples are noisy; we use Bayesian priors but confidence is still lower in March (MLB) and October (NBA)
• Player-specific props are lower-confidence than team-level bets
• Injuries can invalidate predictions between lineup drop and tipoff
• Edges of 1–3% are statistically indistinguishable from zero in small samples
• Sharp books (Pinnacle, Circa) price efficiently; our edge comes mostly from soft books