Methodology

Schedules, scores, results: ESPN public scoreboard feed for NHL, NBA, NFL and MLB, synced hourly and backfilled ~400 days for model training. The provider layer is swappable (e.g., Sportradar) without changing the engine.

Odds: The Odds API (US sportsbooks), stored as append-only, timestamped snapshots.

Not yet connected: injuries, lineups, starting goalies/pitchers, weather, and licensed advanced metrics (xG, EPA, DVOA, FIP). Missing data lowers a pick's data-completeness score or blocks it entirely. It is never invented.

Elo ratings rebuilt from every completed game, with home advantage, margin-of-victory scaling (NBA/NFL) and off-season regression.

Opponent-adjusted scoring model: iterative attack/defence ratings over a season window and a recent-form window.

Seeded Monte Carlo (default 10,000 runs): Poisson goal/run distributions for NHL and MLB; normal margin and total distributions for NBA and NFL. Rest and back-to-back effects applied. Seed, inputs, count and outputs are stored.

Bayesian market prior: each handicapper blends its model with the no-vig market probability using its own weight.

EV = (model win probability × decimal odds) − 1, plus push probability for spreads and totals. No-vig probability removes the bookmaker margin from both sides.

An uncertainty penalty grows as sample size shrinks and as data completeness drops. A pick qualifies only if uncertainty-adjusted EV and edge both clear the handicapper's thresholds.

Stakes: 0.5u / 1u / 1.5u / 2u tiers based on adjusted EV, capped per handicapper. Never escalated to chase losses.

Every handicapper evaluates every market with its own feature weights, market prior, thresholds, sample requirements, leagues and special rules (underdog-only, buy-low regression, line-shopping, required data feeds). Agreement only happens when independent evaluations land on the same side.

Handicappers are AI personas. Modeled experience is not employment history.

Consensus counts handicappers that actually published, shown against the eligible denominator. Because all models share inputs, we estimate average pairwise correlation (ρ) from their weight vectors and report effective agreement = n / (1 + (n − 1)ρ).

Daily schedule (America/Toronto): 05:00 schedule discovery · 06:00 initial modeling · 08:00 morning analysis · 10:00, 13:00, 16:00 publishing runs · hourly score sync, settlement and pre-game checks.

Picks are rejected by the database if the game has started. Selection, price, sportsbook and timestamps are immutable; updates are appended as revisions. Picks settle from final scores; postponed/cancelled games void.

Win % = wins ÷ (wins + losses); pushes and voids excluded. ROI = net units ÷ units staked × 100. A low win rate on underdogs can still be profitable; a high win rate on heavy favourites can lose money.

All published records on this site are live tracked. Any future backtest will be labelled with its period and method and kept separate.

Sports outcomes are highly random. A 60% probability loses 40% of the time. Short samples mislead — always check the number of picks.