How the Model Works

30 enrichment signals, a heuristic base model, a gradient-boosted machine on 500K+ MLB plate appearances, and honest limitations.

The 30 Enrichment Signals

Every bet the Randingers bot considers gets scored through a stack of 30 enrichment factors. Some apply to totals, some to props, some to moneylines. Each is a multiplier on the raw model probability.

IDSignalSourceApplies to
M1-M5Base heuristic model (team form, matchup, etc.)InternalAll
M8Weather (temp, wind vs CF bearing, pressure, humidity, air-density index)OpenWeatherMapTotals, HRs
M9Park factorStatic tableTotals, HRs
M11Team form (last-5 W-L, run diff)MLB StatsAPIML, totals
M12Starting pitcher formMLB StatsAPIAll MLB
M13Home-plate umpire (45d rolling K% + RPG)MLB StatsAPITotals, K props
M14Bullpen fatigue (recent-outing count)MLB StatsAPIML late-game
M15Rest advantageMLB scheduleML
M16Statcast metrics (xwOBA, whiff%, K%)Baseball SavantPlayer props
M20Opposing-pitcher fatigue (pitch count, days rest)MLB StatsAPITeam runs, K props
M21Batter-vs-pitcher career splits (min 15 PA)MLB StatsAPIPlayer props
M22Batting-order slot × Time-Through-OrderMLB StatsAPIHitter props
M23Handedness platoon splitsMLB StatsAPIHitter props
M26Stolen-base opportunityMLB StatsAPITotal bases
M29Injury / DTD veto (M29)MLB TransactionLogPlayer props
M30Career umpire tendency (91 umps, 5,160 games)MLB StatsAPITotals, K props

Plus in-flight: F1 (X/Twitter beat-writer scan for scratches), RLM (reverse-line-movement), Steam (coordinated multi-book moves), Sniper (sharp-edge detection with Pinnacle confirmation).

The Ensemble

Two models compete for every prediction:

  1. Heuristic base model. A weighted formula built from years of manual analysis. Fast, interpretable, resistant to overfitting.
  2. Gradient-boosted machine (model_v2). Trained on 500,000+ MLB plate appearances via time-series cross-validation. Learns nonlinear interactions the heuristic misses.

Final probability = w_gbm × p_gbm + (1 - w_gbm) × p_heuristic, where w_gbm is tuned per market by weekly Brier-loss minimization on historical results.

Model isotonic calibration maps raw scores to empirical hit rates. This is why our published win rates match our probability estimates.

Bet Selection

The AI Trader logs 10 straight bets per day, split into two windows:

  • Morning window (11 AM CT): 6 picks, filed after morning lineup drops
  • Evening window (6 PM CT): 4 picks, filed after confirmed lineups

Every pick must clear:

  1. Minimum edge threshold (varies by market)
  2. Pinnacle confirmation — Pinnacle's no-vig line must corroborate our edge
  3. Injury veto (M29) — player not on IL / DTD
  4. Scratch veto (F1) — no beat-writer scratch signal
  5. Bovada cushion — Bovada bets need extra edge (softer market)
  6. Too-Good-To-Be-True (TGTBT) veto — if edge > 8pp AND Pinnacle doesn't confirm, skip

5 picks are "sharp" (-105 to -120) and 5 are "dogs" (up to +410). Sharp lines give us tighter risk; dogs give us upside on high-conviction plays.

Bankroll Math

Bet sizing uses Kelly Criterion at 0.5× (half-Kelly) for volatility control.

Unit size: $250. Bankroll is always computed idempotently:

bankroll = starting_bankroll + SUM(profit) for each graded bet

Never bankroll += delta — that drifts over time. A nightly reconciliation cron flags any discrepancy between stored and computed bankroll.

Closing Line Value (CLV)

CLV is the single best predictor of long-term betting profitability. It measures whether the price you bet was better than the closing line.

Formula: CLV (bps) = (closing_implied − bet_implied) / bet_implied × 10000

Positive CLV = you got value the market later confirmed. Our platform captures CLV hourly on every bet automatically.

Randingers current all-time avg CLV: +1,989 bps across 113 captured bets, with 77.9% positive-CLV rate.

Honest Limitations

The model is not perfect. Here's what it doesn't do well yet:

  • Live in-game re-pricing beyond basic run-expectancy math (v1 scaffold)
  • Deep NBA/NFL modeling (v1 scaffold — MLB is where the depth lives)
  • Public bet % / handle data (requires paid Action Network feed we haven't wired)
  • Referee tendencies for NBA / NFL / NHL
  • Advanced injury data (currently uses free MLB TransactionLog, not paid RotoWire)

All limitations are on the public roadmap. We ship improvements weekly.

Verify Everything

We don't hide our numbers. Every graded bet is public.