MLB Betting Dashboard — Loading… LIVE

📐 Model Methodology — Why the numbers say what they say

The model is an ensemble of eight inputs that each independently move a game's run, strikeout, and win-probability projections, then get blended with empirically-derived weights. Nothing here is guessed — every component is either a MLB Stats API pull, an Open-Meteo weather pull, or a hand-tuned coefficient derived from multi-year league baselines.

• 25%Starting pitcher quality — season ERA, WHIP, K/9, and BB/9. Each point of ERA above 4.00 adds ~0.25 runs to the team total; WHIP above 1.30 inflates opposing OBP expectation.
• 20%Handedness splits — the team's actual OPS/OBP/AVG/K% against the exact hand of pitcher they're facing. A lineup that OPSes .810 vs LHP but .680 vs RHP is a completely different offense depending on who starts.
• 15%Park factors — 30-park table (runs, HR-L, HR-R, doubles, walks). Coors (1.35) and Great American (1.10) amplify; Oracle (0.91) and loanDepot (0.94) suppress. Home-run factors split by batter handedness because Yankee Stadium boosts LHB HRs 24%.
• 15%Wind vector projection — meteorological wind direction is projected onto each park's home-to-CF bearing so "8 mph NNE" becomes actionable "OUT to LCF." Wind OUT adds up to ~1.2 runs; wind IN subtracts ~0.9. Domes ignore wind.
• 10%Opponent K% and contact profile — high-K lineups inflate pitcher K prop overs and team-K-thrown totals. Low-K lineups suppress them even against ace pitchers.
• 8%Bullpen load — proxied by recent IP; fatigued pens lift late-game totals and hurt the F5 under.
• 5%Temperature and humidity — warm air is less dense (ball travels ~0.4 ft per 10°F) and adds carry, nudging HR projections.
• 2%Home-field baseline — +0.20 runs of model edge for the home team, applied after all other inputs.

What the model believes: most public totals over-react to the last 3 games of either team and under-weight park/wind, which is why it looks for ≥0.8 runs of disagreement before flagging an OVER/UNDER. On strikeout props it trusts K/9 over reputation — a 30% K% SP facing a 26% K% opponent at Tropicana will project higher total Ks than a household-name ace facing Cleveland. On moneylines the model intentionally stays small because public books are sharpest there; it only flags >0.3 ERA-equivalent gaps. Human-element layer (news/injury) is currently a qualitative overlay you should apply after reading the narrative, not auto-fed.