⚾ MLB — All Models

Why three numbers, not one: accuracy alone can mislead — a 76% record profits at -110 but LOSES at -400. ROI is the real profit at the prices we actually posted; CLV (did we beat the closing line?) is the honest leading indicator of edge — beating the close >55% predicts profit even through short-run variance. ROI & CLV fill in as we log each pick's odds; we show them even when they're unflattering.

How we make these predictions. Every MLB pick comes from a model, then is checked against the market so it stays honest.

• Hits (and runs / RBIs): we start from the batter-versus-pitcher matchup (a Log5 calculation of the hitter's rate against this specific pitcher), then adjust for the things that actually move a hit: the ballpark, the opposing starter, the weather (wind and air), platoon splits (lefty/righty), recent form, and Statcast expected batting average (xBA, quality of contact). We then anchor the number partway to the sportsbook's posted price, so a hot model reading never runs away from reality.
• Strikeouts: we project a starter's strikeouts from their strikeout rate and how much the opposing lineup whiffs, then anchor to the DraftKings posted strikeout line and report our probability of the Over.
• NRFI (no run in the first inning): a calibrated probability built from both starting pitchers and the top of each lineup.

Every model is calibrated against real results and graded on profit and closing-line value (did we beat the number the market closed at), not just on how often a pick wins. That scoreboard is right above. Research tool, not betting advice.

AVG shows the season number and H+R+RBI shows tonight's projection, each with the player's last-5-game average (L5) beneath. A rising L5 means the hitter is hot lately. Projected lineups = each team's most recent batting order; actual lineups can differ. Method note: a Log5 base (batter average vs the starter) adjusted by five capped factors — platoon (the batter's split vs the starter's handedness), ballpark (Coors up, pitcher parks down), recent form (last-10 blended in), weather (warm air helps a little), and Statcast xBA (regress lucky/unlucky averages toward the batter's actual contact quality). Following Baseball Musings (David Pinto), hitters facing LOW-strikeout (contact) pitchers also get a boost — see 'opp SP K%' (lower = better for hitters). H+R+RBI = projected hits+runs+RBIs (o1.5 = P(over 1.5)). Runs & RBIs are noisier than hits. Factors are priors, not yet backtested — research tool, not betting advice.

projK = projected strikeouts. DK O/U = DraftKings' actual posted line with our lean (the side our projection favors) and P(over it); this judges us against the number the book really hangs, not a fixed line. o4.5/o5.5/o6.5 = P(over those fixed lines) for reference. Backtested: MAE 1.85 (beats baselines), proj/actual corr 0.39, O/U-5.5 62% — the most reliable model here. Research tool, not betting advice.

Read this: the big % are the raw model lean (ranking only). Out-of-sample backtesting shows these are over-confident — the calibrated estimate (≈50%) is honest, and the model does not beat a typical NRFI price. Research screen, not betting advice. ⚠️ = low sample.

Current consecutive games with a hit (a no-at-bat game doesn't break a streak). Only players in today's projected lineups appear. A streak is a momentum signal, not a guarantee.

L5 = batting average over the last 5 games (the 'hot bat' signal, like Caglianone's .450); AVG = season; P(hit) = model chance of a hit today. Only players in today's projected lineups. Hot stretches regress toward the season number, so a hot bat is a lean, not a lock.

✓ = predictions locked in BEFORE games (trustworthy). ~ = backfilled from current stats (optimistic — ignore for judging accuracy). Strikeouts is the validated model; NRFI is ~a coin flip. Not betting advice.

Read this: if the model is calibrated, 'Model says' ≈ 'Actually'. Overconfident = we claimed more than we delivered (shrink those probabilities). Underconfident = we delivered more than we claimed. Hits ran overconfident historically (Platt calibration was applied 2026-06-28; older logged picks here are pre-calibration). Strikeout overs run underconfident — the overs hit more than stated. Recalibration: hits → keep the Platt shrink and refit weekly on post-6/28 logs only; strikeouts → the o5.5 probabilities can be nudged UP (or fit a Platt correction on P(over)).