MLB Over/Under Predictions with AI: How Our Models Analyze Pitchers, Rotations, and Totals

A starter posting a 4.20 ERA but a 3.30 FIP — the market overvalues that pitcher's danger, which is actually an Under opportunity. That's exactly what our models catch, and what most bettors miss.

MLB totals betting gets sold as the simplest market: take a line (say 8.5 runs), decide whether the combined score goes over or under. In practice, it's one of the hardest markets to beat long-term — because it depends on four variables that shift every single day: the starting pitcher, bullpen health, the ballpark, and weather. Ignore any one of them and the market takes back the edge.

Here's how ProbWin breaks down each lever before every MLB game.

Key takeaways - MLB totals depend on 4 combined variables: starting pitcher (FIP), bullpen fatigue, park factor, and weather — each one alone is not enough. - FIP and xFIP are better predictors of a starter's future performance than raw ERA, which absorbs defensive variance and luck. - In 2026, MLB bullpens account for 45% of all innings pitched — their day-to-day fatigue is consistently underpriced by sportsbooks at the 24-hour mark. - Our TOTALS_FT model hits 59.5% win rate on 1,349 official picks since April 2025; the F5 market reaches 62.1% on 1,192 picks since May 2024.

Why MLB Totals Have a Hidden Value Structure

Baseball is a high-variance sport. Scores like 1-0 and 13-5 are equally part of the game. That variance creates two traps for bettors: overestimating how predictable totals are, or not exploiting them enough.

The reality sits in between. MLB totals are beatable long-term — if you build an independent score estimate before looking at the posted line. That's the core of the method: estimate first, compare second.

The market is efficient on standard cases (an elite pitcher duel between two weak offenses is already in the number). Value shows up in non-linear cases: an undervalued starter, a bullpen running on fumes after extra innings, a 20 km/h wind blowing in at Wrigley, a starter scratched at 10 AM.

The market handles the visible data well — starter ERAs, general bullpen quality. It underprices short-window information (yesterday's bullpen usage) and non-linear combinations (extreme park factor + record heat + depleted bullpen). That's where the edge lives.

Starting Pitcher Impact: ERA vs. FIP and What the Market Misses

This is the most visible variable, so it's the most efficiently priced — on average. "Efficient on average" doesn't mean "always right."

The problem with raw ERA: it absorbs runs that depend on the defense and on the variance of balls in play (BABIP). A pitcher with a 4.50 ERA and a 3.20 FIP has simply been unlucky on batted balls — his ERA will regress toward his FIP over the coming weeks. The market keeps pricing him at 4.50 in the meantime.

FIP (Fielding Independent Pitching) isolates only what the pitcher controls: strikeouts, walks, home runs. xFIP goes further by normalizing home run rates to the league average (removing ballpark effects from HR). These two metrics are our entry points for evaluating a starter.

Under opportunity signal:
  Starter ERA ≥ FIP + 0.80
  → Market penalizes more than the pitcher's actual level
  → Line is too high, Under value

Over opportunity signal:
  Starter FIP ≥ ERA + 0.80
  → Starter performing better than ERA shows
  → Market sees the pretty ERA... and misses the coming regression

For a deeper look at why ERA misleads, read our ERA vs xERA guide.

Mid-season (July–August), rotations thin out. Starters accumulate innings, managers shorten their leashes, and bullpen games or openers appear more frequently. These games require a different approach: forget the fictional starter's FIP, model the expected score directly from each team's combined bullpen output. Our model detects these games automatically and reweights the variables accordingly.

Mid-Season Bullpen Fatigue: The Most Underpriced MLB Betting Lever

In 2026, MLB bullpens account for 45% of all innings pitched league-wide. That share has climbed every year for a decade. The direct consequence: what the starter does matters less than it used to, and bullpen health matters more.

The problem for sportsbooks: bullpen fatigue at the 24-hour mark is calculable from yesterday's box scores — public data — but most pricing models fail to account for 3-4 day accumulation. A bullpen that absorbed 6 innings over 48 hours across two extra-inning games will perform measurably worse tonight. The line won't necessarily move.

Our bullpen tracking table monitors each team's relief corps daily (innings thrown per pitcher over 72 hours, aggregated freshness score). Across 3,150 records since April 2026, the scoring impact is clear:

A +1.1 run swing on an 8.5 line is the difference between cashing the Over and missing. Especially when combined with a hitter-friendly park.

Mid-season Over profile — bullpen fatigue signal:
  ✓ Two consecutive 9+ inning games in the last 72 hours
  ✓ Opposing bullpen freshness score below 40/100
  ✓ Day game (less recovery time since evening action)
  ✓ Ballpark factor ≥ 1.05

The inverse applies too: two fresh bullpens in a pitcher-friendly park is a solid Under profile even with ordinary starters. Combined signals always outperform single-variable reads.

For a full breakdown of bullpen metrics and how they move lines, check our Bullpen ERA guide.

Park Factors and Weather: Coors Field to Petco Park, the Extremes That Change Everything

Park factor is a permanent variable — it doesn't change game to game — and the market generally prices it correctly. Two situations still create value:

1. Extreme parks, where public lines occasionally underreact on lower-profile matchups
2. Weather, which dynamically modulates the park's effect

Parks That Inflate Totals

Coors Field is in a category of its own. Denver's altitude (5,280 feet) thins the air: balls travel farther, breaking pitches break less, and bullpens allow runs at abnormal rates. According to 2026 Statcast data, Coors Field's park factor sits at 110 (10% more runs than average, with a 3-year run index of 128). A Colorado home game with a 10-run line when the MLB median is 9 isn't an outlier — that's the park talking.

But even at Coors, elite starters plus fresh bullpens can produce an Under. Park factor is the multiplier, not the decision itself.

Parks That Suppress Scoring

Petco Park (San Diego) is the perfect opposite of Coors. The marine layer off the Pacific and the stadium's dimensions compress scoring. A 7.5 line in a Padres home game with a quality staff becomes even more interesting when both bullpens are fresh.

Weather adjusts all of this. A headwind at Wrigley Field temporarily converts a neutral park into a pitcher-friendly environment (-0.5 to -0.8 runs). Dry heat in Phoenix extends ball carry (+0.3 to +0.5 runs). These dynamic adjustments don't flip the bet direction — they sharpen the probability around the line.

Our MLB Park Factors guide lists all 30 parks with their coefficients and year-over-year trends.

How Our Model Builds a Total Estimate

Our process is sequential: we never look at the posted line before generating an independent estimate. Looking at the number mid-calculation creates anchoring bias — the book's figure would unconsciously pull our own estimate.

Step 1 — Offensive projection for each team Expected runs created from rolling 15-day offensive metrics, adjusted for the matchup (pitcher handedness, lefty/righty splits). See our Platoon Splits MLB guide for how handedness affects run production.

Step 2 — Starting pitcher adjustment FIP, K/9, BB/9, ground ball rate (GB%) — each metric weighted by season sample size reliability. A starter at 50 innings gets less weight than one at 130 innings.

Step 3 — Bullpen adjustment Proprietary freshness score (innings/72h, recent adjusted ERA). Our tracking covers 3,150 observations since April 2026.

Step 4 — Park + weather adjustment Park factor multiplier, corrected by day-of-game weather forecasts (wind direction, temperature, relative humidity).

Step 5 — Poisson distribution The combined expected score (e.g., 8.3 runs) gets converted into a probability distribution around the market line. If P(total ≥ 9 | line = 8.5) exceeds 57%, that's an actionable Over signal.

Simplified example (illustrative):
  Team A offense: 4.8 runs/game (rolling 15d)
  → Adjusted for opposing starter FIP 3.30: 3.9 expected runs

  Team B offense: 4.2 runs/game
  → Adjusted for opposing starter FIP 3.80 (ERA 4.50): 3.5 expected runs

  Raw total: 7.4 runs
  × Ballpark factor (1.01): 7.47
  + Weather adjustment (12 km/h tailwind): +0.6 → 8.1 runs expected

  Market line: 7.5
  Over 7.5 signal: +0.6 gap, P(Over) ~59% → PICK

Historical model performance On 1,349 official TOTALS_FT picks since April 2025: 59.5% win rate, +201.3 units. On 1,192 TOTALS_F5 picks since May 2024: 62.1% win rate, +236.4 units.

These results aren't linear — there are losing streaks, as with any model facing probabilistic sports. What matters is that the edge shows up across a large sample, not that every week is green.

F5 vs. Full Game: Choosing the Right Market for Your Pick

The First 5 Innings (F5) market is often pitched as "simpler" because it isolates the starters. It's actually a different market with its own value structure.

Our F5 performance (62.1% WR) edges ahead of FT (59.5%) — not surprising. A shorter window means less bullpen noise, and a model well-calibrated on starters has fewer moving parts to absorb over 5 innings than 9.

Practical rule: if the pick's value comes mainly from a FIP/ERA gap on the starter, play F5. If value comes from bullpen fatigue or an extreme park factor combined with a hot offense, play FT — the advantage tends to materialize in the late innings.

For the complete breakdown of the F5 market, read our First 5 Innings (F5) guide.

Frequently Asked Questions

How do you use AI to predict MLB Over/Under totals? The process has 4 ordered steps: (1) assess the starter's FIP rather than ERA, (2) check bullpen health over the past 72 hours, (3) apply the ballpark's park factor, (4) adjust for game-day weather. Our model generates an independent run estimate and compares it to the market line — a gap of 0.5+ runs with P > 57% triggers a pick.

Is ERA enough to analyze MLB totals bets? No. ERA includes runs that stem from defensive performance and luck on balls in play (BABIP). FIP and xFIP are stronger predictors of a starter's future behavior. A pitcher with a 4.50 ERA and a 3.20 FIP is being overvalued as a danger by the market — that's an Under signal.

What is Coors Field's impact on MLB totals betting? Coors Field's park factor is 1.10 (2024), meaning 10% more runs than the MLB average. Lines in Denver typically run 0.5 to 1.5 runs higher than standard. Denver's altitude (5,280 feet) reduces air resistance, letting balls travel farther and making breaking pitches less effective.

When is the best time to bet the Over on an MLB mid-season game? The strongest Over profile: fatigued opposing bullpen (6+ innings/72h) + park factor ≥ 1.05 + temperature > 25°C + tailwind. Our data shows these combined conditions add +1.1 runs above the average line — often underpriced by sportsbooks at 24 hours out.

What is the difference between MLB F5 and full-game totals? F5 covers the first 5 innings, isolating the starter and reducing bullpen variance. Full game includes the complete contest. Play F5 when the edge comes from a starter's FIP divergence. Play FT when it comes from bullpen fatigue or an extreme park factor — those advantages show up late in the game.

Next Step

MLB totals reward methodological discipline over the long run. Starting pitcher (FIP, not ERA), bullpen (72-hour freshness), ballpark (park factor), weather (dynamic adjustment) — these four levers can't be read in isolation. Their combination — or lack of one — determines whether a genuine edge exists.

To go deeper on each dimension: - ERA vs xERA — Why Traditional ERA Misleads Your MLB Bets - Bullpen ERA: The Hidden Factor That Makes or Breaks Your MLB Bets - Park Factors MLB: How the Stadium Influences Your Bets - Confirmed Pitcher MLB: Why This Information Is Worth Gold for Your Bets