Sports

MLB betting models built around fair price

Baseball prices the winner, not a handicap. With no spread to choose, the entire opinion lives in the odds, and the pitcher on the mound accounts for most of it. Model the matchup and the run environment, and the moneyline follows.

Updated October 22, 2026 · 11 min read

Analytics only. quant+ never places bets or holds funds.

  • Start. The listed pitcher. Sets the run environment for the innings he is expected to work.
  • Bridge. Setup relief. Where availability matters more than season-long quality.
  • Close. The ninth and beyond. Only reached in a close game, which is itself a conditional state.

A structural decomposition, not a measurement. No pitcher, team or park figure appears on this page, and no inning distribution is asserted.

Four properties that shape every MLB model

Each of these is a reason a baseball model is built differently from a football, basketball or hockey one, and each is a place where a generic team-strength approach goes wrong.

The moneyline has no spread

Baseball's primary market is a straight two-way price on who wins the game. There is no handicap to choose and no half point to consider. The price carries the entire opinion, which means the discipline of the sport is pricing, not line selection, and it means a small error in the implied probabilities shows up directly in the payout.

The starting pitcher is the market

Nine innings of baseball are not a team-against-team contest in the way a football game is. The starter sets the run environment for roughly the first half of the game, and the identity of that pitcher moves the line more than any position player does. A published modelling guide named the starting matchup first among relevant inputs, and the market agrees with it.

A long, low-correlation season

One hundred and sixty-two games per team means a great deal of data and a very noisy per-game outcome. The regular season is long enough for small edges to compound, which is why a modelling guide on the sport described it as unusually model-friendly: it rewards many small correct probabilities rather than occasional large convictions.

Discrete, countable, and well suited to probability

Runs, hits, strikeouts and plate appearances are events that occur at measurable rates in defined opportunities. There is no clock and no possession structure, so the game decomposes into a sequence of batter-versus-pitcher matchups, and a probability can be attached to each. That structure is why baseball accepted quantitative analysis earlier and more thoroughly than its peers.

In baseball, the price is the whole opinion

In football and basketball, a model can be right about the game and still lose because it picked the wrong number on the spread. The handicap absorbs some of the error. Baseball has no such cushion. The moneyline states the market's probability directly, and any disagreement the model has is a disagreement about a probability rather than about a point value.

That makes the conversion step unusually important. American odds, decimal odds and implied probability are three representations of the same number, and the sum of the two implied probabilities exceeds one hundred percent by the book's margin. Everything a baseball model produces has to survive that conversion unchanged, because there is no intermediate object to absorb a mistake.

It also makes the run line a genuinely different bet rather than a variation. Fixing the handicap at one and a half runs means the run line is a bet on the margin of victory exceeding a constant, which is a different question from who wins. The favourite on the run line is frequently an underdog by price, and the two markets can disagree about the same game without either being wrong.

  • Moneyline. Two-way, no spread. The price on each side reflects the market's view of the win probability, with the margin embedded. Everything in the model ultimately has to express itself here.
  • Run line. A fixed handicap, conventionally one and a half runs, applied to the favourite. The favourite gives the runs and is typically priced at plus money as a result. It is functionally a point spread with a fixed number rather than a moving one.
  • Game total. Combined runs from both teams. This is where park, weather and the two starters' run-prevention profiles land most visibly, and it is often the market where a good model disagrees most with the price.
  • Team total. One side of the game's scoring. Internal consistency with the game total and the moneyline is checkable without any forecast, which makes team totals a useful coherence test on a book's own board.

The listed pitcher moves the number more than anything else

Baseball's rotation is public and known in advance, which makes the starter the single most forecastable large input in any major sport. It is also the input that most often voids or reprices a market.

A starter is typically worked deep into the game but rarely beyond the seventh or eighth inning, and modern usage is monitored closely with pitch counts. Modern rotations run five starters, with a five-man cycle now standard, and the workload per start has fallen substantially over the decades. The opener, in which a reliever starts and hands off after an inning or two, is a recognised variant that changes the structure of the game rather than only the quality of the first inning.

The practical consequence for a model is that the game has to be decomposed by innings rather than treated as a single contest. The first six innings belong largely to the starter and the lineup he faces. The remainder belongs to a bullpen whose composition depends on what happened in the previous few days. Modelling one aggregate run rate for the whole game blends two different processes.

There is a data-hygiene point buried in this. Because the starter is named in advance, some markets are listed with that pitcher and are adjusted or voided if he does not start. Others are action regardless of who pitches. Treating an action-regardless market as a listed-pitcher market, or the reverse, is a specification error that costs money in exactly the games where the pitcher changes.

Components rather than outcomes

  • Strikeout rate per batter faced, which stabilises quickly and is largely under the pitcher's control.
  • Walk rate per batter faced, which is both a control measure and a workload driver.
  • Home run rate, which is the component most sensitive to park and weather and therefore the one that needs the environment applied.
  • Pitch count and expected innings, which determine how much of the game the starter actually covers.
  • Platoon splits, because a lineup constructed against a specific handedness is a different problem from a generic lineup.

Six inputs, each measurable before first pitch

Baseball has an advantage here: the inputs that matter most are public, structured and available well before the game.

Starting pitcher, beyond the ERA

Strikeout rate, walk rate and home run rate per batter faced are more stable and more predictive than earned run average, which is influenced by defence, sequencing and official scoring decisions. A model should use the components and treat the ERA as an outcome rather than an input.

Bullpen quality and, crucially, availability

Who is available tonight is a different question from who is good in general. A relief corps that worked three extra innings the previous night is not the same bullpen at the same quality. Usage is public and recent, which makes it a usable input rather than a guess.

Park effects

Stadiums differ in how much they suppress or amplify run scoring. The standard park factor compares runs scored at home with runs scored on the road and indexes the result so that one hundred is neutral. Park effects are large enough to move a total materially, and they interact with the pitcher's own profile.

Weather and wind direction

Wind blowing out turns fly balls into home runs; wind blowing in does the reverse. Temperature affects carry. These are measurable before the game and are among the most concretely attributable adjustments available in any sport.

Lineup construction and handedness

A lineup's composition against a particular pitcher matters, including how many left-handed and right-handed hitters it contains and how those splits interact with the starter's own platoon tendencies. The starting nine also changes with rest and injury, and the change is usually smaller than the starter change and not negligible.

Umpire and, at the margin, catcher

The home plate umpire's strike zone tendencies affect walk and strikeout rates at the margin. This is a small effect compared with the pitcher and the park, and it is included for completeness rather than as a headline input.

Park and weather, treated as multipliers rather than footnotes

Baseball is the one major sport where the venue and the weather measurably change scoring, and where the effect can be attributed to a specific physical mechanism.

Input Mechanism Which market responds most
Park factor Stadium dimensions and configuration change how often batted balls become hits and home runs, expressed as an index where one hundred is neutral. Game total, then team totals
Wind direction and speed Wind blowing out carries fly balls over the fence; wind blowing in suppresses them. The effect is concentrated in the home run component. Game total, and any market involving home runs
Temperature and humidity Warmer, more humid air is less dense, so batted balls carry further. The effect is real and smaller than wind. Game total
Bullpen availability Recent workload determines which relievers are usable tonight, which is a different question from how good the bullpen is on average. Moneyline, especially in the late innings
Umpire tendencies Strike zone width shifts walk and strikeout rates at the margin, which propagates into run expectation slowly. Marginal effect on totals and pitcher props

The rankings in the final column are ordinal judgments about relative sensitivity. No numerical effect size is claimed for any input, because none was measured for this page.

Three ways a baseball model looks better than it is

The long season is a gift for sample size and a hazard for self-deception, because there is enough data to fit almost anything.

Using season-long pitcher stats early

A pitcher's full-season rate line is not available in April, and neither is his opponent's. A backtest that uses season aggregates to predict games within that season has looked forward, and it will report an edge that cannot be reproduced in real time.

Ignoring the pitcher change rule

Some markets void or adjust when the listed starter does not pitch; others do not. A backtest that silently assumes every market settles on the listed pitcher will produce results that do not match what a bettor would actually have collected on the days the pitcher changed.

Aggregating across eras

Run environments, ball construction, opener usage and relief workloads have all changed across seasons. Fitting a single model across many years blends regimes that are not comparable, and the blend can appear statistically strong while being wrong about the current season.

The standard that matters is unchanged: evaluate in point-in-time order, score against closing prices, and require a large sample. For the general method see the model building guide.

Where an MLB model fails

The sport is unusually good at making a correct probability look wrong. A model can assign a seventy percent chance and lose, and across a three-game series that will happen. Baseball's per-game randomness is high enough that a genuine edge is essentially invisible in a week, a month, or sometimes a single season, which is precisely why the compounding-across-many-bets framing matters and why grading on win-loss over a short stretch is close to meaningless.

The most liquid baseball markets are also priced competently. A model that reports a large edge on a widely traded moneyline is more likely to have a data leak than to have found something the market missed. The places where a well-built model has more room are the markets that require several inputs to be combined correctly at once: totals where park, wind and two starters interact, and team totals where internal consistency with the other markets is checkable.

And the model cannot see the game. Weather forecasts are forecasts, lineups change after they are posted, and a manager's decision in the seventh inning is not predictable from a spreadsheet. A model that behaves as though all of that were settled is overstating what it knows.

quant+ publishes the research surface. It does not place bets and has no account connection to any sportsbook. For the player-level side of baseball markets, see MLB player prop research.

Scan the baseball board · Compare prices across books

Model the pitcher and the park

Decompose the game by innings, apply the environment, and see where the moneyline sits relative to a probability you derived yourself.

Analyze the MLB slate · MLB player props

Questions people ask

What is an MLB betting model?

A system that converts the starting pitcher's component rates, the opposing lineup, the ballpark and the weather into a probability distribution over runs, and from that into a fair moneyline, run line and total. Because baseball's primary market carries no spread, the model's output has to be a probability rather than a point estimate, and the price is where the whole opinion is expressed.

Why does the starting pitcher matter so much in MLB odds?

Because he sets the run environment for the innings he works, which is several times the influence of any single position player. The rotation is announced in advance, so the market can price it precisely and adjust when it changes. It is the largest forecastable input in the sport, and the market treats it accordingly.

How do park factors and weather affect MLB totals?

Stadiums differ in how readily batted balls become hits and home runs, expressed as an index with one hundred as neutral. Wind is the weather input that matters most, because wind blowing out carries fly balls over the fence and wind blowing in suppresses them, which lands directly on the home run component. Temperature has a smaller effect through air density. These inputs are measurable before the game, which is unusual among major sports.

Does MLB have a spread?

The primary market, the moneyline, does not. Baseball prices the winner directly. The run line is a separate market with a fixed handicap of one and a half runs, applied to the favourite, which is functionally a spread with a constant number. The two markets ask different questions about the same game and can disagree without either being wrong.

Can an MLB model beat the closing line?

Occasionally and modestly on the most liquid markets, and with somewhat more room in markets that require several inputs to be combined correctly. The long season means a small edge can compound, but it also means a small edge is difficult to verify from a single season of results. Scoring in point-in-time order against closing prices, across a large sample, is the only test that separates a real edge from a fitted one.

Sources

  1. What a sports betting model outputs | Rithmm, rithmm.com
  2. How to build an MLB betting model | Rithmm, rithmm.com
  3. Starting pitcher | Wikipedia, en.wikipedia.org
  4. Batting park factor | Wikipedia, en.wikipedia.org
  5. Sabermetrics | Wikipedia, en.wikipedia.org
  6. Ballpark factor (attempted) | Wikipedia, en.wikipedia.org
  7. Market rules | DraftKings (attempted), sportsbook.draftkings.com