The lefty-righty note is the first line in too many MLB prop writeups and the last line in too many models. “Good side of the platoon” is not a price. It is one branch in a sequence that can change when the starter leaves, the lineup turns over, or the manager reaches for a different pitch profile.
Start with the actual matchup, not the label
Handedness matters because pitchers and hitters do not interact symmetrically, but the mechanism sits in the repertoire. A hitter may handle velocity from one side and struggle with a breaking ball that moves away. Another hitter may show a broad split because of lineup usage or a small cluster of plate appearances rather than a stable skill.
The model should represent pitcher handedness, pitch types, locations, velocity bands, and the hitter’s responses to those offerings. A headline OPS split can be a useful summary, but it hides which pitches produced the difference and whether those pitches are likely today.
The feature work can be explored in the workshop, but the model card should state which pitch and handedness inputs are available before the prop is offered.
Project exposure before projecting outcome
A batter prop depends on opportunities. The confirmed lineup determines batting order, teammates, and whether the player is starting. The opposing starter’s expected workload shapes how many plate appearances occur against the advertised handedness. Game state and substitution risk can change the rest.
Do not hardcode the starter across every projected plate appearance. Build an exposure distribution: likely appearances against the starter, likely appearances against relief groups, and a branch for early removal or lineup change. Then apply outcome probabilities inside those branches.
This is especially important for full-game props. A starter matchup can be real and still govern only part of the ticket.
The bullpen is a sequence, not an average
Team bullpen splits flatten the decisions that matter. Availability changes with recent workload. Leverage changes which reliever is likely to appear. A manager may prefer handedness, pitch shape, or a designated inning role. The next batter in the order can influence the choice as much as the current one.
A useful bullpen state includes available relievers, throwing hand, repertoire, recent workload, role, and likely entry conditions. Missing data should widen the range of possible outcomes. It should not be replaced with an invented “typical” reliever.
When the bullpen feed is stale, the correct boundary result is unavailable bullpen state; cure is refresh or pass. A full-game prop priced from an old availability assumption is not a controlled wager.
Shrink noisy splits instead of worshipping them
Observed platoon performance mixes talent, opponents, parks, sequencing, and luck. The smaller the sample, the more aggressively it should be pulled toward a broader prior. The prior can reflect league, position, pitch type, or player history, but it must be chosen before the current result is admired. The calibration math behind that pull is in our Brier Score for Bettors explainer.
Recent changes deserve investigation, not instant belief. A new stance, pitch mix, injury, or role can make older data less representative. It can also create a temporary run that looks structural. The model should record the change and test it out of sample.
For live prop review, the player props page should surface the source timestamp and lineup status rather than presenting a split as timeless.
Keep park and game context in separate lanes
Park, weather, expected game length, and lineup quality affect opportunity and contact. They should enter the model without being mistaken for platoon talent. A hitter’s raw split compiled across parks is not directly portable to a single environment.
Separate the layers: pitcher-hitter interaction, opportunity, park context, and market price. That decomposition makes double counting visible. If park-adjusted outcome data is used, adding the full park adjustment again can inflate the same signal twice.
Price the exact prop contract
Hits, total bases, home runs, and combined counting stats answer different questions. They use different event definitions and respond differently to opportunity. A model trained for one target should not be copied onto another because both concern the same hitter.
Read the settlement rule, including participation and postponement handling. Match the probability to the offered line and price. The selections on the MLB picks board are useful only when the market definition and model target remain aligned.
Make the failure states visible
The model should return a typed reason when it declines: lineup unconfirmed, starter changed, bullpen unavailable, target unsupported, price stale, or uncertainty too wide. Each reason needs a cure. Refreshing a lineup can cure missing confirmation. Nothing cures an unsupported target except a model built for it.
Performance reporting must be equally plain. Show the ATS wins and losses, win rate, named window, and graded sample size. Keep the
The platoon edge is not “lefty versus righty.” It is a priced chain of expected encounters. Model the chain, or admit the ticket contains matchups you never projected.
Breakeven win rate at recorded American prices
Breakeven probability is calculated only from American prices that were actually captured in the odds-history table.
