MLB Player Props Guide: How Batter Props Actually Work

Every batter prop has three components: the player, the stat, and the line. For example:

The line (1.5 in this case) is set by the sportsbook based on the player's historical production, tonight's matchup, and the amount of money already bet on each side. This is where the analytical work begins: is the sportsbook's line accurate, or does the data suggest a different probability?

Three structural reasons why player props offer better opportunities than moneyline or totals:

The hits market is the most straightforward batter prop. The two standard lines are O/U 0.5 (does the batter record at least one hit?) and O/U 1.5 (does the batter record two or more hits?).

The O/U 0.5 line is where sportsbooks make the most money on batter props. The public loves taking “over 0.5 hits” on star hitters because it feels like a lock — of course Juan Soto is going to get at least one hit. But the book prices this at -180 to -250, which implies 64-71% probability. Since the actual rate is roughly 67% for most regulars, the edge is paper-thin and often negative after vig.

The analytical action is at O/U 1.5. This is where the base rate is close enough to 50/50 that small differences in a player's true rate create meaningful edges. A player averaging 1.0 hits per game has roughly a 26% chance of going over 1.5. A player averaging 1.4 hits has roughly a 41% chance. That 15-point spread is where prediction precision becomes profitable.

Runs scored is a more volatile stat than hits because it depends not just on the batter's performance but on what happens around them. A batter can go 0-for-4 and still score a run (walk + advanced on hits behind them). Conversely, a batter who goes 3-for-4 might score zero runs if nobody behind them drives them in.

The standard line is O/U 0.5 runs. The league-wide base rate for scoring at least one run is approximately 38-42% for lineup regulars, varying significantly by lineup position. Leadoff hitters and #2-3 hole batters score more frequently because they get more plate appearances and bat in front of the best hitters in the lineup.

Because runs depend on context (lineup protection, team offensive quality, game flow), models that encode team-level offensive features alongside individual player stats have a significant advantage over models that only look at the batter in isolation.

RBIs are the most context-dependent batter stat in baseball. A player cannot drive in a run if nobody is on base when they bat. This makes RBI props heavily influenced by the quality of the hitters batting ahead of them — specifically, how often those hitters reach base.

The standard line is O/U 0.5 RBIs. Top-of-order hitters (1-2 spot) typically have lower RBI rates despite high overall production because they bat with fewer runners on base. Middle-of-order hitters (3-5 spot) have higher RBI opportunities by design — they bat behind the guys who get on base.

For modeling purposes, the key feature is opportunity: how many runners are likely to be on base when this batter comes to the plate? This is a function of the OBP of the batters ahead of them in the lineup, the number of plate appearances the batter is expected to get, and the overall offensive environment of the game (park factor, opposing pitching quality).

HRR is the combined stat line that sums a player's hits, runs scored, and RBIs for the game. It is the highest-volume batter prop market in MLB — nearly every sportsbook offers HRR lines, and the combined nature of the stat creates more interesting odds than any single component.

The standard line is O/U 1.5. For a league-average regular, the probability of accumulating 2+ HRR in a game is roughly 50-55%, which means the sportsbook prices it near even money. This near-coinflip base rate is what makes HRR the most analytically interesting batter prop — small model edges translate directly to actionable bets because the juice is lower than on heavily skewed lines like O/U 0.5 hits.

HRR also has a feature that makes it mathematically distinct from the other three props: overdispersion. Because it combines three different stats — some of which are correlated (a hit that also scores a run counts twice) — the variance of HRR is roughly twice its mean. This has major implications for how you model the probabilities, which we cover in detail in section 6.

An important note on HRR odds: most sportsbooks do not publish HRR lines directly through odds APIs. HRR is typically available on the sportsbook app or website, but the data feeds that odds aggregators and models pull from often omit it. This means HRR lines need to be constructed synthetically — by identifying players who appear in hits, runs, or RBI odds and computing the combined probability. This extra step creates an information barrier that keeps some analytical competition out of the market.

On standard -110/-110 sides and totals bets, the sportsbook's vig (margin) is about 4.8%. On player props, it is significantly higher. Here is what typical batter prop juice looks like:

Notice the pattern: the more “obvious” the bet feels, the more juice the book embeds. O/U 0.5 hits feels easy — everyone takes the over on their favorite player — so the book charges a premium. HRR at 1.5 is closer to a coinflip and draws more balanced action, so the vig is lower. Lower vig means the breakeven accuracy threshold is lower, which means your model edge goes further.

Sportsbooks generate hundreds of player prop lines daily. They cannot spend 30 minutes hand-tuning each one. Instead, they use baseline models calibrated to season averages, then adjust based on matchup, park, and sharp action. The places where this process breaks down:

Heavily juiced lines on star players are the biggest trap in batter props. When a book offers Shohei Ohtani Over 0.5 hits at -250, they are implying 71.4% probability. His actual hit rate might be 70.5%. After the vig, you are paying a 71.4% price for a 70.5% outcome — a guaranteed losing proposition over volume.

The lesson: never evaluate a prop bet by how likely it is to win. Evaluate it by whether the price is right. A 55% prop at -110 is a better bet than a 70% prop at -250, because the first one has positive expected value and the second one doesn't.

A classifier trained to predict “Over 1.5 hits: yes or no” learns a single threshold. If the sportsbook moves the line to 0.5 or 2.5, you need a completely different model. This approach is brittle and wasteful.

A regressor trained to predict “how many hits will this player get tonight?” outputs a continuous value — for example, 1.35 hits. That single number can be evaluated against any line. One model, every possible threshold.

More importantly, the regressor output contains magnitude information that a classifier throws away. A classifier might say “62% chance of going over 1.5 hits” — but it cannot tell you whether the underlying prediction is 1.6 (barely over the line) or 2.3 (comfortably over). The regressor can, and that magnitude is what determines the actual probability.

Baseball stats are counts — 0, 1, 2, 3 hits, not 1.7 hits. The Poisson distribution is the natural probability model for count data. It takes a single parameter, lambda, which represents the expected average count. From lambda, the Poisson CDF calculates the exact probability of any count occurring.

This is the power of the regressor + CDF approach: one prediction (1.35 hits) produces accurate probabilities at every line. You do not need separate models for O/U 0.5, O/U 1.5, and O/U 2.5 — the Poisson distribution handles all of them from a single output.

A model that predicts exactly 2.0 hits might make the Poisson CDF output an artificially confident probability at the 1.5 line. But the model's prediction has its own uncertainty — no model is perfectly accurate.

The solution is a standard deviation floor. The system calculates the model's actual prediction error on historical validation data and enforces a minimum variance that prevents any single prediction from producing probabilities that real-world variance would not support. In practice, this means a hits prediction uses a minimum standard deviation of 0.7, runs uses 0.5, and RBIs uses 0.5. These floors are calibrated from thousands of historical predictions and prevent the model from ever claiming 90%+ confidence on a proposition that actually hits at 75%.

The largest feature group captures the player's recent performance across multiple time windows. Using multiple windows simultaneously is critical — a 10-game window captures current form, while a 3-game or 5-game window captures micro-streaks.

A batter facing a top-5 pitching staff has a different stat ceiling than one facing the worst pitching staff in baseball. These features encode the quality of the entire opposing pitching unit — not just the starter.

When a batter has faced tonight's starting pitcher before, that history is encoded as features: batting average, home run rate, strikeout rate, RBI rate, OPS, and total at-bats in the matchup. A confidence weight scales these features based on sample size — a 5-AB sample gets minimal weight, while a 30+ AB history gets treated as a real signal.

BvP data is one of the most overrated features in casual baseball analysis and one of the most useful features when properly weighted. The key is not using it in isolation — it is one signal among 72, and the model learns exactly how much to trust it based on the sample size and context.

Beyond the team-level pitching quality, the model encodes the specific starter the batter will face tonight. These 12 features are computed from the pitcher's last 5 starts:

These opposing pitcher features account for approximately 16% of total feature importance in the model — a substantial signal that was added based on the intuition that who is pitching tonight matters as much as how the batter has been hitting.

Raw features in isolation miss the most important dynamics. A batter with a 1.2 hit average facing a team that allows 9.5 hits per game is a different situation than the same batter facing a team allowing 7.0 hits per game. Interaction features capture these compound effects:

These three features let the model distinguish between a good hitter in a bad matchup and a good hitter in a great matchup — context that individual features alone cannot express.

The Poisson distribution has one core assumption: the variance of the data equals the mean. For individual stats like hits or runs, this holds reasonably well. A player averaging 1.0 hits per game has a variance close to 1.0.

HRR violates this assumption. When you add hits + runs + RBIs together, the variance of the sum is roughly 2x the mean. This happens because the three stats are correlated — a hit that also scores a run counts in both the “hits” and “runs” buckets, and an RBI single counts in both “hits” and “RBIs.” These correlations inflate the variance beyond what Poisson expects.

If you naively apply the Poisson CDF to HRR predictions, the distribution underestimates the probability of extreme outcomes (0 HRR and 4+ HRR) and overestimates the probability of moderate outcomes. At the critical O/U 1.5 line, Poisson overestimates the over probability by approximately 20%.

In practical terms: if the true probability of Over 1.5 HRR is 52%, a Poisson-based model might report 62%. That 10-point overestimate turns an edge into a mirage — you think you have a strong play when the true probability is barely above breakeven.

The solution is to use different probability distributions at different lines, selected based on backtested calibration:

The Negative Binomial distribution generalizes the Poisson by adding an extra parameter (r) that controls how much extra variance the model allows. When r is large, the Negative Binomial converges to Poisson. When r is small (like 1.83), it allows the fat tails that HRR's overdispersion creates — correctly modeling the probability of 0 HRR games and 5+ HRR games that Poisson underestimates.

At the 1.5 line specifically, the Normal distribution with a calibrated standard deviation of 5.7 produces the lowest mean absolute error in backtesting. This is an empirical result — the math says “use the distribution that produces the most accurate probabilities at each line,” not “use one distribution everywhere.”

The starting point for every prediction is the projected value — the model's best estimate of the actual stat the player will produce. For example, “Aaron Judge: 1.45 hits.” This is the lambda that feeds into the probability distribution.

The projected value is useful on its own for context — it tells you whether the model thinks the player is likely to clear a line comfortably (1.45 vs. a 0.5 line) or just barely (1.45 vs. a 1.5 line). The further the projection is from the line, the more confident the probability estimate.

From a single projected value, the model computes over/under probabilities at every standard line. Here is what a complete hits output looks like for a hypothetical player:

The probabilities at each line tell you where the edge is, not just whether one exists. In this example, the model sees the strongest signal at the 1.5 line (under at 60.8%). If the sportsbook is offering Under 1.5 at -110, the implied breakeven is 52.4% — and the model says the true probability is 60.8%. That 8.4-point gap is the edge.

Not every line is equally useful. The primary line is the standard sportsbook default — usually O/U 0.5 for hits, runs, and RBIs, and O/U 1.5 for HRR. This is the line most bettors see and bet on. The best line is whichever threshold produces the highest model probability in either direction.

Sometimes these diverge. The model might show modest confidence at the primary 0.5 line but high confidence at the 1.5 line (or vice versa). Knowing which line to target — rather than always defaulting to the sportsbook's primary — is a meaningful edge in itself.

The 10-game sparkline is included for context — it shows whether the player has been hot, cold, or consistent. But it is not an independent signal. The model already encodes this information through its rolling stats, trend features, and streak features. The sparkline exists so you can quickly sanity-check the model's projection against what you know about the player's recent form.

If the model says Over 1.5 HRR at 65% but the sparkline shows 8 out of 10 games below 2 HRR, the model might be seeing something the raw game log is not showing — an improving matchup, a weak opposing pitcher, or a favorable park. Trust the probability, use the sparkline for context.

Not every prediction above 51% is worth acting on. The breakeven threshold at -110 odds is 52.4%. At -130 (common prop juice), it climbs to 56.5%. The model needs to clear the breakeven plus a margin for model error before a prediction becomes a genuine play.

In practice, different confidence thresholds produce very different accuracy-vs-volume tradeoffs. A higher threshold means fewer plays but higher accuracy. A lower threshold means more volume but thinner edge per play. The right threshold depends on your risk tolerance and bankroll strategy — and it differs by market because the base rates and juice structures are different for each stat.

The biggest advantage of the regressor + CDF approach is the ability to evaluate every line, not just the sportsbook default. A player might show no edge at the primary O/U 0.5 hits line but a significant edge at O/U 1.5. Or the model might find that Under 2.5 HRR is a stronger play than the conventional Over 1.5.

This flexibility is something that casual bettors — who only look at the default line the app shows them — systematically miss. A model that evaluates all lines simultaneously finds 2-3x more actionable plays per day than one that only checks the primary.