What Is HRR in Baseball? Hits + Runs + RBIs Explained

Imagine Mookie Betts in tonight's game. He goes 2-for-4 with a single, a home run, scores 2 runs, and drives in 3 RBIs. His HRR for the game:

Notice that a single home run with a runner on base counts in all three categories — 1 hit, 1 run, 1 RBI for the runner, plus 1 RBI for the home run itself. This correlation between the three stats is what makes HRR mathematically distinct from any single component, and it is the basis for the math problem we cover in section 3.

For an average MLB lineup regular hitting in the 1-6 spots, the probability of accumulating 2 or more HRR in a game sits between 50% and 55%. League-wide, the average across all qualified hitters is approximately 52% over 1.5. This is essentially a coinflip — and books love coinflips because they let them charge a thin vig and still draw balanced action.

The 1.5 line works specifically because it isolates lineup regulars at near-50/50 — exactly the population that gets the most prop volume. The same line on a 7-hole hitter would be heavily skewed under, and the same line on Aaron Judge would be heavily skewed over. Books adjust the price (juice) on individual players to account for these tier differences, but the line itself stays at 1.5 because that is what casual bettors expect to see.

Because HRR 1.5 draws balanced action, the typical juice is -115/-105 or -110/-110 — roughly 4-5% vig. Compare that to:

The breakeven win rate at -115 is 53.5%. At -250, breakeven is 71.4%. Lower juice means lower required accuracy to be profitable, which means analytical models with even modest edges can extract value at HRR 1.5 in a way they cannot at heavily juiced individual-stat props.

The Poisson distribution is the natural starting point for modeling count data like baseball stats. It takes one parameter — lambda, the expected count — and produces a probability for every possible outcome (0, 1, 2, 3, etc.). The catch is that Poisson assumes variance equals mean. For individual stats like hits or runs, this assumption holds reasonably well. For HRR, it does not.

HRR is overdispersed. Its variance is roughly twice its mean. This happens because the three component stats are correlated — an RBI single counts in both “hits” and “RBIs,” a home run with a runner on counts in “hits,” “runs,” and twice in “RBIs.” Those correlations inflate the variance of the sum well beyond what Poisson expects.

If you naively apply the Poisson CDF to an HRR projection, the distribution underestimates the probability of extreme outcomes (zero HRR games and 5+ HRR games) and overestimates the probability of moderate outcomes clustered near the mean. At the critical 1.5 line specifically, 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 reports 62%. That ten-point overestimate turns an actual coinflip into an apparent strong play. You bet, lose half the time at -110, and slowly bleed the bankroll. Most public HRR models that use Poisson naively produce confident-looking outputs that have no real predictive power at the threshold that actually matters.

Before May 2, 2026, Prediction Engine's HRR model used a Normal distribution at the 1.5 line with a calibrated standard deviation of 5.7. This was an empirical fit — backtesting on thousands of historical games found that Normal at scale=5.7 produced lower mean absolute error than naive Poisson at the 1.5 line.

The problem with Normal at 5.7 was that the standard deviation was too narrow at the threshold. The Normal distribution has thin tails relative to the actual empirical distribution of HRR. At 1.5, the Normal CDF correctly avoided the Poisson overestimate, but it produced over-probabilities that were still slightly too confident — and the calibration drifted as season-on-season HRR variance shifted.

The Negative Binomial distribution generalizes Poisson with an extra parameter, r, that controls how much extra variance the model allows. When r is large, Negative Binomial converges to Poisson. When r is small, it allows the fat tails that overdispersed data require.

For HRR, backtested calibration on thousands of historical games converged on r = 1.83. At this dispersion parameter, the Negative Binomial CDF correctly models both the elevated zero-HRR rate (about 24% league-wide) and the long right tail (4+, 5+, 6+ HRR games). On May 2, 2026, the production model switched the 1.5-line probability calculation from Normal(scale=5.7) to NegBin(r=1.83). The 0.5, 2.5, and 3.5 lines were already using NegBin since the original HRR upgrade.

The result is a unified probability framework for HRR at every line, calibrated against actual baseball outcomes rather than a Poisson assumption that breaks on combined stats. For most picks, the new NegBin probabilities are 2-4 percentage points lower at the 1.5 line than the old Normal output — closer to ground truth, which means fewer false-confidence plays and tighter alignment with where the real edge lives.

When sportsbooks know the public will hammer one side of a line, they widen the vig to compensate. The classic example is Over 0.5 hits on a star player. Most casual bettors take it on instinct — “of course Juan Soto gets a hit” — and books price it accordingly.

The HRR market sits at near-coinflip pricing, while the single-stat props charge between 56% and 69% breakeven on the over. To beat -220 on Hits 0.5, your model has to be extraordinarily right. To beat -115 on HRR 1.5, you only need to be modestly above the implied 53.5%. That is the structural arbitrage.

HRR is also the only batter prop that captures the full picture of an offensive contribution. Consider three different game scripts for the same player:

Notice that in Script B, the player went 0-for-3 — Over 0.5 Hits would be a loss, but Over 1.5 HRR is a win because walks and runs scored are still captured. HRR reflects multiple paths to value: hot hitting, lineup protection (runs from teammates' hits), or driving in runs without necessarily reaching base often. Single-stat props miss two of those three paths every night.

The HRR model uses 82 engineered features (upgraded from 72 with the addition of nine Baseball Savant Statcast features on April 29, 2026). For Judge tonight, the most influential inputs:

Feeding these features into the trained XGBoost regressor produces a single point estimate:

The point projection of 2.7 feeds into the Negative Binomial CDF with r=1.83. The CDF computes the probability mass for every possible outcome (0, 1, 2, 3, ...) and sums the over-1.5 region:

At -115 (53.5% breakeven), a true 64.4% probability is a 10.9-percentage-point edge. That converts to a positive expected value of roughly 23 cents on the dollar per unit bet at standard juice — a strong play, but not a lock. Judge will still go under 1.5 HRR in roughly 36% of these projected scenarios, which is the irreducible variance of the game.

Under the pre-2026-05-02 Normal(scale=5.7) approximation, this same projection of 2.7 would have produced an over-1.5 probability of approximately 67-68% — about 3 percentage points higher. That sounds like a small change, but compounded across hundreds of HRR plays per season, it is the difference between a model that confidently reports 65%+ on plays that hit at 62% and a model whose stated probabilities match actual outcomes. The NegBin migration tightens calibration, which directly improves Kelly sizing and unit allocation.

HRR projections tend to cluster — at any given slate, the model produces a tier of 15-20 hitters with over-1.5 probabilities between 56% and 62%. Most of those plays are profitable in expectation but small. The real edges come from confluent picks where three independent signals align:

When all three align on a top-of-order or middle-of-order bat, the model produces 65-72% over-1.5 probabilities at -110 to -115 juice. These are the picks where HRR becomes a genuine cash-cow market. They appear several times per week during the season — not every day, but consistently enough to build a portfolio around.

On April 30, 2026, Prediction Engine deployed a three-pronged screen on top of the raw model output to surface the highest-conviction HRR plays. The filter requires:

Picks that clear all three flags are tagged as “elite” in the model output. On a typical day this surface 4-8 plays from a slate of 200+ projections. Solving the “everyone bunched at the top” problem makes the model's output actionable rather than overwhelming.

HRR singles are the most reliable expected value play. Parlaying multiple HRR overs sounds attractive — three legs at +100 each pays 7-to-1 — but the leg-by-leg correlation is not as low as books price it. When the same weather front sweeps multiple parks, multiple legs covary; when an unexpected bullpen game emerges, the entire leg slate shifts. Independent confluent singles, sized via fractional Kelly or flat units, produce smoother bankroll growth than HRR parlays. This is the documented #1 edge in the user's personal betting data — singles dominate parlays on long-run ROI.

A player who is 9-for-his-last-15 with three multi-HRR games feels like an automatic over. The trap is ignoring the matchup. A red-hot bat facing a Cy Young-caliber starter (Skubal, Skenes, Sale) is not a good HRR bet — the pitcher quality dominates the recent batter form. The model assigns roughly 16% feature importance to opposing pitcher individual stats specifically because pitcher quality is one of the strongest signals on any given night.

The correct framing: hot streaks help, but they only matter when paired with a favorable matchup. A hot bat against a weak starter is the elite play. A hot bat against an ace is closer to a coinflip with juice — typically negative EV at -115.

Season ERA is a lagging indicator. A pitcher with a 3.50 season ERA who has allowed 12 earned runs across his last two starts is not a 3.50 ERA pitcher tonight — he is a struggling pitcher whose season number has not caught up to recent form yet. Books often price props using season-level pitcher stats because that is what their templates use, which means recent-form deltas are systematically underweighted.

Always cross-check the opposing starter's last 5 starts before betting an HRR over. If his last 5 ERA is 2+ points higher than his season ERA, the matchup is more exploitable than the line suggests. The model captures this through rolling pitcher features, but a quick manual sanity check on the boxscore page never hurts.

A player who has gone 0-fer in two consecutive games “feels due” for a big game. There is a small statistical reality to this — players coming off multiple zero-HRR games do hit at slightly elevated rates the next day, which the model captures through the “zero-HRR games in last 5” feature. But the effect is roughly 2-3 percentage points, not enough to overcome a bad matchup.

The trap is taking the over on a struggling player against a strong pitcher purely because they are due. The matchup still dominates. Only bet the bounce-back angle when the player is in a favorable matchup AND coming off multiple zeros — that double-edge is real. Without the matchup, the bounce-back signal is noise.

Books occasionally post HRR lines on hitters who are not confirmed in tonight's lineup. Limits are typically lower on these props because the books themselves are uncertain whether the player will start. Even if the line looks juicy, the variance from a possible bench scratch is enormous — a player who pinch-hits once for a single AB has almost no path to over 1.5 HRR.

Always verify the lineup is confirmed before betting HRR. Prediction Engine's prop pages only display lines for confirmed-lineup hitters specifically to avoid this trap. Roughly 19% of model defaults on weather and probable starters happen because cron timing precedes lineup hydration — the late-day refresh is what delivers actionable HRR projections, not the morning run.