Reliever CSR

Motivation

Why Relievers Break Every Standard Tool

Watch a one-run game end. The closer gets the final three outs, the save goes in the book, and the broadcast calls him the best arm in the bullpen. Two innings earlier, a different pitcher walked into a tie game with the bases loaded and one out, the genuinely harder assignment, escaped it without a run, and got a hold for his trouble, a statistic most fans cannot define. The save and the hold are bullpen-chart labels. They record the inning a manager chose, not the difficulty of the work or the quality of the pitching. Built on them, the entire popular understanding of relief value is sorting by job title.

Earned run average does no better, and in one specific way it does worse. ERA charges inherited runners to the pitcher who put them on base. When a reliever enters with two on and strands them both, the escape that defines his value leaves no trace on his ERA, and when he lets them score, the damage lands on someone else's line. The single most important thing a high-leverage reliever does, cleaning up the mess he was handed, is invisible to the number every casual evaluation leans on. The reliever conversation runs on statistics that cannot see the most valuable work being done.

Leverage Is the Point, and Leverage Is a Trap

Leverage Index already knows that a bases-loaded eighth in a tie game matters more than a clean ninth with a three-run lead. Any honest reliever framework has to put leverage at its center. The obvious move is to rank relievers on leverage-weighted value, win probability added, which credits the firefighter's escape and the closer's finish on a single scale. That move is a trap, for two reasons. First, leverage value does not repeat. A reliever's win probability added correlates with his own the following season at roughly one tenth, the noisiest signal in the entire dataset. Second, crediting a pitcher for the leverage he was handed credits the manager who handed it to him. Leverage is necessary, and leverage cannot be the rank.

The Orienting Question

How do you credit high-leverage relief work without rewarding mere deployment, and without ranking on a number that does not survive to next season? The framework answers by splitting leverage into its two honest jobs. Leverage sets eligibility: a reliever has to be trusted in high leverage to appear at all. Skill sets the rank: among the trusted, the order is decided by command and contact, the things that carry forward. The value a pitcher delivered is shown beside his name, credited and visible, but it does not move him up or down.

Methodology

The Trust Gate: gmLI ≥ 1.50

Reliever CSR is computed only for relievers whose average game-entry leverage, gmLI on FanGraphs, is at least 1.50, meaning they enter games in situations fifty percent more critical than league average. That threshold is the trust gate, and it does the work an innings cutoff does for starters. A mop-up arm who pitches in decided games is not failing the metric, he is outside its question. The gate is not arbitrary. Game-entry leverage is one of the most stable reliever signals there is, correlating with itself year over year at better than one half, so it selects a real and durable role rather than a one-week accident.

Component 1: K-BB%

K-BB% is strikeout rate minus walk rate, the cleanest public measure of command and stuff. A strikeout and a walk are settled between the pitcher and the hitter with no fielder involved, and the gap between them is among the most stable skills a pitcher owns. In the backtest it is the single most predictive reliever input, and it anchors the rating.

Component 2: Barrel% (Inverted)

Barrel% is the share of batted balls struck at the launch angle and exit velocity that produces extra-base damage, inverted so that suppressing hard contact scores positive. This is the contact-suppression core that gives the framework its name. It catches the arm who limits damage through weak contact rather than through swing-and-miss, a profile that command alone would miss.

Component 3: SIERA (Inverted)

SIERA is an earned run estimator that folds batted-ball regression into its calculation while staying defense-light, inverted so that a lower estimator scores higher. It adds a stable, run-prevention-oriented signal that neither command nor contact fully captures, and in the backtest the three-component rating forecasts next-year ERA better than any single input alone.

Why Win Probability Added Is Shown, Not Ranked

Every arm on the board carries its win probability added, displayed as the value it delivered, because that value is real and worth seeing. It simply does not forecast. A reliever's win probability added predicts his next-season value worse than his command does, worse than his contact quality does, worse even than his saves and holds. Putting it into the rating made the rating a worse predictor, so it stays beside the rating rather than inside it. Value delivered and value sustained are different questions, and the framework refuses to let the noisier one set the order. This is the one place Reliever CSR departs from the starter version of CSR, which does rank partly on contextual value, and the departure is principled: for starters that contextual term validated, and for relievers it did not.

Z-Score Standardization

Each component is converted to a z-score against the full reliever pool for the season, with Barrel% and SIERA negated so that lower is better. Standardizing puts command, contact, and the estimator on one scale, so a one-standard-deviation edge in any of them counts the same. Because the baseline is the full reliever pool, a Reliever CSR of 0.000 is an exactly average reliever, and the trusted arms that clear the gate tend to sit above it.

Formula

Worked Example: Mason Miller

Mason Miller leads all trusted relievers in 2026, and he passes every test at once. The computation runs against the 2026 reliever pool, which carries a mean K-BB% of 14.6 percent, a mean Barrel% of 6.6 percent, and a mean SIERA of 3.69.

Miller's command score is the most extreme value in the pool, a strikeout-minus-walk rate nearly three times the league average. His contact suppression is elite and his estimator is the best among trusted arms. No single component carries him, and the value is matched by the skill underneath it, which is exactly the case where every framework should agree.

The 2026 Reliever CSR Board

The twelve highest-rated trusted relievers through June 24, 2026. The rating column orders the board; the WPA column is the value each arm delivered, shown but not ranked. The order tracks skill rather than saves: a setup man with one save ranks third, while several closers with gaudy save totals sit lower because the skill underneath does not support the results.

Case Studies

The Firefighter Surfaces: Adrian Morejon

Adrian Morejon has one save. By the popular ledger he is a middle reliever, a name most fans outside San Diego would not place. Reliever CSR ranks him third among every trusted arm in baseball, ahead of established closers, because the skill is plainly there: a 25.6 percent command rate, a 1.9 percent barrel rate among the lowest in the sport, and a sub-2.00 estimator. He enters at a 1.84 average leverage, deep in the firefighting band, and holds thirteen leads doing it. The save column cannot see him. The rating can, because it never asked what inning he pitched, only how well.

The Regression Catch: Kevin Kelly and Riley O'Brien

The skill anchor protects the board from the arms a leverage-only metric would crown. Kevin Kelly entered games at higher average leverage this season than Mason Miller did, and on a leverage-weighted board he would rank near the top. His command sits at 15.9 percent, his estimator at 2.92, and his earned run average has run more than a run below what his contact quality supports. He ranks twentieth, flagged. Riley O'Brien carries nineteen saves and a 4.05 earned run average on a 15.4 percent command rate, and lands fourteenth. Both delivered real win probability value, which the board shows, and both carry profiles that say the results are on loan.

The Honest Down Year: Andres Munoz

The gate admits a struggling arm when the trust is real, and the rating does not hide the struggle. Andres Munoz enters at a 1.91 average leverage with a closer's pedigree and a 24.6 percent command rate, but a barrel rate above ten percent and a 5.08 earned run average drag his contact score down. He sits in the middle of the board with negative win probability value displayed beside his name. The framework neither buries him on one bad stretch nor pretends it did not happen.

Validation: The Backtest

Reliever CSR was tested against five complete seasons, 2021 through 2025. For each pair of consecutive seasons, the question was simple: how well does a number from year one predict results in year two, across every reliever who threw at least thirty innings in both seasons. The four transitions were pooled, roughly three hundred player-pairs, and association was measured with the Spearman rank correlation.

The Predictive Horse Race

The skill rating forecasts next-season earned run average better than every alternative tested, including the sophisticated incumbent, reliever WAR. The leverage statistics the sport actually uses finish last. For ERA a more negative number is better, since a stronger year-one number should mean a lower year-two ERA.

On next-season WAR, command alone edges the composite by a hair, and this paper states that plainly rather than hiding it. The composite earns its place anyway, because it wins outright on earned run average, the cleaner measure of run prevention, and because it is more robust across both outcomes than any single statistic, folding in contact and estimator signal that command by itself cannot see.

Does Each Signal Even Repeat?

Before asking whether a number predicts results, ask whether it predicts itself. A signal that does not survive to next season cannot anchor a rating.

Reading the Numbers

Because every z-score is measured against the full reliever pool, a Reliever CSR of 0.000 is a dead-average reliever, and the scale reads in plain tiers.

And the translation that gives each z-score a face. A plus-one is not an abstraction once you know it means a twenty-one percent command rate.

One detail keeps the reader honest. The 0.000 baseline is the average of all relievers, but the arms that clear the trust gate average roughly plus four tenths, because trust is earned before a pitcher ever reaches the board. A mid-pack rating on the board is still an above-average reliever across the league.

Limitations

• Reliever samples are small and volatile. Through late June these are thirty-inning seasons of the highest-variance arms in the sport, and the board will move week to week.
• There is no prestige award to validate against. Relief pitching has no honor with a leverage backbone, and the popular reliever awards reward saves, which would make any such test circular. The validation rests entirely on the out-of-sample predictive backtest.
• This is a value-and-sustainability tool, not a one-number oracle. The rating says who is good and likely to stay good; the displayed win probability says what a pitcher delivered. The framework does not claim the value figure forecasts, because the data says it does not.
• The gate depends on deployment. A genuinely excellent arm buried in low-leverage innings will not appear, by design. Reliever CSR measures trusted relievers, and a manager who misuses a good pitcher can keep him off the board.
• The components are park-light, not park-neutral. Command and contact travel well and the estimator folds in batted-ball context, but no public reliever framework fully removes environment.

Data Sources

• K-BB% and SIERA: FanGraphs Advanced Pitching leaderboard
• Barrel%: FanGraphs Statcast Pitching leaderboard (sourced from Baseball Savant / Hawk-Eye)
• WPA, gmLI, Shutdowns, Meltdowns: FanGraphs Win Probability leaderboard
• Reliever WAR (validation reference): FanGraphs reliever WAR
• Eligibility: FanGraphs reliever pool per season, with the gmLI ≥ 1.50 gate applied

Frequently Asked Questions