Bullpen Usage and the Reliever Prop Problem

Start with the uncomfortable version of the problem. Take the single most dominant arm in a bullpen, give him the best matchup on the slate, and then let the game go 9 to 1 by the fifth inning. He throws nothing. His strikeout rate, his velocity, his platoon splits, every measurable quality that made him interesting: all of it multiplies by zero, because a manager looked at the scoreboard and decided to save him for tomorrow.

That is not an edge case. It is the ordinary condition of relief pitching, and it is why relievers are the clearest illustration in any sport of opportunity dominating skill. In most prop markets, opportunity is uncertain within a range. Here it is frequently binary and decided by a third party during the game you are researching.

The first question is whether he pitches at all

Every reliever projection has to begin with an appearance probability, and that probability is rarely close to one. It is driven by three things, none of which is the pitcher's ability.

1. Recent workload. Whether he threw yesterday, whether he threw the day before that, and how many pitches were involved.
2. Role definition. Whether he occupies a specific late inning job or floats inside a committee that could use any of four arms in the same spot.
3. Expected game state. Whether the game is likely to be close enough, and in the right direction, for his usual role to come up.

Get the appearance probability approximately right and the rest of the projection becomes tractable. Skip it and everything downstream is a conditional number pretending to be an unconditional one. This is a stronger version of the exposure argument in the starting pitcher strikeout article. A starter's exposure varies between roughly 18 and 27 batters. A reliever's varies between zero and four or five, and the zero is common.

Back to back appearances and workload limits

Relief usage is governed by informal but real fatigue constraints. An arm that pitched the previous day is less available today; one that pitched the previous two days is often unavailable outright regardless of the game situation. Managers also watch a rolling window, so three appearances in four days tends to trigger a planned rest day even when nothing about today's game argues against using him.

None of this is published as a rule, and clubs differ in how strictly they apply it, which means workload research is inference from recent usage rather than lookup. But the inference is unusually stable. A reliever who has thrown on consecutive days is a materially worse candidate to appear today than the same reliever with two days of rest, and that difference swamps almost any matchup consideration you could put next to it. When a heavy usage stretch and an attractive matchup point in opposite directions, workload wins more often than researchers expect.

The same logic runs in reverse. A bullpen that has been idle for two days, or one that just watched its starter go deep twice in a row, arrives at today's game with more availability than usual, which raises appearance probability across the whole group and slightly lowers it for any one specific arm, because there are more alternatives to choose from. Availability is a team level condition before it is an individual one.

A defined role and a committee are different research objects

The single most useful thing you can learn about a reliever is whether he has a job. A pitcher with a defined late inning role appears in a predictable game state, which means his appearance probability can be estimated from the probability of that game state occurring. That is a hard number to get exactly right, but it is a real quantity you can reason about: how likely is this game to arrive at the situation in which he is the designated answer.

A committee removes that structure. If three arms are interchangeable for the same innings, the probability that the situation arrives is unchanged, but it now splits across the group in a way that depends on matchups, handedness, who threw yesterday, and preferences that are not visible from outside. The situation probability was already uncertain; the allocation probability is uncertain on top of it, and the two multiply. This is the same failure mode described in the article on role changes: when a role is undefined, the opportunity term of the projection stops being estimable, and no amount of skill measurement compensates.

Roles also move during a season without announcement. A pitcher who held a defined job in June may be sharing it in August, and the usage pattern usually shows the change before anyone describes it. Reading the last two weeks of appearances by inning and score state is a better guide to the current role than any depth chart.

Leverage and score state decide the usage

Bullpens are deployed by situation, and the situation that matters most is the margin. Managers reserve their best arms for close games and spend their least valuable arms when the outcome is close to settled. The result is a usage pattern that is essentially a function of the score, which makes game script the central input in reliever research rather than a contextual footnote.

The crucial and frequently missed point is that blowouts remove relievers in both directions. Researchers naturally think about a team falling behind and shelving its closer. The mirror case is just as common and just as destructive to a projection: a team leading by seven does not use its highest leverage arm either, because there is nothing to protect. So an implied total or a matchup that suggests a lopsided game reduces appearance probability for the top of both bullpens simultaneously, while raising it for the arms further down that most people are not researching.

Conditional twice over

Put the pieces together and a reliever prop is conditioned at two separate stages. First, does he pitch. Second, given that he pitches, how much: one batter as a specialist, one inning as a defined role, or two innings because the game went long and the bullpen ran short. Those are different exposure levels, and they are not equally likely for every pitcher.

Two stages of conditioning make the resulting distribution genuinely awkward. It is not a bell shape with a mean you can compare to a line. It has a large mass at zero, a cluster around the outcome of a typical single inning, and a thin tail for the long appearance. A line sitting anywhere near the middle of that shape is not asking what the average outcome is; it is asking about mass in specific places, which is the point made generally in the article on reading a distribution and which relievers demonstrate more starkly than almost any other market.

This shape is also why the same reliever at two different numbers is two entirely different questions. A very low line is mostly a question about whether he appears. A higher line is a question about whether he appears and then goes long, which is a compound event that almost never happens in a defined single inning role. Researching the pitcher instead of the posted offer will get this backwards regularly.

Void handling changes the product

Because zero appearance is a common outcome rather than an unusual one, the settlement rule for a pitcher who never enters the game does more work in reliever markets than almost anywhere else. If the offer voids when he does not pitch, you are buying a conditional product: his stat given that he appeared, with the appearance risk removed. If the offer settles as a loss in that case, you are buying the unconditional version, which is a strictly harsher product at the same number.

The gap between those two products is not marginal. It is roughly the entire appearance probability, which can be the largest single term in the whole projection. A researcher who prices the conditional version and then buys the unconditional one has made an error larger than any matchup edge could recover. The general treatment is in the void rules article; the specific instruction here is to confirm the rule on the platform before doing any other work, because the answer determines which quantity you are supposed to be estimating.

Thin markets, and the honest default

Reliever offers tend to appear on fewer platforms and at fewer numbers than starter offers, with less cross platform disagreement available to check any of them against. That thinness is not an invitation. As covered in the thin markets article, a number that few operators post is a number few operators have scrutinized, and the reason it is uncertain to them is often the same reason it is uncertain to you. On relievers the shared uncertainty has a specific name: nobody outside the clubhouse knows who is available today.

Which points to a plain default. Relievers are a pass unless two conditions hold together: the role is defined enough that you can say which situation brings him into the game, and the expected game script makes that situation reasonably likely. One without the other is not enough. A defined role in a game that projects lopsided produces the zero outcome. A likely close game with an undefined committee produces an appearance by somebody, which is not a projection of anybody.

• Check the last several days of usage before anything else, and treat consecutive day appearances as a strong availability signal.
• Establish whether the role is defined or shared, using recent appearances by inning and score rather than a listed depth chart.
• Estimate the probability the game reaches his situation, remembering that a blowout in either direction removes him.
• Read the projection as a distribution with mass at zero, not as an average.
• Confirm the void rule, because it decides whether you are pricing a conditional or an unconditional product.

Slateline simulates each game rather than adjusting a season average, so relief exposure emerges from the same simulated games that produce the score, and the resulting distributions carry the zero mass instead of smoothing it away. Where the inputs do not support a projection, the output is a pass rather than a thin number with a grade attached. Our graded record and calibration are open in the Model Room, the MLB research library sits at the MLB props pages, and the current offers are on the Signal Board. Relief pitching is the part of baseball where the honest research answer is most often that the question belongs to the manager.