How MLB Projection Models Handle Pitcher Fatigue

A projected pitching performance is only as good as its treatment of fatigue. Every model that puts a number on a starter's next outing has to answer a simple question — how tired is he, and by how much does that shift what he is likely to do? The answer, it turns out, is a stack of overlapping signals that each capture a different slice of the same phenomenon.

Fatigue in baseball is a strange concept to model. It is real — a body of research spanning MLB team analytics groups, Baseball Prospectus, FanGraphs contributors, and biomechanics journals confirms that pitchers throw slower, spin the ball less efficiently, and give up harder contact as workload accumulates^[1][2]. But it is also invisible in the traditional box score, and the signals that reveal it operate on at least three different timescales at once. A rigorous projection has to handle all three, weight them intelligently, and translate the resulting fatigue estimate into concrete adjustments to expected performance.

This piece walks through the pieces a modern projection framework has to account for. It is written for readers who want to understand the mechanics of the models they see quoted — not to serve as a betting or wagering recommendation. It covers, in order: what "fatigue" actually is inside a model, the three timescales of the problem, the specific Statcast signals that carry the most weight in current research, and the honest limitations of every method described.

What "fatigue" means inside a projection framework

A projection model is fundamentally a conditional expectation. It says: given everything I know about this pitcher, this opponent, this park, this weather, and this game state, what is the distribution of outcomes I expect? Fatigue enters that expression as a set of adjustments to the pitcher's baseline talent. It does not replace the talent estimate. It modulates it.

The clearest way to think about it: a pitcher has an underlying "true talent" level — the version of him that would show up if he were perfectly rested, in his usual role, with his mechanics at their normal baseline. Fatigue is the estimate of how far below that baseline he is likely to be right now. A well-built projection treats fatigue as an offset to expected wOBA-allowed, strikeout rate, and walk rate, not as a binary "tired or not" flag^[3].

This framing matters because it separates two kinds of error a model can make. The first is talent misestimation — thinking a pitcher is better or worse than he actually is. The second is state misestimation — correctly identifying his talent level but incorrectly adjusting for his current condition. These require different signals and different treatments.

Three timescales of fatigue

Every projection framework worth its salt handles fatigue at three timescales simultaneously, because each captures a different physical process.

Within-game fatigue: pitch count and times through the order

The oldest fatigue signal is also the simplest: as a starter throws more pitches in a game, he gets progressively worse. Baseball Prospectus's Pitcher Abuse Points framework, which dates to the early 2000s, formalized the observation that fatigue starts to accumulate meaningfully around 100 pitches and that each additional pitch beyond that point causes disproportionately more damage than the pitch before^[2]. That specific threshold has been refined by more recent research, but the core finding holds — within-game workload degrades performance in a roughly monotonic way.

The Times Through the Order Penalty, popularized by Mitchel Lichtman and colleagues, provides a second within-game signal. The observed pattern is that a pitcher's expected wOBA-allowed rises meaningfully each time he sees a batting order — the second time through is worse than the first, the third worse than the second^[4]. A 2023 Bayesian re-analysis by Ryan Brill at Wharton put the average per-TTO increase at roughly 13 wOBA points, with the important caveat that the increase is smooth and continuous rather than discontinuous at any single cutoff^[1]. That finding matters for how a model implements the adjustment: a step function that suddenly downgrades a pitcher when he enters the third TTO is technically wrong, even though it may be operationally convenient.

The practical implementation inside a projection typically looks something like this:

A common mistake in publicly quoted models is treating these three signals as independent when they are not. Pitch count, TTO, and innings pitched all move together over the course of an outing. Adding penalties for each without adjusting for their correlation double- or triple-counts the same underlying fatigue, which is one of the reasons some published projections seem to punish workhorse starters more than the data actually supports.

Between-start fatigue: rest days and recent workload

The second timescale is between-start recovery. A starter typically pitches on four days of rest in a five-man rotation. Research on how deviations from that baseline affect performance is mixed but converges on a few points:

• Three days of rest (short rest) is associated with measurable degradation — historical data suggests ERA increases of roughly 0.3 to 0.5 runs, fastball velocity drops of 0.3 to 0.7 mph, and elevated walk rates^[3].
• Five days of rest is often, though not always, associated with slight performance gains over the standard four — a finding that seems counterintuitive until one considers that the extra day allows recovery without disrupting mechanical rhythm^[5].
• Extended rest (seven or more days) shows mixed effects. Some pitchers benefit; others lose feel for their release point and command^[6].

A 2017 FanGraphs Community analysis found that while there are statistically significant differences between specific rest-day counts, the coarser buckets of "short," "normal," and "extended" rest do not differ meaningfully as groups — the effect is granular and pitcher-specific^[6]. A modern projection therefore has to model rest at the individual pitcher level, not as a blanket league-wide adjustment.

The relief pitcher version of this signal is more direct. High-leverage relievers who have thrown 35 or more pitches across the previous three calendar days are meaningfully impaired; 50 or more across three days is a common threshold for treating a reliever as effectively unavailable^[7]. Those thresholds are not league rules — they are working numbers developed from usage-pattern observation — but they reflect the underlying biological constraint that a major-league arm needs roughly 24 hours to recover from a normal one-inning outing.

Season-long accumulation: workload trajectory

The third timescale is the slowest and the least well-modeled. Over the course of a 162-game season, some starters accumulate workload faster than others — high pitch counts, deep starts, and the cumulative wear of six months of maximum-effort throwing. Research by Bradbury and Forman among others has documented that season-long workload correlates with elevated ERA and home run rates late in the year^[8]. The observation that starters exceeding 100 pitches in three or more of their last five starts tend to see performance decline in August and September is a rule of thumb that appears in multiple analytics traditions^[3].

The 2025 season is notable here: only 11.6% of games started that year featured a starting pitcher throwing 100 or more pitches, compared to 38.3% in 2015^[9]. Teams have adjusted their usage patterns, in part, to preempt exactly the fatigue accumulation that models were catching after the fact. A projection framework calibrated on 2015 usage patterns will systematically over-fatigue modern pitchers if it does not update its baseline expectations.

The Statcast signals: velocity and spin-rate decay

The historical fatigue signals — pitch count, rest days, seasonal workload — are indirect. They correlate with fatigue because heavier workloads make fatigue more likely, but they do not measure the fatigued state itself. The Statcast era has added a set of direct measurements that observe fatigue as it happens.

Velocity decay

Fastball velocity is the most-cited real-time fatigue indicator. Most starters show a characteristic within-game velocity curve — peak in the first or second inning, plateau through the middle innings, and a decline of one to two miles per hour by the 85-to-100 pitch mark^[3]. The shape of that curve varies by pitcher: power arms with high peak velocities tend to show steeper declines, while soft-tossing control pitchers may hold velocity longer but drop off more suddenly when fatigue does hit.

A projection model can learn a pitcher-specific velocity decay function from historical Statcast data, then compare the pitcher's current in-game velocity against that expected curve. When the observed velocity drops below expectation faster than the historical baseline, the model infers above-normal fatigue and adjusts downward for the remaining plate appearances. When the observed velocity holds above expectation, the model can hold or even nudge upward the remaining projection.

Spin rate decay

A more recent thread of research suggests that spin rate, not velocity, may actually be the more sensitive fatigue marker. A 2025 Statcast-based analysis by Sirsi Bhargav found that spin rate showed the clearest correlation with fatigue events across a range of pitcher clusters, while velocity drift within games was often too small and too noisy to be a reliable predictor^[10]. The proposed mechanism is that forearm muscle fatigue reduces grip strength, which reduces spin rate — a physiological chain distinct from the general arm fatigue that shows up in velocity.

This finding runs contrary to conventional expectation, but it fits a growing body of biomechanics research showing that different fatigue components (arm musculature, forearm musculature, mechanics maintenance) manifest in different measurable outputs^[11]. A model that tracks both velocity and spin-rate decay curves has a more complete picture of within-game fatigue progression than one that tracks either alone.

Release-point drift

A third Statcast signal, less commonly discussed publicly but increasingly used in team analytics groups, is release-point drift. When a pitcher fatigues, his release point (measured as the three-dimensional coordinate where the ball leaves his hand) tends to drift from its early-game baseline^[12]. The drift is small — typically less than an inch — but the direction and magnitude carry information about which muscle groups are fatiguing first, and how command is likely to degrade in the remaining pitches.

Putting it together: how the signals combine

None of these signals is decisive on its own. A modern projection framework treats them as overlapping evidence about the same underlying latent state — the pitcher's current fatigue level relative to his baseline — and combines them through some form of weighted aggregation.

A simplified sketch of the logic:

1. Start with the pitcher's true-talent baseline (expected wOBA-allowed, K%, BB%) derived from his season-long and multi-year performance.
2. Apply the season-long workload adjustment based on his cumulative pitch count and start pattern to date.
3. Apply the between-start rest adjustment based on how many days since his last outing and how heavy that outing was.
4. During the game, apply the within-game adjustment based on pitch count so far and times through the order.
5. Layer on the Statcast state adjustment based on observed velocity, spin rate, and release-point drift relative to the pitcher's own historical curves.

The most sophisticated frameworks use Bayesian methods to update the fatigue estimate continuously as new data arrives during the game. A pitcher who is throwing his 85th pitch at the same velocity he was throwing his 15th, with spin rate stable and release point unchanged, is producing evidence that his fatigue level is below what a naive pitch-count model would predict. A well-calibrated projection updates on that evidence in real time rather than sticking to the pregame estimate.

The limits of every method above

Everything described in this piece is subject to real limitations. A serious reader should hold each of them in mind.

Small samples inside single games

A pitcher throws 90 to 110 pitches in a typical start. That is a small sample from which to infer a fatigue trajectory. A velocity dip of 0.4 mph over five pitches could be fatigue; it could equally be a change in pitch mix, an intentional velocity variation to disrupt hitters' timing, or measurement noise. Models that overreact to small in-game samples produce forecasts with high variance and low predictive value.

Pitchers are not identical

Every research finding cited in this article is a league-average observation. Any given pitcher may deviate substantially. Some starters seem genuinely resistant to fatigue and hold performance across 110+ pitch outings. Others decline at 80. A projection that uses only league-average penalties will systematically miss both types.

Reverse causation and confounding

Consider a pitcher who consistently reaches 110 pitches. Is he throwing 110 because he is well-conditioned and therefore performing well, or is he performing well and therefore his manager lets him throw 110? The two hypotheses have very different implications for how much of his late-game performance to credit to fatigue resistance versus selection. Untangling this requires careful causal modeling that many public projections skip.

Modern usage has shifted the baseline

The dramatic decline in 100-pitch starts between 2015 and 2025^[9] means that fatigue thresholds calibrated on older data may not transfer cleanly. Managers now pull starters preemptively at points where 2015 managers would have let them continue, which means the historical data on 100+ pitch outings is increasingly skewed toward pitchers with unusually strong late-game performance — the ones who earned the leash. A model that treats that skewed sample as representative of typical late-game outcomes will produce optimistic forecasts for the rare modern starter allowed to reach that pitch count.

Injury risk is not performance risk

Some of the most-cited fatigue research — including MLB's Pitch Smart guidelines^[13] — is oriented toward injury prevention, not performance projection. A pitcher pushed past a fatigue threshold is more likely to get hurt, but the immediate next-outing performance projection is a different question. A projection framework should not conflate the two.

Why this matters for readers of any projection

When you see a pitching projection quoted anywhere — this publication or another — the question worth asking is not "what number did the model output?" but "which fatigue signals does it use, and how does it combine them?" A projection built only on pitch count and rest days is describing 2005-era fatigue understanding. One that incorporates within-game Statcast signals is closer to the current state of the art but has to justify its weighting choices. One that claims certainty about any pitcher's fatigue level on any given day is overreaching, because the underlying phenomenon is intrinsically noisy and pitcher-specific.

The honest posture is calibrated skepticism. Fatigue is real, its effects on performance are measurable, and the signals described here are the best current tools for quantifying it. But every projection that includes a fatigue adjustment is making assumptions about which signals to weight, how much correlation to allow between them, and how strongly to update in-game. Those assumptions are model choices, not physical laws, and they shape the projection's output as much as the pitcher's actual state does.