Fatigue and Injury Risk Modeling in MLB Prediction
A starting pitcher on June 1 is not the same pitcher on September 1. The cumulative toll of throwing a baseball at maximum effort, start after start, over a six-month season degrades velocity, sharpens the decline of command, and elevates the risk of injury in ways that static talent models cannot capture. Prediction systems that ignore fatigue are essentially treating every start as if it occurs in a vacuum, disconnected from the 800 pitches thrown in the preceding two weeks or the cross-country flight completed twelve hours ago.
Fatigue modeling attempts to quantify these accumulated effects and feed them into game-level predictions as dynamic, time-varying adjustments to a pitcher's expected performance. It is one of the more difficult modeling challenges in baseball, because fatigue is not directly observable. It must be inferred from proxy signals: pitch velocity trends, spin-rate decay, release-point drift, and the historical relationship between workload and performance degradation.
Pitch Count as a Crude Fatigue Proxy
The simplest and most widely used fatigue indicator is pitch count, both within a single game and across recent starts. Within a game, the relationship between pitch count and performance degradation is well-documented: by the time a starter reaches 90 to 100 pitches, his average fastball velocity has typically declined by 0.5 to 1.5 mph relative to his first-inning baseline, strikeout rate drops, and the probability of allowing hard contact increases.
But pitch count alone is a blunt instrument. It does not differentiate between 95 pitches thrown in five stressful, high-leverage innings and 95 pitches thrown across seven efficient innings. A pitcher who faces the minimum through five innings has experienced a fundamentally different physiological workload than one who has thrown 40 pitches in a single inning while navigating a bases-loaded jam. Stress-adjusted pitch counts, which weight pitches by game context (runners on base, close score, high-leverage situations), represent an attempt to refine this metric, though the empirical evidence for stress-weighting remains mixed.
Workload Decay Functions
More sophisticated fatigue models move beyond simple pitch counts and track how specific measurable outputs decay as workload accumulates. The key decay signals are velocity, spin rate, and command precision.
Velocity Decay
Fastball velocity is the most reliable real-time fatigue indicator. Most starting pitchers show a characteristic velocity curve: peak velocity in the first or second inning, a plateau through the middle innings, and a decline of 1 to 2 mph as they approach the 85-to-100 pitch threshold. The shape of this curve varies by pitcher. Power arms with high peak velocities tend to show steeper decline curves, while soft-tossing control pitchers may maintain velocity longer but with a sharper cliff when fatigue finally hits.
Prediction models can learn pitcher-specific velocity decay functions from historical Statcast data. When a pitcher's in-game velocity drops below his personal decay curve faster than expected, the model infers above-normal fatigue and adjusts his projected performance downward for the remaining plate appearances in that start.
Spin Rate and Movement Degradation
Spin rate decline is a subtler fatigue signal that often precedes velocity loss. As the muscles in the forearm and fingers fatigue, the ability to generate maximal spin on off-speed pitches deteriorates. A curveball that spins at 2,800 RPM in the first inning may drop to 2,650 RPM by the sixth inning, resulting in less vertical break, a flatter trajectory, and a higher probability of being squared up by the hitter. Models that track spin-rate decay curves alongside velocity provide a more complete picture of within-game fatigue progression.
Release Point Consistency
Fatigue manifests mechanically as inconsistency in the release point. A well-rested pitcher delivers the ball from a tight cluster of release points, making it difficult for hitters to distinguish between pitch types. As fatigue accumulates, the release-point cluster expands. The variance in both horizontal and vertical release position increases, which has two consequences: command deteriorates (pitches miss their intended location by wider margins), and the pitcher becomes more readable to hitters because different pitch types begin releasing from detectably different positions.
Days-of-Rest Effects
The interval between starts is a critical fatigue variable with a nonlinear performance relationship. On standard four days of rest (the modern five-man rotation), most pitchers perform at their baseline level. On short rest (three days), historical data shows measurable degradation: ERA increases of 0.3 to 0.5 runs, velocity drops of 0.3 to 0.7 mph, and elevated walk rates. The effect is not enormous for any single start, but it compounds across a season and becomes significant in playoff contexts where short-rest starts are more common.
Extra rest (five or more days) produces a more complex pattern. One extra day of rest typically has a neutral-to-slightly-positive effect. Extended rest of seven or more days, which occurs after the All-Star break or during injury returns, often produces a one-start adjustment period where the pitcher's command is worse than baseline before returning to normal. Models that treat days-of-rest as a linear variable miss this curvilinear relationship. Piecewise functions or categorical binning (short rest, normal rest, extra rest, extended rest) capture the pattern more accurately.
Cumulative Season Workload
Beyond game-to-game rest intervals, the total workload accumulated across a full season affects pitcher performance in the later months. The concept of "innings-pitched thresholds" has been debated extensively. While there is no universal cliff at exactly 180 or 200 innings, aggregate data shows that pitchers who exceed their previous career-high workload by a significant margin tend to show degraded performance in September and elevated injury risk in the following season.
The mechanism is cumulative microtrauma: the repeated stress of throwing at high effort damages connective tissue, tendons, and ligament fibers at a rate that exceeds the body's ability to fully repair between starts. This damage is not visible in a single-start stat line. It accumulates over weeks and months, eventually manifesting as reduced velocity, diminished breaking-ball quality, and the dreaded "dead arm" period that many pitchers experience in late August or September.
Workload models track cumulative pitch counts, innings pitched, and a weighted stress index across rolling windows of 30, 60, and 90 days. When a pitcher enters territory that historically correlates with performance decline, the model reduces his projected effectiveness accordingly, even if his most recent start looked sharp.
Travel and Scheduling Fatigue
Baseball's 162-game schedule imposes travel demands that no other major sport matches. Cross-country flights, time-zone changes, day games following night games, and extended road trips all contribute to fatigue that extends beyond the pitcher to affect the entire roster.
The measurable effects of scheduling fatigue in aggregate data are modest but statistically significant. Teams playing day games after night games show a small decline in offensive production, likely driven by disrupted sleep patterns. Teams on the final game of a long road trip (10 or more days) show slightly elevated error rates and reduced plate discipline, suggestive of mental and physical fatigue. West Coast teams playing early-start East Coast games show reduced first-inning performance, consistent with circadian rhythm disruption.
For prediction models, scheduling variables enter as categorical features: day-after-night flag, road trip length, time-zone crossing count, and days since last off-day. These are not high-magnitude effects individually, but they accumulate and interact with pitcher workload in ways that a comprehensive fatigue model should capture.
Injury Risk Probability Models
The most ambitious application of fatigue modeling is predicting when a pitcher is at elevated risk of injury. This moves beyond forecasting performance degradation and attempts to forecast the probability of a pitcher missing time entirely, which has an outsized effect on team-level predictions.
Injury risk models typically combine workload variables (cumulative pitch count, innings trajectory relative to career norms, recent rest patterns) with biomechanical signals derived from Statcast data. Declining velocity trends, inconsistent release points, and changes in arm slot angle can serve as early warning indicators that a pitcher's mechanics are compensating for underlying fatigue or discomfort. When multiple indicators align, the model flags elevated injury probability.
These models are inherently probabilistic. They cannot predict that a specific pitcher will get hurt on a specific date. They can estimate that a pitcher in a given workload and biomechanical state has, for example, a 12 percent probability of an IL stint in the next 30 days versus a baseline rate of 4 percent. This probability then weights the prediction model's expected contribution from that pitcher, blending his projected healthy performance with the probability that a replacement-level arm takes his start instead.
Fatigue and Bullpen Deployment
Pitcher fatigue modeling does not exist in isolation. When a starting pitcher is projected to fatigue earlier than normal (due to short rest, high recent workload, or in-game velocity decline), the model must also project when the bullpen will be called upon and which relievers are likely to be available. This creates a direct interaction with bullpen usage modeling, where the projected quality of available relievers depends on their own recent workload.
The fatigue cascade is a system-level phenomenon: a fatigued starter exits early, shifting innings to the bullpen. The bullpen was already taxed from covering extra innings in the previous game. The team is forced to use lower-leverage relievers in higher-leverage situations. Run prevention degrades not because any single pitcher is dramatically worse, but because the aggregate quality of pitching deployed across nine innings is lower than the sum of its rested parts.
Position Player Fatigue
Whether position player fatigue exists in a statistically measurable form remains an open question. Unlike pitchers, whose workload involves explosive, high-stress movements with clear physiological costs, everyday position players perform activities (running, throwing, swinging) that are within normal athletic capacity. The 162-game schedule is grueling, but the evidence for systematic within-season performance decline among position players is weak after controlling for injuries and normal statistical variance.
Some research has identified small effects: catchers show performance decline in the second game of doubleheaders, players with heavy recent playing time show slightly reduced sprint speed in Statcast tracking data, and plate discipline metrics may erode marginally during extended stretches without days off. These effects are small enough that most prediction models treat position player fatigue as negligible relative to pitching fatigue, though this may change as more granular biomechanical tracking data becomes available.
Integrating Fatigue into Game Predictions
The practical output of fatigue modeling is a set of adjustments applied to a pitcher's baseline projection for a specific start. A pitcher with a season-long 3.50 ERA might be projected at 3.80 for a start on short rest after 110 pitches five days ago, or at 4.20 for a September start after exceeding his career-high innings threshold. These adjustments feed into simulation systems that run thousands of game iterations using the fatigue-adjusted pitcher performance as input.
The models that handle fatigue well treat it as a continuous, multi-dimensional state rather than a binary fresh/tired classification. A pitcher is always somewhere on a fatigue continuum that reflects his recent pitch count, his cumulative season workload, his rest interval, his travel schedule, and the biomechanical signals from his most recent outings. Prediction systems that encode this full state vector produce more accurate game-level forecasts than those that ignore the accumulation of stress across a long season.