BASEBALL OPERATIONS · PUBLIC-DATA RESEARCH
An auditable bullpen research desk. Trace a pitcher’s recent usage, inspect the performance evidence, and test whether workload improves the next-appearance forecast.
No clear predictive gain from workload. Recent usage adds context, but this model did not improve on prior performance and entry conditions in 2025.
See the test ↗01 / HISTORICAL APPEARANCE BOARD
Explore actual 2025 relief appearances. Workload columns use earlier dates only. This is an appearance log, not a live roster or availability recommendation.
| Date | Pitcher | MLB off-days | Prior 3d | Prior 7d | Usage note |
|---|
Rest days = full calendar days without a recorded MLB pitching appearance. “0” means pitched the day before. Counts omit minor-league appearances, warm-ups, bullpen sessions and other unrecorded throwing; an MLB off-day is not proof of physical rest. Filters are descriptive—not validated fatigue thresholds.
PITCHER DOSSIER / AS OF THE SELECTED APPEARANCE
EARLIER MLB THROWING EXPOSURE
All recorded MLB pitching, including starts and games with other clubs. No pitches from the selected appearance or later dates enter this view. A blank bar means no recorded MLB pitches, not confirmed rest.
| Date | Recorded pitches |
|---|
BEFORE THIS APPEARANCE
Workload percentiles describe usage relative to this pitcher’s earlier eligible appearances. They are not probabilities of fatigue, injury or poor performance.
EARLIER OUTCOMES · ALREADY KNOWABLE
| Date | Club | MLB off-days | Prior 7d pitches | Outing pitches | BF | K | uBB | K−uBB rate |
|---|
This table applies the research cohort rules. The daily chart above includes recorded pitching outside that cohort. One short outing can swing a rate sharply; use the denominators.
02 / LEAGUE-WIDE EVIDENCE
Strong relievers may be used more often. Raw results alone can confuse pitcher quality with workload. Start with a pitcher’s prior performance, then adjust for context.
DESCRIPTIVE COMPARISON
Strikeout-minus-unintentional-walk rate, in percentage points above or below the pitcher’s expanding, shrunk baseline from earlier appearances in the same season. Higher is better.
| Rest | Apps | Difference | 95% interval |
|---|
Whiskers show 95% intervals. Pitcher-cluster bootstrap; 500 resamples. These are within-stratum intervals, not tests of pairwise differences. Baseline adjusts for prior performance, not every deployment decision.
ADJUSTED ASSOCIATIONS · 2023–2024
Pitcher-season fixed effects, entry conditions and recent performance. Effects are percentage points in K−uBB rate, holding included variables fixed.
95% pitcher-clustered intervals. Correlated workload variables can be difficult to separate. An interval crossing zero is compatible with effects in either direction; it does not prove no fatigue exists.
We observe pitchers who were used. A manager may already have rested those most fatigued. Unavailable pitchers, warm-ups and private recovery measures are missing.
03 / THE TEST THAT COUNTS
Fit on 2023–2024, freeze coefficients, then predict 2025. Earlier 2025 appearances may update the prior-performance baseline; no later outcomes enter a prediction.
Prior K−uBB rate, entry inning, score margin, inherited runners, home/away, temperature and month.
Add rest groups, prior 3- and 7-day pitches, 20+ pitches yesterday and usage on both prior days.
INCREMENTAL VALUE
RMSE change · positive is worseDeployment decision: do not promote this model into a pitcher-ranking or availability system. Use the workload log to structure questions and gather better evidence.
RMSE measures typical prediction error; lower is better. Both models use the same fixed ridge penalty and batters-faced weights. Uncertainty uses 1,000 paired resamples of held-out pitchers. The target is a noisy single-appearance rate, not an injury outcome.
POST-HOC CHECKS / FROZEN MODELS
Same two fitted models, viewed through different slices of 2025. These checks were added after the first result; they are diagnostics, not a fresh test or model-selection exercise.
| Evaluation slice | Apps / pitchers | Context RMSE | + workload RMSE | Change / 95% interval |
|---|
All errors are percentage points; positive change means worse performance. Paired pitcher-cluster bootstrap, 1,000 resamples per slice. Intervals are pointwise, not adjusted for looking at multiple slices. “Late & close” is not a formal leverage index.
CALENDAR STABILITY
| Month | Apps | Change, pp | 95% interval |
|---|
INTERPRETING THE TEST
This checks whether the chosen workload features add value to this ridge model for K−uBB rate. It does not test every possible fatigue signal, model or baseball outcome.
Managers may have rested the pitchers most affected. We cannot observe how those rested pitchers would have performed.
A workload feature can predict without causing an outcome—or have a real causal effect that this dataset cannot identify.
04 / TORONTO · STATCAST CASE STUDY
Track four-seam velocity against the pitcher’s prior five qualifying Toronto relief appearances. Keeping pitch type constant avoids mistaking a different pitch mix for lost velocity.
| Rest | Apps / pitchers | Δ mph | 95% interval |
|---|
PITCHER EXPLORER
At least five measured four-seam pitches in the current appearance and three prior qualifying appearances (up to five averaged, equal appearance weight). Baselines reset by season and only use Toronto appearances. This small, exploratory sample is not an independent validation of the league model.
Recalculate the zero-MLB-off-day mean after excluding current appearances where the two feeds differ by more than two pitches. Keep the original trailing velocity baselines fixed. This checks sensitivity to one data-quality rule; it does not validate the missing-pitch mechanism.
| Season | Current-appearance filter | Apps / pitchers | Mean velocity change |
|---|
05 / REPRODUCIBLE RESEARCH
Download the analysis tables, inspect the definitions and reproduce the results from the public source files.
This project measures recorded workload and subsequent performance. It cannot identify a pitcher’s underlying fatigue, diagnose injury risk, or establish what would happen if a manager changed the rest schedule.
The practical use is an auditable conversation: who pitched recently, what was recorded, what changed, and what additional information would a coach need?
Read the analysis specification ↗Read the research brief ↗Download held-out appearances ↓Download frozen predictions ↓Download modeling table ↓Download Toronto velocity data ↓Retrosheet daily logs and parsed plays: 2023–2025 regular seasons. Baseball Savant: Toronto pitching, 2024–2025 regular seasons. Field meanings follow the Retrosheet dictionary and Statcast CSV documentation. Retrieved September 17, 2026.
The information used here was obtained free of charge from and is copyrighted by Retrosheet. Interested parties may contact Retrosheet at 20 Sunset Rd., Newark, DE 19711.
Independent portfolio research. No affiliation with MLB or the Toronto Blue Jays. Source files are downloaded by the reproduction script; the site distributes derived tables.
Relief appearances are pitching sequence greater than one, with positive batters faced and at least five earlier relief appearances in that season. A retrospective season-role filter requires pitching appearances to be at least half of all games played. This excludes position-player pitching; it also limits evaluation to this defined cohort.
Workload sums recorded pitches over the previous calendar dates, including starts and actual continuation dates of suspended games. Suspended-game outcomes and pitcher-days with two appearances are excluded from modeling. Minor-league and rehab appearances, warm-ups and bullpen sessions are not available; gaps in MLB records must not be interpreted as confirmed rest. Missing pitch counts on non-plate-appearance events count as zero; every completed plate appearance has a recorded count. Missing temperature uses the development-sample median.
Outcome = (strikeouts − walks + intentional walks) / all batters faced. The prior baseline pools earlier non-suspended relief outcomes with 100 pseudo-batters at a fixed 15% rate; it resets each season. Both ridge models use development-only standardization, penalty 100 and batters-faced weights. No tuning on the 2025 test set.
Association estimates use weighted demeaning within pitcher-season, WLS and pitcher-clustered confidence intervals with absorbed-effect degrees-of-freedom correction and a cluster-t critical value on development seasons. They are distinct from the predictive ridge coefficients. Descriptive and velocity intervals resample pitchers. Fixed random seed: 71. Entry context is measured when a pitcher enters; these are not day-before forecasts.
Integrate warm-up counts, exact recovery hours, travel, role and roster availability. Add opponent and handedness adjustment, pitch shape and movement, and pitch-type-specific league-wide velocity. Use rolling-origin validation and a genuinely new season before operational use.
Study the decision to deploy as well as the performance of deployed pitchers. A causal rest-policy claim needs a design that addresses that selection process; a workload correlation alone is insufficient.
Run python download.py --raw raw, python analyze.py --raw raw, python toronto.py --raw raw, python diagnostics.py --raw raw, and python build_data.py. Dependencies and detailed instructions are in README. Source manifest records file hashes.