The Central Pattern — Raw Tool vs. Specific Skill Across 3 Domains

This portfolio has repeatedly tested the same question in different areas of the game: does a raw physical tool predict outcomes as well as a more specific skill built on top of it? Pulling the original datasets from three prior repos and re-verifying every number here, the honest answer is: it depends on the domain — and hitting is the one place where the raw tool wins, decisively.

Part of the Emerson Performance analytics portfolio (EP-TSP framework). This is a synthesis repo — it doesn't introduce new data, it re-verifies and compares findings already published in ep-swing-intelligence, vaa-approach-angle-study / fastball-shape-outcomes-study, and ep-defense-intelligence.

Versión en español disponible aquí

---

The question

Three separate projects in this portfolio asked a version of the same question in a different domain. This repo puts all three side by side, using each project's original dataset, to see whether there's a real, general pattern — or whether each result was domain-specific all along.

The finding: it depends on the domain

Domain	Raw Tool	r	Specific Skill	r	Winner
Hitting	Bat Speed	0.74	Squared-Up%	-0.46	Raw tool
Pitching	Induced Break (IVB)	-0.03	VAA (height-adj.)	0.34	Skill
Defense	Sprint Speed	0.39	Burst	0.73	Skill

In pitching and defense, the more specific, harder-to-scout skill predicts the outcome better than the simple raw-tool measurement. In hitting, the opposite is true — and not by a small margin. Bat speed correlates strongly and positively with Barrel% (r=0.74); squared-up%, the metric this portfolio's earliest work treated as the more sophisticated predictor, is actually negatively correlated with Barrel% (r=-0.46) in this same 205-player sample.

Why hitting is different: a real trade-off, not noise

This isn't a data error. Bat speed and squared-up% are themselves negatively correlated with each other (r=-0.65) — hitters who swing harder tend to square the ball up less consistently, a real and well-documented power/contact trade-off. Because bat speed drives Barrel% so strongly on its own, squared-up%'s raw correlation with Barrel% gets pulled negative through that indirect relationship, even though — as the standardized-coefficient analysis in ep-swing-intelligence shows — squared-up% still carries real, independent predictive weight once bat speed is statistically controlled for. Simple correlation and controlled effect size are telling two different, both-true stories here.

Why this matters

The lesson isn't "skill always beats tool" — an earlier draft of this portfolio's language said exactly that, and it was wrong. The real lesson is more useful and more honest: a raw physical tool and a more specific skill usually both carry real information, and which one "wins" in a simple comparison depends on the domain, the outcome being predicted, and how the two metrics relate to each other. For a front office, this argues against a blanket rule ("always weight skill over tool," or the reverse) in favor of testing the specific relationship in each domain being evaluated — exactly the exercise this repo performs.

Repo structure

ep-tsp-central-pattern/
|-- data/
|   |-- hitting_data.csv
|   |-- pitching_vaa_data.csv
|   |-- pitching_ivb_data.csv
|   `-- defense_data.csv
|-- scripts/
|   |-- central_pattern_analysis.py
|   `-- ep_chart_style.py
`-- outputs/
    |-- central_pattern_{EN,ES}.png
    `-- central_pattern_summary.csv

Reproduce the analysis

git clone https://github.com/ejimenezperformance/ep-tsp-central-pattern.git
cd ep-tsp-central-pattern
pip install pandas matplotlib
python scripts/central_pattern_analysis.py

Methodology

• Hitting data: the 205-player merged dataset from ep-swing-intelligence (bat speed, squared-up% per swing, and Barrel%, all official Baseball Savant 2026 figures). Correlations recomputed directly from that dataset for this repo, not copied from the other repo's reported summary numbers.
• Pitching data: vaa-approach-angle-study's merged VAA/outcomes dataset (418 pitchers, VAA Above Average vs. SwStr%) and fastball-shape-outcomes-study's merged IVB/outcomes dataset, both recomputed directly from their original files.
• Defense data: ep-defense-intelligence's merged dataset (90 outfielders, Sprint Speed and Jump-Burst vs. OAA), recomputed directly.
• Correlation: simple Pearson r in every panel — the same metric used consistently across all three source repos.

Limitations

• Different sample sizes and seasons per domain (n=205 hitters vs. n=418 pitchers vs. n=90 outfielders, all 2026) — this is a synthesis of three independently-scoped studies, not one unified dataset with matched samples.
• Only three domains tested. This portfolio has also examined pitch recognition (ep-pitch-recognition), which doesn't fit the same "raw tool vs. specific skill, single correlation" structure as cleanly and isn't included here for that reason — see that repo for a related but structurally different finding.
• Correlation, not causation, in every panel — none of these comparisons establish that training one metric over another would causally change a player's outcome.
• The hitting domain's history is disclosed for a reason. An earlier version of ep-swing-intelligence reported the opposite result due to a since-corrected coefficient-scaling bug, and that earlier (incorrect) version was cited in this portfolio's other repos before being caught and fixed. This repo exists partly to close that loop with a single, consistent, independently-reproducible reference point going forward.

Contact

Emerson Jiménez — Strength & Conditioning Coach, Baseball Performance Specialist. Emerson Performance · @emersonperformance

---

EP-TSP framework and design © Emerson Performance. Statcast/Baseball Savant data is public domain, non-commercial use.