Bat speed as an early signal — pre-registration

Pre-registration — bat speed as an early-season signal (H-BT)

Frozen 2026-07-13, before any 2026 bat-tracking or 2026 outcome data was examined. The git timestamp of this file is the freeze receipt. One confirmatory run; results published regardless.

Motivation

Statcast bat tracking (public since 2024) measures bat speed on every competitive swing — roughly 3–4x the April sample size that EV95 gets from balls in play. Exploratory work on the train years (2024, 2025; catalyst/analysis/bt_exploratory.py) found:

• Bat speed is the most stable metric we have measured: YoY r = 0.914 (2024→2025), above EV95's 0.884 in our build.

• At ANNUAL granularity it adds nothing beyond EV95 (r ≈ 0.89 collinearity; incremental betas insignificant). Season-scale integration is NOT on the table.

• At APRIL granularity it adds real signal. Predicting May-onward value on contact from April-only inputs (non-overlapping design):

  2024 (n=198): bat_speed beta +0.413, p < 0.0001
  2025 (n=208): bat_speed beta +0.231, p = 0.015
  

  In both years April bat speed out-predicted April results (aval) and rivaled April EV95.

Data (all archived before the freeze)

• Savant bat-tracking leaderboard CSVs, raw archived at data/raw/battracking/ (seasons 2024–2026; April windows 2024/2025/2026), parquets at data/parquet/bat_tracking_*. The 2026 files were fetched but ONLY row counts observed.
• Outcomes from the engine's own lake: bip_{season}.parquet joined to game dates via events_{season}.parquet; value map {1B .9, 2B 1.25, 3B 1.6, HR 2.0}, the same linear weights the engine uses everywhere.

PRIMARY hypothesis (H-BT-E, confirmatory)

On 2026 data, untouched at freeze time:

z(aval_May1_to_Jul12) ~ z(aval_April) + z(ev95_April)
                        + z(bat_speed_April)     [OLS]

Sample: batters qualified on the Savant April-2026 window (minSwings=q) AND >= 30 April balls in play in our lake AND

= 60 balls in play from 2026-05-01 through the data on hand at run time (ingested through last night's press).

SUCCESS = bat_speed coefficient positive AND two-tailed p < 0.05. Anything else — wrong sign, p >= 0.05 — is a MISS and gets published exactly like a hit would.

SECONDARY (reported, with criterion)

Stability 2025→2026: YoY r of avg bat speed among batters qualified in both seasons. SUCCESS = r >= 0.75.

TERTIARY (reported, no criterion — expected null)

2026 season-to-date aval ~ 2025 aval + 2025 ev95 + 2025 bat_speed. The train years say bat speed adds nothing at annual scale; we report this to keep ourselves honest about scope, not to claim it.

Integration gate

If PRIMARY passes: bat speed earns candidate status as an EARLY-SEASON de-noiser only — an April/May adjustment to the boards' contact-quality estimate while EV95 is still stabilizing. Any product integration is designed separately, ratified by the founder, and validated on its own terms before it touches a board. If PRIMARY fails: the finding dies in public, filed next to the clutch null.

Caveats declared in advance

• Two train pairs is thin history; bat tracking simply hasn't existed longer. The 2024 leaderboard covers from early April 2024 (public rollout backfilled), so "2024 April" is slightly truncated at the front.
• The confirmatory outcome window ends at run time (~2026-07-12), not season's end. It measures 2.5 months of "the rest of the season," which matches the operational use (early boards), not a full-season claim.
• Savant qualification thresholds ("q") are Savant's, not ours; they gate the sample identically in train and test.
• One evaluation. No re-runs, no threshold shopping. The script is catalyst/analysis/bt_confirmatory.py, committed before it was executed.