Datasets:

yasumorishima
/

mlb-stats

pretty_name: MLB Shared Stats
tags:
  - baseball
  - mlb
  - statcast
  - sabermetrics
language:
  - en
configs:
  - config_name: catcher
    data_files: catcher.parquet
  - config_name: fg_batting
    data_files: fg_batting.parquet
  - config_name: fg_pitcher_plus
    data_files: fg_pitcher_plus.parquet
  - config_name: fg_pitching
    data_files: fg_pitching.parquet
  - config_name: oaa
    data_files: oaa.parquet
  - config_name: oaa_team
    data_files: oaa_team.parquet
  - config_name: park_factors
    data_files: park_factors.parquet
  - config_name: sc_bat_tracking
    data_files: sc_bat_tracking.parquet
  - config_name: sc_batted_ball
    data_files: sc_batted_ball.parquet
  - config_name: sc_batter_exitvelo
    data_files: sc_batter_exitvelo.parquet
  - config_name: sc_batter_expected
    data_files: sc_batter_expected.parquet
  - config_name: sc_pitcher_arsenal
    data_files: sc_pitcher_arsenal.parquet
  - config_name: sc_pitcher_exitvelo
    data_files: sc_pitcher_exitvelo.parquet
  - config_name: sc_pitcher_expected
    data_files: sc_pitcher_expected.parquet
  - config_name: sprint_speed
    data_files: sprint_speed.parquet
  - config_name: statsapi_batting
    data_files: statsapi_batting.parquet
  - config_name: statsapi_pitching
    data_files: statsapi_pitching.parquet
  - config_name: mart_batter_season
    data_files: marts/mart_batter_season.parquet
  - config_name: mart_pitcher_season
    data_files: marts/mart_pitcher_season.parquet
  - config_name: mart_batter_aging_pairs
    data_files: marts/mart_batter_aging_pairs.parquet
  - config_name: mart_pitch_arsenal_scouting
    data_files: marts/mart_pitch_arsenal_scouting.parquet
  - config_name: mart_scouting_reliability
    data_files: marts/mart_scouting_reliability.parquet
  - config_name: mart_batter_process_reliability
    data_files: marts/mart_batter_process_reliability.parquet
  - config_name: mart_pitcher_process_reliability
    data_files: marts/mart_pitcher_process_reliability.parquet
  - config_name: mart_fielding_running_reliability
    data_files: marts/mart_fielding_running_reliability.parquet

MLB Shared Stats

Season-level MLB tables, refreshed weekly by yasumorishima/mlb-data-pipeline and published here as Parquet. One file per table at the repository root.

Read the freshness column before using a table. Not everything here is current, and the rows in a stale table look exactly like the rows in a fresh one.

Tables

File	Source	Seasons	Refreshed
`sc_batter_exitvelo.parquet`	Baseball Savant	2015–	weekly
`sc_pitcher_exitvelo.parquet`	Baseball Savant	2015–	weekly
`sc_batter_expected.parquet`	Baseball Savant	2015–	weekly
`sc_pitcher_expected.parquet`	Baseball Savant	2015–	weekly
`sc_pitcher_arsenal.parquet`	Baseball Savant	2017–	weekly
`sc_batted_ball.parquet`	Baseball Savant	2015–	weekly
`sc_bat_tracking.parquet`	Baseball Savant	2024– (Hawk-Eye)	weekly
`sprint_speed.parquet`	Baseball Savant	2015–	weekly
`oaa.parquet`	Baseball Savant	2016–	weekly
`oaa_team.parquet`	Baseball Savant	2016–	weekly
`catcher.parquet`	Baseball Savant	2015–	weekly
`park_factors.parquet`	Baseball Savant	2015–	weekly (added 2026-09-23)
`statsapi_batting.parquet`	MLB Stats API	2015–	weekly (added 2026-09-23)
`statsapi_pitching.parquet`	MLB Stats API	2015–	weekly (added 2026-09-23)
`fg_batting.parquet`	FanGraphs	2015–2025	⚠️ frozen 2026-04
`fg_pitching.parquet`	FanGraphs	2015–2025	⚠️ frozen 2026-04
`fg_pitcher_plus.parquet`	FanGraphs	2020–2025	⚠️ frozen 2026-04

statcast_pitches (pitch-level, ~6.8M rows) is fetched by manual dispatch only and is not part of the weekly refresh.

Marts (`marts/`)

Analysis-ready tables built from the tables above by the dbt project and republished after every weekly refresh, only when all of its data tests pass. A build that fails leaves the previous marts in place, so they can lag the raw tables: marts/_manifest.json records when they were built, the pipeline commit, and the revision of this dataset they were built from. Rows for a season whose regular season is not over carry is_partial = true.

File	Grain
`marts/mart_batter_season.parquet`	batter-season (PA > 0): wOBA, wRC+, WAR next to xwOBA, batted-ball mix, bat speed, sprint speed, OAA
`marts/mart_pitcher_season.parquet`	pitcher-season: K%, BB%, K-BB%, FIP, xERA, pitch-mix breadth
`marts/mart_batter_aging_pairs.parquet`	same batter, season and season + 1: input for aging curves
`marts/mart_pitch_arsenal_scouting.parquet`	pitcher-season-pitch: usage rank, whiff and run-value percentiles within type
`marts/mart_scouting_reliability.parquet`	metric x sample-size bin: year-to-year correlation of each pitch metric, raw and within pitch type
`marts/mart_batter_process_reliability.parquet`	metric x PA bin: year-to-year correlation of each batting number and its correlation with next season's wOBA
`marts/mart_pitcher_process_reliability.parquet`	metric x BF bin: year-to-year correlation of each pitching number and its correlation with next season's ERA
`marts/mart_fielding_running_reliability.parquet`	metric x position or competitive-run bin: year-to-year correlation of OAA, fielding runs prevented, catch rate above expected, sprint speed and home to first

Why the three FanGraphs tables are frozen

FanGraphs refuses the GitHub Actions runner. The cause is the address, not the client: pybaseball sets no User-Agent, so it sends the honest python-requests default, and it still got HTTP 403 for all 12 seasons on run 35565978836 (2026-09-21). The same request from a home connection returns 200.

There is a second, separate layer that is easy to confuse with it: a client claiming to be a browser gets 403 with cf-mitigated: challenge from anywhere, including a residential line. Measured 2026-09-23.

So these three tables hold a rescue snapshot taken in 2026-04, which ends with the 2025 season. They contain no 2026 rows. The pipeline reports them on every run and the exemption has an expiry date, so this cannot quietly become permanent.

For current wOBA / wRC+ / WAR / FIP / xFIP, use statsapi_batting and statsapi_pitching below. They are not FanGraphs tables.

`statsapi_batting` / `statsapi_pitching`

From the MLB Stats API, keyless: stats=season (counting and rate stats) joined with stats=sabermetrics (wOBA, wRAA, wRC, wRC+, WAR and its components for hitters; FIP, xFIP, FIP-, ERA-, WAR, RA9-WAR, leverage for pitchers), playerPool=ALL. Column names are the API's own (plateAppearances, wRcPlus, xfip, …). Responses carry "Copyright MLB Advanced Media, L.P." and point to the terms at http://gdx.mlb.com/components/copyright.txt.

MLB does not document how its sabermetrics feed is computed or where it comes from. What we measured, 2026-09-23, joining on MLBAM player_id against the frozen FanGraphs snapshot above (every one of its 5,703 hitter and 4,648 pitcher rows matched):

Counting stats agree exactly (HR, SO; PA within 1).
2015–2021: wOBA within 0.002, wRC+ within 0.5 (FanGraphs stores integers), WAR within 0.2, FIP / xFIP within 0.005 — i.e. rounding.
2022–2025: wOBA still within 0.006, but wRC+ differs by up to 1.0 (2022, 2023), 3.4 (2024) and 5.6 (2025), and the 2025 difference is lined up by club (ATH +5.1, CIN −3.1). Park factors were revised after the snapshot. WAR follows: up to 0.34 for hitters in 2024, and 0.39 for hitters and 0.40 for pitchers in 2025. One 2024 pitcher differs by 0.31 in FIP and 0.32 in xFIP; the other 433 by ≤ 0.005.

So past seasons can change from one week to the next; every run refetches every season.

Rows and columns:

One row per (player_id, season), every player who appeared, not a qualified subset (e.g. 1,252 hitters in 2015 against 546 in the FanGraphs snapshot). The hitter count falls from ~1,250 to ~770 in 2022 because the universal DH ended pitchers batting.
A player who changed clubs has one row with his season total. last_team_id is the club he finished with and num_teams how many he played for. It is not "his stats for that club".
is_partial is true for the current season until every regular-season game is final. fetched_at is when the row was read.
inningsPitched is kept as the API's string: "5.1" means five and one third innings, not 5.1. Use outs / 3.
Rates are absent where the denominator is zero: woba / wRcPlus for hitters with 0 PA (288 of 1,252 in 2015), fip / xfip for a pitcher with 0 outs. Placeholder strings such as ".---" become null.
In the current season a pitcher can appear in stats=season before stats=sabermetrics (one in 2026); his sabermetric columns are null.

The fetch fails and these two tables are not published if a response was truncated (totalSplits ≠ rows returned), answered for another season, lacks a required stat, repeats a player, has a player only the sabermetrics side knows, or if a completed season now has fewer players than the copy already published here. A run that checked the ids and knows MLB removed a player (e.g. merged a duplicate id) can publish with the allow_statsapi_shrink workflow input.

For every table except statcast_pitches (one file per season, so a narrower run cannot remove the others), the output audit also refuses to publish a file that lacks a season the published copy has, so a manual run over a narrower year range cannot delete seasons from here. A deliberate removal has to name the table in the allow_lost_seasons workflow input.

The weekly run covers seasons up to the latest one in which every club has played a game, read from the Stats API standings; before that (and through an overseas opening series) it stops at the previous season.

Why you can trust the "weekly" column

Every run audits what it produced. A table that fetched nothing used to be skipped silently — the upload step pushes whatever files exist, so the copy here simply stayed as it was and the job still went green. That is how park_factors never arrived at all until 2026-09-23, and how the FanGraphs tables went stale without an alarm.

Now scripts/check_outputs.py prints one row per expected table, publishes only the tables that passed, and turns the run red otherwise.

`park_factors` columns

One row per club per season, all 30 clubs in every season. 100 = neutral, above 100 favours hitters. From Baseball Savant's Statcast park factors.

season, team, venue_id, venue_name, n_pa_1yr, n_pa_3yr, pf_1yr, pf_3yr, pf_3yr_years, then pf_hr, pf_1b, pf_2b, pf_3b, pf_so, pf_bb, pf_obp, pf_hits, pf_woba, pf_wobacon, pf_xwobacon, pf_bacon, pf_xbacon, pf_hardhit, pf_wobatto.

Everything except pf_1yr and n_pa_1yr comes from a 3-year window; pf_3yr_years records which window that was.

Two NaN patterns are expected, not defects:

8 rows of 360 have no 3-year window at all, because the park has no three-year history: 2017 ATL, 2018 ATL, 2020 TEX, 2020 TOR, 2021 TEX, 2025 OAK, 2025 TB, 2026 OAK. Their pf_3yr* columns are NaN while pf_1yr is filled. The club is kept rather than dropped, so a team never silently disappears from a season.
pf_xwobacon, pf_xbacon and pf_hardhit are NaN for all 30 clubs in 2015 and 2016, because those windows reach back before Statcast measured batted balls. Those three columns have 68 NaN, not 8.

Subdirectories

mlb_bat_tracking/ and mlb_wp/ come from other projects and are not touched by the weekly refresh. Their freshness is unrelated to the table above.

Usage

import pandas as pd

base = "https://huggingface.co/datasets/yasumorishima/mlb-stats/resolve/main/"
pf = pd.read_parquet(base + "park_factors.parquet")
print(pf[pf["season"] == 2026].nlargest(3, "pf_3yr")[["team", "venue_name", "pf_3yr"]])

Sources and terms

Baseball Savant — MLB Advanced Media
FanGraphs — the three fg_* tables
MLB Stats API — statsapi_batting and statsapi_pitching; responses state "Copyright MLB Advanced Media, L.P." and refer to http://gdx.mlb.com/components/copyright.txt
Fetched with pybaseball and savant-extras

This is a derived, aggregated mirror published for research and reproducibility. No licence is asserted over the underlying data, which remains subject to the terms of the sources above; check those before redistributing. Open an issue on the pipeline repository for corrections.