MLB WAR by direct regression

A wins-above-replacement-style baseball metric computed by direct ridge regression of half-inning run differential on player presence — not assembled bottom-up from box-score components like Fangraphs / Baseball-Reference WAR.

The methodology is regularized adjusted plus-minus (RAPM, well-known from basketball) adapted to baseball. See REPORT.md for the full writeup, including all design choices, calibration, and caveats.

Browse the leaderboards interactively → — all-time, current season with WAR-over-time chart, every individual season since 1910, and click-through to per-player career detail. Refreshed daily.

Top 15 all-time (career-WAR = sum of per-season fits)

#	Player	total_war	off	pit	fld	seasons
1	Willie Mays	63.8	39.7	—	24.2	22
2	Walter Johnson	59.0	-8.2	67.2	—	18
3	Roger Clemens	54.8	-0.7	55.6	—	24
4	Tom Seaver	52.4	-8.3	60.8	—	20
5	Eddie Murray	51.7	47.7	—	4.0	21
6	Cal Ripken	49.2	41.8	—	7.4	21
7	Brooks Robinson	49.2	44.3	—	4.9	23
8	Frank Robinson	49.0	43.1	—	5.9	21
9	Hank Aaron	48.3	48.6	—	-0.3	23
10	Dave Winfield	47.8	46.0	—	1.8	22
11	Lou Gehrig	47.4	41.5	—	5.9	17
12	Pete Alexander	45.9	-7.1	52.9	—	20
13	Gil Hodges	44.2	36.3	—	7.9	18
14	Babe Ruth	43.2	30.0	9.5	3.7	22
15	Lefty Grove	42.3	-4.4	46.7	—	17

The headline career number sums each player's per-season WAR — one ridge fit per season, then aggregated. This matches the Fangraphs/B-Ref convention and avoids the cross-era identifiability issues of fitting one model across all 116 seasons jointly (see REPORT.md §cross-era and §centering). The all-time single-fit version is still available in the webapp under "All-time (single-fit)" with a caveat banner.

Data sources

The pipeline auto-routes per year:

• Retrosheet (data/raw/<YYYY>/) — fully detailed historical event files. Used wherever a season's zip exists on disk. download_all.sh walks 1910 through the current year and fetches everything Retrosheet has published, then clears any superseded statsapi feed cache for those years.
• MLB Stats API (statsapi.mlb.com) — used for any year that doesn't have Retrosheet data yet (typically just the current in-progress season). Per-game feed/live responses are cached under data/cache/feeds/<year>/.
• Chadwick Bureau register — MLBAM ↔ Retrosheet player-ID crosswalk so the two sources can be unioned in a single fit. Cached locally on first use.

build_dataset.py --years 1910-2026 --tag all HEADs retrosheet.org for any year not on disk; if available it downloads on the fly, otherwise it falls back to statsapi. Output: half_innings_<tag>.parquet (the row-per-half-inning table the fits consume), game_park_<tag>.csv, and rosters_<tag>.csv.

How to reproduce

Requires chadwick (Retrosheet's cwevent C tool, brew install chadwick) and Python with pandas, scipy, scikit-learn.

# 1. Download every available Retrosheet season (walks 1910 → current year)
bash download_all.sh

# 2. Build the half-innings dataset for the years you want (~5 min on cold cache)
python3 build_dataset.py --years 1910-2026 --tag all

# 3. All-time single-fit (~2 min) -- produces coefficients_all{,_enriched}.{csv,parquet}
python3 fit_ridge_all.py --tag all
python3 make_views.py    --tag all

# 4. Per-season fits (~3 min for 1910-2026) -- produces season_war_all.csv and
#    the per-season-sum career table that the webapp uses as its headline.
python3 fit_per_season.py --tag all --extra-tags 2026

# 5. Regenerate the webapp manifest
python3 snapshot.py --update-manifest-only

Output leaderboards under data/events/:

File	What
career_seasons_sum_<tag>.csv	Headline career WAR, derived from the per-season fits.
coefficients_<tag>.csv / _enriched.csv	Single all-time fit with per-season/peak-rate/per-position views.
season_war_<tag>.csv	Long-format (season × player) WAR table. ~97K rows for 1910-2026.
park_effects_<tag>.csv	Per-park run environment (single-fit output).
manifest.json	Index the webapp reads to discover available views.

Daily refresh

refresh.sh is the cron entry point. It pulls the repo, runs build_dataset for the current season (statsapi), refits the current-season models, updates the year's row inside season_war_all, regenerates the manifest, and pushes the changed CSVs. Suggested crontab (also checked in as crontab):

PATH=/opt/anaconda3/bin:/usr/local/bin:/usr/bin:/bin
0 6 * * * /Users/vbaratham/claude/mlb_war_regression/refresh.sh >> /Users/vbaratham/claude/mlb_war_regression/refresh.log 2>&1

Install with crontab /path/to/repo/crontab. On macOS, grant /usr/sbin/cron Full Disk Access in System Settings → Privacy & Security.

Webapp

Static HTML/JS in webapp/. No backend. Loads CSVs from data/events/ via relative paths (or a configurable DATA_BASE URL in app.js for hosts other than GitHub Pages). Three view modes:

• All-time (per-season-sum): the headline.
• All-time (single-fit): legacy single ridge fit across all seasons, with a banner explaining the cross-era caveat.
• Each individual season (1910 → current): per-season leaderboard. The current season's entry shows " (through MM/DD)" with a WAR-over-time chart driven by daily snapshots.

Clicking any player row opens a modal with a season-by-season WAR table and a bar chart (per-season WAR) + line (cumulative). Other features: filter by name, multi-position chip filter, click-to-sort columns, dark mode, modern team abbreviations (NYY/NYM/LAD/etc. translated from Retrosheet's NYA/NYN/LAN).

Files

File	Purpose
download_all.sh	Walk 1910..current year, fetch any Retrosheet zips not on disk.
loaders/common.py	Source-agnostic per-event → half-inning aggregation.
loaders/retro.py	Retrosheet path (cwevent + park + roster parsing + auto-fetch).
loaders/statsapi.py	MLB Stats API path (game feed/live + roster API).
loaders/crosswalk.py	MLBAM ↔ Retrosheet ID mapping via Chadwick register.
build_dataset.py	Top-level builder; auto-routes years to retro or statsapi.
fit_ridge_all.py	Single all-time ridge fit (player + season FE + park FE + home).
fit_per_season.py	One ridge fit per season; emits long-format season_war.
make_views.py	Per-season-rate / peak-rate / position-z views from the single-fit table.
snapshot.py	Daily snapshot of current-season coefs + manifest.json regeneration.
refresh.sh	Cron entry point.
webapp/	Static frontend.
REPORT.md	Full methodology, calibration, and the cross-era discussion.