Why aWAR
What I read before building Armside’s WAR, and what each complaint about WAR changed in it.
By Noah Mitchem · October 2026
Why build one
Fans go to FanGraphs and Baseball-Reference for WAR. Armside had no stat that answered “is he good.”
Baseball Savant publishes two parts of one, fielding run value and baserunning run value, but no WAR board. In Bradford Doolittle’s June piece on the state of WAR, MLB.com’s Mike Petriello said, “Nothing technically really stands in the way,” and “the main goal is quantifying all the cool stuff along the way.”
Armside doesn’t show or store data from FanGraphs, Baseball-Reference or Baseball Prospectus, so aWAR is built from Savant’s boards and Armside’s own pitch data.
What I read
Before building it, I went through the WAR literature. What shaped the design:
- FanGraphs’ WAR primer: “WAR is an estimate. You should not use WAR with the expectation that it is precise to the decimal point.” Baseball-Reference’s WAR explainer says not to take “any full-season difference between two players of less than one to two wins to be definitive (especially when the defensive metrics are included).”
- Baseball Prospectus. Colin Wyers put a standard error on hitters’ batting runs when he reworked WARP in 2013. Jonathan Judge wrote that a stat shown without its uncertainty “leaves users to guess at what that uncertainty might be,” and BP published DRC+ with a standard deviation beside it.
- openWAR, by Benjamin Baumer, Shane Jensen and Gregory Matthews. It says current versions of WAR depend “upon proprietary data, ad hoc methodology, and opaque calculations,” and it puts an interval on its own numbers by simulation. In their simulated 2012 seasons, Miguel Cabrera beat Mike Trout nearly 32% of the time.
- Tom Tango, who wrote in 2025 that there is “one WAR framework” with “many implementation choices.”
- statcast_xwar, the one community Statcast WAR we found. It takes its position and baserunning values from FanGraphs. Armside can’t.
- The arguments. Dave Cameron on why FanGraphs’ pitcher WAR uses FIP, against Baseball-Reference’s pitcher method, built on runs allowed. Catcher framing, which Baseball-Reference left out of WAR in 2018 and FanGraphs added in 2019. The 2025 AL MVP race, which Tango walked through as Aaron Judge’s bat against Cal Raleigh’s defense, with “a reasonable argument for either.” And Sean Smith, who built Baseball-Reference’s original WAR, in Doolittle’s piece: “Almost every time you see a difference between players, it comes down to the defensive number.”
What kept coming up
One number, to a tenth
Both sites say a small gap in WAR means little, then print one number with no range. aWAR prints a 90% range under every value. Aaron Judge’s 2024 reads 10.9, with 8.5 to 13.4 under it.
The batting part of the range has been tested. Across 1,389 hitter-seasons of 500 PA or more, split into first and second halves, 12.0% of the gaps fell outside it. A perfect range misses 10%. The fielding and baserunning parts haven’t been tested that way, and the methods page says so.
Defense
Defense is the noisiest part of WAR, and the systems disagree. Armside has one, Savant’s fielding run value, so aWAR shows it on its own line: Pete Crow-Armstrong’s 2026 carries +20.9 fielding runs. Its noise goes into the range, sized by how well Savant’s values repeat from one season to the next. Infielders’ repeat less than outfielders’ (a correlation of .447 against .698), so their ranges run wider. Tango’s September comparison with DRS found the same soft spot: Savant wins at catcher and in the outfield, and the infield is about even.
Each fielder is measured against the average at his position. Savant rates all outfielders on one scale, so in 2026 the average center fielder came out 5.6 runs above average over a full season. The position value already pays center field for being hard to play, so aWAR subtracts it.
FIP or runs allowed
Pitcher WAR splits on one question: how much of a ball in play belongs to the pitcher. Cameron wrote in 2010 that “there is no right way to do this, based on the statistics we currently have available.” Baseball-Reference measures “the value of the recorded performance.” aWAR’s headline uses FIP, because it repeats better from one season to the next, and a runs-allowed aWAR sits beside it. Its FIP is ifFIP, which counts infield popups as strikeouts. Emmanuel Clase’s 2024 reads 2.1 on ifFIP and 3.8 on runs allowed. FIP ignores contact quality and may undervalue relievers, and the methods page says that too.
Position values
FanGraphs’ position values are the ones Tango set in 2006. His August primer says the infield-outfield gap behind them has “crashed down to about 4 runs,” and FanGraphs’ library says its DH value “might be too negative.”
aWAR still uses FanGraphs’ values. Our first try at our own, from five seasons of our batting by position, failed a stability test: the DH value swung 3.1 times what sampling explains. The methods page now prints a second set beside FanGraphs’, from Tango’s bench method, which compares how each position’s bench players hit, on seven seasons of our data. For 2026 it puts the infield-outfield gap at 3.9 runs and DH at −8.8 runs per 162 games. Whether to switch is still open.
Parks, relievers and the fielders behind a pitcher
My first draft had no park factors and no reliever leverage. Park factors now come from Armside’s pitches: a club’s rate at home over its rate on the road, pooled over up to five seasons in the same park and pulled toward average by how steady each factor is from year to year. Hitters get a factor for the value of their batting events. Pitchers get one for each FIP event, so Coors costs a pitcher strikeouts and popups as well as home runs, and aWAR credits him for all three.
Relievers get half credit for leverage, as at both sites, from a table built on our own win probabilities. Clase’s 2024 gets +4.7 runs from it.
Baseball-Reference charges each pitcher a share of his team’s defense, by his balls in play. One community analysis says that “assumes that defensive performance is uniform among a team’s pitchers.” aWAR’s runs-allowed line charges each pitcher for the fielders who were behind him, from who was in the field for every pitch.
Double plays
aWAR has a double-play line. Baseball-Reference’s WAR has one too, and FanGraphs’ baserunning numbers carry none. Speed and hustle keep a hitter out of double plays. Slow runners and ground-ball hitters roll into them. A hitter gets credit for grounding into fewer than an average hitter would in his chances (a runner on first, fewer than two outs), at 0.40 to 0.45 runs each, about what Baseball-Reference charges. The line is smaller than baserunning: since 2016, among hitters with 400 PA or more, its spread is about two thirds of baserunning’s, and about half of it is luck, which the range carries.
Methods nobody can rebuild
That was openWAR’s complaint. aWAR’s parts are on every player page and in the Explorer, and every constant is on the methods page, with the twelve checks a season must pass before it’s written. Two inputs can’t be rebuilt, Savant’s fielding and baserunning values, and the page says so.
What it isn’t, and what’s next
aWAR will differ from fWAR and bWAR. Where it does, its parts and its range are on the page, and the methods page lists every place it departs from FanGraphs.
Still open:
- Position values: whether to switch to our bench values.
- Pitchers: a version that credits contact quality, and a blend of FIP and runs allowed.
- Double plays: counting only ground balls as chances, or an expected-double-play model from exit velocity, direction and sprint speed.
- Comparisons: a chart of the chance one player’s season beats another’s, built on these ranges.
Every change to the method adds a changelog entry naming each player whose aWAR moved by 0.5 or more.