Does WAR Actually Predict Team Wins?

Every winter, a front office adds up projected WAR for forty players, adds the replacement-level baseline, and produces a win total. Every October, some of those teams finish six games off the projection in one direction or the other, and everyone acts surprised.

Both things are normal. They are, in fact, the same fact stated twice.

WAR was built to answer a roster question, not a standings question. FanGraphs' primer on the framework is explicit that replacement level is a chosen convention: it is set so that a roster of freely available players would win somewhere in the neighbourhood of 47 to 48 games over a 162-game season, and so that total league WAR reconciles with total league wins above that floor. Baseball-Reference's version makes the same reconciliation with different pitching and defensive inputs.

That reconciliation is the whole trick. Because the baseline is calibrated to league totals, league-wide summed WAR has to line up with league-wide wins. Nothing in that construction promises it lines up for the Brewers.

Team wins, in WAR's language, are: replacement baseline, plus position-player WAR, plus pitching WAR. Each player's contribution is a run total divided by roughly ten runs per win.

Three things about that arithmetic matter more than they get credit for.

The components are added as if independent. Bat, glove, legs, and position are summed with no interaction term, even though a defensive alignment is a team property and a lineup's run scoring is not the sum of nine solo performances.

The two public versions disagree with each other on purpose. fWAR grades pitchers on FIP — strikeouts, walks, home runs — while bWAR grades them on runs actually allowed, adjusted for the defense behind them. Baseball-Reference publishes a direct comparison of the two frameworks. A rotation can be worth several wins more in one system than the other, which means "the team's WAR" is not one number.

Context is stripped deliberately. WAR does not know that the home run came with the bases loaded. That is a design decision that makes the metric more predictive going forward and less descriptive of what already happened — and team wins are, by definition, a description of what already happened.

The residual is not random noise scattered evenly. It clusters in identifiable places.

Sequencing. A team that hits with runners on base outperforms its component stats; one that does not, underperforms. Neither is in the model, and neither reliably repeats.

Bullpen leverage. Relievers pitch a small number of innings that swing win probability enormously. WAR applies a leverage adjustment, but the year's actual distribution of high-leverage outcomes is far coarser than any season-total adjustment can capture.

Playing time. Preseason WAR projections assume a distribution of plate appearances and innings. Injuries redistribute them to worse players. The projection was not wrong about the players; it was wrong about who would play.

Defensive uncertainty. The fielding term carries the widest error bar of anything in the equation — a point we work through in full in the WAR explainer. Aggregated across a roster, those errors partially cancel. Partially.

A version of team WAR that reconciled to individual clubs — not just the league — within two games in most seasons, using only preseason inputs, would mean the residual is model error rather than genuine in-season variance. Public tracking data on defensive positioning and reliever deployment is the most likely source of that improvement, and it would move this from "useful with a range" to "usable as a forecast."