Projecting Pitchers with the Marcel Method

Tom Tango's Marcel system was built to be the simplest projection that still works: weight recent seasons more, and regress everything toward the league average. It is famous for being hard to beat by much more complicated systems. BullpenEdge uses a version of it for every pitcher and hitter.

Weight the seasons

The current season counts 1.25, last season 1.0, two years ago 0.8 and three years ago 0.6. Every season is measured relative to its own league average first. That matters: the 2023 rule changes lifted scoring and stolen bases, and 2022 had the deadest ball of the decade, so a 4.00 ERA did not mean the same thing in both.

Regress by how noisy the stat is

Some pitching skills show up quickly and some take years. The model adds a fixed amount of league-average performance to each stat before dividing: about 70 batters faced for strikeout rate, 170 for walks plus hit batters, and 1,000 for home runs. So a pitcher with 150 batters faced keeps most of his strikeout signal but almost none of his home-run luck. Runs allowed per batter are the noisiest of all and, in testing, added nothing once strikeouts, walks and Statcast quality of contact were known — so the model does not use them.

Add Statcast, but only from last season

Expected wOBA (xwOBA) and CSW% (called strikes plus whiffs) from Baseball Savant describe how hard a pitcher is to square up. The model uses the previous season's leaderboard only, so a projection never contains information that was not public at the time.

How long will he pitch?

A great starter who lasts four innings hands the game to the bullpen. Expected innings come from a pitcher's recent starts, weighted toward the latest and regressed toward the league norm (around 15 outs per start in 2026, down from 15.5 in 2019). That share decides how much of the game the starter's quality counts for and how much falls to the relievers. Every projection is listed on the Pitcher projections page.