Keeping the core together

The question

"They kept the core together" is offered every offseason as a reason to trust last season's team, and "they lost half the roster" as a reason to discount it. Both sentences assume that how much of a roster returns says something about how much of its strength returns. This study tested that assumption in three leagues.

What was tested

Continuity was measured from the game log and the current roster: the share of last season's minutes (NBA), plate appearances and batters faced (MLB) or attempts (NFL) that was produced by players still on the roster, with arrivals bringing their own history from their previous club. The claim, written before the run, was that carrying last season's rating into the new season in proportion to continuity beats a flat discount at predicting early-season results. It was tested leave-one-season-out on five NBA, five MLB and five NFL seasons. A follow-up let the data choose the direction of the effect rather than assuming it.

A later claim tested the other way of asking the question, in the NBA only: instead of counting how much playing time returned, sum the measured impact of each player expected to take the floor, so that losing the best player costs his number and losing the bench costs almost nothing. The expected lineup is the players who appeared in the team's previous game, at their minutes, which from a team's second game on is its roster as it actually plays. For the first game of a season it is last spring's final game, because no season before 2026-27 has a record of a roster as it stood on opening night: a player who left over the summer still counted for his old team for one night, and an arrival did not count until the second. Pricing that first game from a roster that peeks a week ahead instead changed the error by two thousandths of a Brier point. The sum was scored against a flat discount on 1,161 early-season games, with the 2025-26 season held out.

What turned up

Continuity by playing time carries nothing. Weighting the carryover by it made early-season prediction worse in all three sports, by 0.7% to 5.4%. In football the relationship ran backwards: last season predicted low-continuity teams better than high-continuity ones (rank correlation +0.42 against +0.25), which fits churn concentrating on rebuilding teams whose direction persists whoever left. Letting the data pick the sign lost or tied in every sport, football worst at 8.4%, with the coefficient flipping between sports, which is what noise being memorized looks like.

Who takes the floor, weighted by impact, carries a lot. Summing last season's per-player impacts over the expected lineup beat the flat discount in three of the four NBA seasons that have a previous season and on the holdout, with a forecast error of 0.195 against 0.218 at Brier. The market's number on the same games was 0.192.

What it means when you ask

An offseason or opening-weeks question about a team gets the roster's measured loss or gain, never a continuity percentage. "The player they lost was worth four points a game to that sum" is a fact about the NBA rating. "They return 80% of their minutes" explains nothing about any rating, in any sport, and DataBaller does not cite it as moving one.

What it does not mean

The roster-sum result is NBA only. A football lineup cannot be identified from seventeen games, and baseball's version of the question is the starting pitcher, so the MLB and NFL ratings carry last season forward with a flat discount and say so. A turned-over roster's continuity figure is not a judgment of its talent, in either direction. Coaching changes are invisible to every version tested.

The record

The continuity claim ran 23 August 2026 with backtest_team_rating.py and the lineup-sum claim on 7 September 2026 with backtest_player_rating.py. A follow-up on 13 September 2026 with backtest_roster_terms.py put four continuity terms beside the lineup sum on the same early-season games, a discount on a rating earned for another team, the lineup's returning-minutes share, tenure with the team, and shared minutes between every pair of players, and none improved it. Five seasons per sport means five offseason boundaries each. The lineup-sum result belongs to Power Rating v2, built and awaiting release, and meets its first unseen season in 2026-27. The full portfolio and its testing rules are on the decision metrics page.