⚾ MLB 2026

TestedThe Elo, the series odds and the bracket pass property tests on the live bracket; the regular-season sim has no invariant row and no backtest is published. About the tiers

Each club's rating starts from its run differential per game across the last two seasons (2025 and 2024), shrunk toward the league average (×0.62) because rosters turn over. This season's real run differential folds in at a weight that climbs with games played and dominates by midsummer - baseball gives you 162 games of evidence, and by August the standings have earned the right to outvote history. Run differential becomes a true-talent win percentage through the classic ten-runs-per-win rule, so a club outscoring opponents by a run a game reads as a .600 team. Those percentages are held as log-odds, which makes every head-to-head exactly the log5 formula plus a home-field term (53.5%, the long-run MLB home win rate).

Every one of the 0 games left on the actual schedule is then simulated 20,000 times, with each simulated season drawing every club's rating afresh from a distribution (σ 0.07 in log-odds) so the output carries the model's own uncertainty rather than pretending to none. The bracket is the real one: byes for the top two seeds, a best-of-three Wild Card round played entirely at the higher seed, a 2-2-1 Division Series and 2-3-2 Championship Series and World Series, with home games where they actually fall rather than a coin flip. Division ties are broken on record then head-to-head, an approximation of the full ladder. The market is not blended in at this point of the season: with this much of the schedule already played, the standings carry more information than the futures do.

Uncertainty shrinks as the season does - the spread of outcomes each simulated season draws from narrows week by week as fewer games remain, and widens back out for a club the stats and the market disagree about most. Each simulated season also draws one correlated error for the whole league, one for each division, and one for each club, rather than treating every game as its own coin flip - real seasons run hot or cold together, not independently.

The regular season has no game-by-game ledger, on purpose. The NFL plays sixteen games a week and each one is an event worth calling in advance; baseball plays fifteen a day and almost none of them decide anything, so the honest unit of prediction here is the season.

The postseason above is a real, graded ledger, and the market is the headline forecast. A sportsbook line comes first when one is posted (ESPN's moneyline, de-vigged); failing that, the Polymarket exchange. The run-differential model - the same regressed rating through log5 plus home field, priced at the actual site of each game (the wild card is entirely at the higher seed; the Division Series is 2-2-1; the Championship Series and World Series are 2-3-2) - only prices a game until a real line exists for it, and stays on as the yardstick the Ledger grades the market against. Series odds get the same treatment: the exact best-of-N recursion runs twice, once model-only and once substituting a real line for any remaining game that has one (shown as Market alongside Model). Bracket odds (reaching each round, winning it all) are still a 20,000-run Monte Carlo over the real reseeding rule, model-only; the Road to the World Series table adds a Market column from Polymarket's title futures where that team is still listed there. There is no pitcher data anywhere in this model - every rating is team-level.