World Series Forecast Methodology

Methods and sources for “What the Market Knew.” The forecast covers the 2026 MLB postseason and uses only information available at the end of the regular season.

1. Data and information boundary
2. Method 1: simulation
3. Method 2: machine learning
4. Starting pitchers and bullpens
5. Backtest
6. Market comparison
7. Where each number comes from
8. Figures
9. Limitations

Data and information boundary

The 2026 forecast uses the published AnalytIQ MLB team ratings (method mlb-team-ratings-raw-v1, batch e52ea2ef) as of October 6, 2026. The batch covers 2,429 of the 2,430 regular-season games through September 27; one game was cancelled. These ratings measure run scoring and run prevention against the league rate, without park or opponent adjustment. Every 2026 regular-season starting pitcher and relief appearance comes from AnalytIQ’s game data, with public MLB box scores filling the games it lacked; where both sources covered a game, they matched exactly.

History comes from Retrosheet game logs and pitcher lines for 2015 through 2025. 2020 is excluded because of its 60-game season and one-off postseason format. Historical seasons are rated with an exact reproduction of the AnalytIQ method, which reproduces every field of the published 2026 ratings. The postseason brackets rebuilt from regular-season standings match every real pairing and home-field team in all ten seasons. The information used here was obtained free of charge from and is copyrighted by Retrosheet. Interested parties may contact Retrosheet at 20 Sunset Rd., Newark, DE 19711.

Method 1: simulation

Each team’s strength is its Pythagorean winning percentage from runs scored and allowed, regressed toward .500. The amount of regression comes from split-half reliability: each team-season is split into alternating games, the halves are correlated, and the Spearman-Brown formula gives full-season reliability. A game’s probability combines the two teams’ strengths with the log5 method, an odds factor for home field and the starting-pitcher adjustment below. Series are played game by game in their real formats: a best-of-three Wild Card round at the higher seed, a best-of-five Division Series and best-of-seven League Championship and World Series, with World Series home field to the better record. The reported odds come from 200,000 simulated postseasons, each drawing every team’s true strength from its rating uncertainty.

Method 2: machine learning

A logistic regression predicts single games from both teams’ ratings as of the day before each game: Pythagorean strength, offense and defense, record against Pythagorean expectation, games played and both starting pitchers. It was trained on about 22,000 regular-season games from 2015 through 2025 and chosen over a boosted-tree model on held-out late-season games before any postseason was scored. Its game probabilities run through the same bracket as Method 1, and the reported odds are exact bracket probabilities.

Starting pitchers and bullpens

A starter is described by strikeouts minus walks per batter faced and earned runs per 27 outs, each relative to his own team’s rotation and shrunk toward the league average with fixed weights chosen before fitting. Postseason starters are not known in advance, so each team gets a projected rotation: pitchers with at least two starts in its final 30 days, ordered by the fitted effect, with the ace starting Game 1 of every series. Because starters pitch fewer postseason innings, the effect is scaled by the ratio of their postseason to regular-season share of innings. Bullpens were measured the same way against league relievers, both as whole bullpens and as each team’s six most-used relievers in its final 30 days. They added no measurable signal and are not part of the headline forecast.

Backtest

For each season from 2015 through 2025, except 2020, every parameter and model was fit on the other nine seasons, ratings were frozen at that season’s end and the real bracket was replayed. Scores are log loss on late regular-season games, postseason games and postseason series, the probability given to the eventual champion and paired differences with 95 percent intervals.

Market comparison

Game lines are closing moneylines stored by AnalytIQ from The Odds API: for each book, the last no-vig price before the scheduled first pitch, with the median taken across books. Games whose closing prices were more than a day old, mostly postponements, are excluded. The information test fits logit P = logit(market) + b × (logit(model) − logit(market)) by maximum likelihood; a weight b near zero means the model adds nothing to the market price. World Series futures are no-vig prices from The Odds API’s historical outrights at 12:00 UTC on the day after each regular season ended. Each book that quoted the whole playoff field is normalized over it, and the median is taken across books. The provider has no World Series futures for 2021, so futures cover 2022 through 2026.

Where each number comes from

The analysis is in the AnalytIQ repository, services/mlb-research, with its generated report at docs/validation/mlb-world-series-2026/REPORT.md. A copy of its results is in this repository at research/mlb-world-series-2026/results.json, with provenance in that folder’s README. scripts/insights/world-series-figures/build.py reads that file and writes research/mlb-world-series-2026/key_numbers.csv, the source of this table. Sources below are paths inside results.json. The training-set size of about 22,000 games is stated in the generated report.

Figures

All four figures and the share card are generated by scripts/insights/world-series-figures/build.py, which reads every value from research/mlb-world-series-2026/results.json; no figure value is typed by hand. Figure 1 plots each playoff team’s title odds from both methods and the futures. Figure 2 shows the paired change in late regular-season log loss from adding starters, and from adding bullpens on top of starters. Figure 3 shows the probability each method and the futures gave each eventual champion, with a uniform pick for reference. Figure 4 shows each model’s 2026 log loss minus the closing line.

Limitations

The ratings are team-level measures of run scoring and prevention. Injuries, lineup changes and roster moves after the end of the regular season are not modeled, and the projected rotation is a rule, not a prediction of managers’ choices. The market comparison covers one season of game lines and four completed seasons of futures, which is enough to show the market predicting games better but not enough to grade title odds. The backtest fits each season on the other nine, including later seasons, and assumes the relationship between ratings and outcomes is stable over the period. Historical ratings come from Retrosheet and the 2026 ratings from MLB data; only the 2026 reproduction is checked directly.

Return to “What the Market Knew”