How we project MLB lineups and start probability
This page explains how our numbers are produced. Right now that means one thing: a deliberately simple baseline, tested on the 2025 season. Every figure below is a backtest of that baseline, not a live track record.
What we estimate
- Start probability: for each player on a team, the chance that he is in the starting nine for a given game, shown as a percentage.
- Batting order: the spot (1 to 9) where each projected starter is most likely to bat.
- Probable starting pitcher: the pitcher most likely to start the game for each team, and how confident we are.
All of these are our own estimates. They are not official information, and they can be wrong. The official lineup is the one a manager hands to the home plate umpire before the game; see when MLB lineups are released.
The baseline
Before building anything clever, we built something simple and measured it. A baseline tells us how hard the problem is and gives every later model a bar to clear. If a more complex model cannot beat it, we will not use that model.
The baseline uses only historical data from Retrosheet: who started each game, at which position, in which batting slot, and who the starting pitchers were, for the 2023, 2024 and 2025 seasons. For each game in the 2025 season, the baseline was allowed to see only games played before that date.
Starting pitcher rule
Look at the pitchers who started the team's last five games. Among those with at least four days of rest, choose the one whose most recent start is the oldest. In the second game of a doubleheader, take the next candidate. We also tested an even simpler rule for comparison: "whoever started five games ago".
Lineup and batting order rule
Look at the team's last 10 games against a starting pitcher who throws with the same hand as today's opposing starter. The nine players with the most starts in those games are the projected lineup (ties are broken by starts in the team's last 10 games overall). The nine are then ordered by their average batting slot in those games. A player's start probability is simply the share of those games he started. We also tested windows of 20 and 30 games.
Because the lineup rule depends on the opposing starter's handedness, we tested it two ways: (A) using the hand of the pitcher who actually started, which is close to the real situation when the probable pitcher has been announced, and (B) using our own predicted opposing starter.
How the baseline scored on 2025
The test covers 4,860 team-games: 2,430 regular-season games, counted once for each team.
| Measure | Result |
|---|---|
| Starting pitcher correct, rotation rule | 67.8% |
| Starting pitcher correct, "started five games ago" rule | 66.4% |
| Opposing starter's throwing hand correct (from our predicted starter; always guessing right-handed would have been right 73.7% of the time) | 86.4% |
| Players in the projected nine who actually started, average out of 9 (A, 10-game window) | 7.15 |
| Same, using our predicted opposing starter (B, 10-game window) | 7.04 |
| Same, 20-game window (A / B) | 6.82 / 6.72 |
| Same, 30-game window (A / B) | 6.57 / 6.49 |
| Brier score for start probability (scenario A, 10-game window; lower is better) | 0.153 |
| Brier score of a constant base-rate forecast on the same players | 0.250 |
The Brier score is the average squared difference between the probability we gave and what happened (1 if the player started, 0 if not). It was computed over every recent starter for the team (anyone who started in its previous 30 games or in the games the projection was based on), plus anyone who actually started, which averaged about 17 candidates per team-game. The comparison forecast gives every candidate the same probability, equal to the overall share of candidates who started. The base rate was computed over the whole season, so the comparison forecast had the benefit of hindsight.
Exact batting slot
| Slot | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 |
|---|---|---|---|---|---|---|---|---|---|
| Correct | 70.4% | 55.7% | 51.9% | 44.5% | 32.2% | 26.1% | 23.6% | 25.3% | 39.8% |
The leadoff spot was the easiest to get right and the middle-to-bottom of the order (slots 5 to 8) the hardest.
By month
| Month | Mar | Apr | May | Jun | Jul | Aug | Sep |
|---|---|---|---|---|---|---|---|
| Starting pitcher | 9.7% | 76.2% | 74.7% | 72.2% | 65.9% | 68.5% | 58.2% |
| Lineup, of 9 | 5.32 | 6.30 | 7.01 | 7.04 | 7.14 | 6.77 | 6.91 |
Both rules lean on recent games, so they do poorly in the first days of a season, when the most recent games are from the previous year. That is exactly where a better model has to add information.
What the baseline does not know
- Injuries, trades, call-ups and other roster moves. A player who just landed on the injured list can still be projected to start.
- Announced probable pitchers and posted lineups. The baseline only uses past games.
- Rest patterns such as day games after night games, catcher rest, or doubleheaders, beyond what shows up in recent starts.
- Its start probabilities are plain frequencies, and they are overconfident at the edges: in 2025, players it gave 0% still started 9.0% of the time, and players it gave 100% started 93.9% of the time. The full calibration table is in our guide to what a start probability means.
- It has only been tested on the 2025 regular season. At the time of writing, Retrosheet's data runs through 2025.
What comes next
We are working on a model that adds the information above. Before any new model is used on the site, we will describe it on this page and show its backtest next to the baseline's, labelled as a backtest. Live projections for the 2027 season will need current-season data, and we will name that source on our data sources page before we use it.
How we will grade live projections is described in our accuracy policy.