MLBengine v14
PriorsSep 17, 202695,949 held-out plate appearances · 95 dates

Last season is the biggest untapped signal

How much should one month of at-bats move a projection? We tested the engine's pull-to-average amounts, its handedness splits, time decay and last season. Last season won by a distance.

The question

Every projection starts from a hitter’s and a pitcher’s record: how often each plate appearance ends in a strikeout, walk, out, single, double, triple or home run. Small samples lie, so the engine pulls them toward a starting point. How hard should it pull, toward what, and should recent games count more?

We scored every option walk-forward through 2026: each prediction uses only data from before its date, graded on the next two weeks of plate appearances.

What we found

  1. The pull-to-average amounts were already right. Estimating them from the data instead of setting them by hand changed nothing measurable. The engine trusts a hitter’s strikeout rate halfway after about 60 plate appearances (the data says about 46) and his home-run rate after about 170 (the data says about 200).
  2. A player’s own handedness splits made predictions worse. They came from raw results, and were trusted far too early. Counting the league-wide platoon gap once, and applying it to each player’s overall rate, was better than both the old splits and no splits at all.
  3. Recent games don’t deserve extra weight. Given a free choice of how fast old games fade, the data chose a half-life of 365 days: effectively no fading. What helped was keeping the whole season rather than cutting at 90 days.
  4. Last season is the biggest signal. Starting each player from his previous season, instead of the league average, was the largest single gain. It helped at every sample size, most for players with little this year, and still for regulars with 300+ plate appearances.

All together

Last season as the starting point plus the league platoon gap improved plate-appearance log loss by 0.0057 per plate appearance. The engine’s entire edge over a league-average guess was 0.0161, so this is roughly a third more skill in the rate layer. Strikeout, hit and home-run probabilities all improved, and their calibration moved much closer to ideal.

What shipped

The league platoon gap and the prior season both went into the engine in engine v11 on this plate-appearance evidence. Later engine versions were then measured through every market, as every version now is. Keeping the full season instead of a 90-day window is still on the list.

A note on hot and cold streaks: this is why our matchup pages show a hitter’s last five plate appearances as context, not as a reason.