We Fixed Where Lineup Slot Comes From. It Made Our Projections Worse.

We tested it, it did not survive the evidence, and the projections did not change. Here is the whole result, including what it does not prove.

Batting order matters. Everyone knows it: a hitter moved from seventh to second gains runs, and one moved out of the middle of the order loses RBI. So a projection system that guesses a hitter's lineup slot from his own projected rates, which is circular reasoning, looks like an obvious thing to fix. Feed it the slot he actually hit in, and the runs and RBI numbers should get better. That was the claim, and it had never been tested on its own.

We tested it. It failed.

The result

We compared four versions on 919 season-to-season player pairs across four seasons the model had never seen, using both mean absolute error (MAE, the average miss in runs or RBI) and rank correlation, or ρ, which measures whether the players are ordered correctly and ignores whether the overall level is too high or too low.

Using real slots moved rank correlation the wrong way for both categories: R −0.0064, RBI −0.0110. In the framing most generous to the layer, using slot data from the same season being projected, runs got worse in all four seasons. Strictly forward, using the prior season's slot, also 0 of 4. It lost the fight it was allowed to rig.

The slot term was doing nothing to begin with

Look at the first two rows. Deleting the slot term entirely produces identical ρ for both categories and identical runs MAE to three decimals. That is an identity, not luck.

Every pair in this sample required 250 plate appearances the prior season, so the guessed slot never leaves the 3-to-5 bucket: slot 3 for 849 players, slot 4 for 64, slot 5 for 6. The runs multiplier in that bucket is 1.00. So for runs the slot adjustment is arithmetically a no-op. The RBI multiplier there is 1.075, applied to essentially everyone, which is a flat 7.5 percent inflation that reorders nobody. Hence the identical ρ.

Why deleting the RBI term is not the free win it looks like

Dropping the RBI term appears to gain 1.151 in MAE. Do not act on that. It moves zero ranks, which means it is purely a level shift, and the test rig here has no step that corrects the overall level while the real engine does. A uniform 7.5 percent RBI inflation is exactly what a calibration step absorbs. That 1.151 measures a hole in the test, not a flaw in the projections. Trust ρ. On ρ, real slots lose.

Why accuracy hurt

The layer's own RBI gain was 0.434. Deleting the constant gains 1.151. So the part you can attribute to slot accuracy is −0.716. It subtracts.

The reason is double-counting. Last season's rate stats already contain last season's batting order effect. Applying next season's multiplier on top counts it twice. With the degenerate guess that is harmless, because a constant cannot double-count anyone differently from anyone else. Give it real, varying slots and the double-count becomes real and uneven. We also tested and killed the tidier theory that runs broke because accurate slots activate a badly signed 6-to-9 constant: the re-estimated value made it worse, −1.194 versus −0.825, and neutralizing it to 1.00 did not help at −0.875.

What survived, and what this does not settle

One variant improved a level-proof metric: divide out last season's slot effect before applying next season's. It gained R ρ +0.00566 against both baselines, positive in 4 of 4 held-out seasons, with a confidence interval of +0.00326 to +0.01002, and a shuffled control that inverts the gain to −0.00485. RBI unchanged. That passed a pre-registered gate.

It is also tiny, roughly 0.6 percent, in-season only by construction, and unvalidated inside the live engine. This does not prove the RBI 1.075 should be removed. It does not prove the strip version should ship. It says nothing about plate appearances, pitchers, or true in-season projection. And 54 percent of pairs changed multiplier bucket, so the harm was not confined to a fringe.

Nothing changed in production. The observed-slot layer stays off, and both routes to reviving it are now closed.

All articles · Redraft rankings · Dynasty rankings