The Innings Formula That Never Ran, And Why We Left It Broken

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.

Our innings projections have a piece of machinery in them that has never once fired. It was written to reshape every pitcher's innings forecast using prior workload, age, injury history and year-to-year volatility. It reads a field of pitcher data that does not exist under that name, so every row it checks fails. Not most rows. All of them: 9,266 rows parsed, 0 pass. Live, across 646 pitchers, the only innings adjustments actually applied are the three others in the chain.

The reasonable belief here is obvious, and I held it too. Prior innings, age, time on the injured list and volatility are exactly the things you would use to forecast innings. A projection system that ignores them is leaving accuracy on the floor, and the fix is one word in one line. So the honest question was: is there a free accuracy gain sitting there, switched off?

What we actually tested

Rather than fix it and see, we rebuilt the dormant formula exactly as written and ran it against real outcomes across five seasons of pitcher pairs, 2021 through 2025. Nothing was tuned. The formula weights prior innings at 0.5 for last year, 0.3 for two years back, 0.2 for three, then multiplies by penalties for age 30 and up, for heavy injured-list history, and for erratic innings totals. The result is blended half and half with the existing projection, and applied only when the resulting ratio lands between 0.6 and 1.4.

That guard turns away almost nobody. The block reaches 1,623 of 1,652 pairs. Only 29 are blocked, and none are missing prior history. Switching this on is not a targeted tweak, it changes nearly every pitcher on the board.

The result is not close

Spearman is rank correlation, how well the ordering of pitchers matches the ordering they actually finished in, where 1.0 would be perfect. MAE is average miss in innings. The revived version improves rank correlation in 1 of 5 seasons, improves the makeup of the projected top 30 in 1 of 5, and improves average miss in 1 of 5. Pooled average miss goes from 80.45 innings to 81.62, a 1.17-inning degradation. Mean rank-correlation change is −0.00066.

The single win is 2021. It wins on all three measures at once, and the four more recent seasons lose on nearly all of them. That pattern has a name in plain English: a formula built for an older workload world. Innings distributions in 2021 do not look like innings distributions in 2025, and the age and volatility penalties are calibrated to the former.

The strongest argument the other way

Here is the case against my read, made properly. Mean rank-correlation change is −0.00066, which is essentially zero. Four of the five losses are small. And 2022 is a tie to three decimal places. You could look at this table and say the formula is roughly neutral on ordering and only mildly worse on absolute innings, while adding real information the projection currently ignores. That argument is not stupid. It is just not enough to ship a change that touches 1,623 of 1,652 pitchers with nothing in place to catch a slow bleed. A 1.17-inning drift in average miss is invisible until it costs you an innings-limited streamer decision in September.

What this does not prove

It does not prove the idea is wrong. Prior innings, age, injury history and volatility are sensible inputs. What failed is this specific set of numbers against the current game. It also does not prove that a re-tuned version would fail, because we deliberately did not tune anything, and a fitted version would have to be tested on seasons it had never seen. And it says nothing about the pitcher skill term, which is separate, independently validated, and untouched by any of this.

What would change the answer

Three conditions, together. A re-tuned version would have to improve rank correlation in a majority of unseen seasons, move the projected top 30 closer to the pitchers who actually finished there, and not worsen pooled average miss. All three. A single-metric gate is exactly how a bad candidate once looked shippable.

The decision

The broken read stays broken. It has been quietly protecting the innings projections from a lever that would make them worse, and it is now marked as deliberately dead so nobody tidies it up on sight. For you as a reader, nothing on the board changes. The practical takeaway is narrower and more useful: our pitcher innings forecasts are not using a prior-workload decay curve, by choice, because the version we have would cost 1.17 innings of accuracy per pitcher. If you want age and injury-history discounts on innings, apply them yourself, case by case, on the arms where you have a specific reason.

All articles · Redraft rankings · Dynasty rankings