Strikeouts at the book's own line

Plain English

A book sets its strikeout line near its own median, so at that line a coin flip is the honest “know nothing” answer. On the live record, our model does no better than the coin at the book’s line, and the book does better than both.

Worse, the model’s biggest disagreements with the book point the wrong way. When it was 10 or more points higher than the book on the Over, the Over hit less often than even the book said.

So strikeout gaps are no longer shown as plays. The strikeout projection, its full distribution and the Strikeouts board are unchanged. Should you trust a strikeout “edge” from us today? No. Should you trust the strikeout count? As much as before.

What the record says

Every starter from September 17 to 24 with a de-vigged opening price, graded after the game. Brier score at the book’s main line, lower is better.

Rows	Starts	Model	Book	Coin
All settled rows	160	.259	.243	.250
Written before first pitch	138	.254	.242	.250

• Model minus coin: +.004, with a 95% interval of −.012 to +.018. No better than a coin.
• Model minus book: +.013, interval −.002 to +.026. Behind the book, though not by a margin we can call certain yet.

Where the model disagrees

Where the model sat	Starts	Model said Over	Over hit
10+ points above the book	28	65.5%	42.9%
Within 10 points	96	51.3%	50.0%
10+ points below the book	14	31.6%	42.9%

The level is fine: the model served 4.83 strikeouts per start against 4.85 actual, and its projection ranks pitchers about as well as the book’s line does. The two agree closely (correlation .83), so most disagreement between them is noise. The problem is the shape of the distribution: a near-Poisson tail turns a 0.8-strikeout disagreement into 15 points of probability.

The likely fix is to model how long the starter stays in the game (pitch counts, rest, a short leash) as a distribution of its own, rather than a single number. That work is next.

What changed

• Wherever a strikeout gap is shown, it now reads Pass, with this reason.
• Two defects in the ledger were fixed while measuring this. The graded model probability had gone stale on 11 of 160 rows; it is now re-read from each row’s own frozen prediction. And 22 rows had been written after first pitch; they are excluded from the pre-game numbers above.

When plays come back

Automatically, not by judgment: when the model beats a coin at the book’s line with the 95% interval clear of zero, on 100 or more pre-game starts over 7 or more dates, under the engine that is serving. The current numbers are on the strikeout model card.