The Lab · The math · updated 2026-10-22
Signals that predict the line, not the winner
Two of our signals predicted where the closing line would go, at more than three standard errors. Adding either one to our probabilities made them worse. Here's how both can be true, and what a line-predicting signal is actually worth.
A signal can see where the price is heading without knowing anything the final price won't. Two of ours led the closing line by about half a point, and adding them to the closing probability made it worse. Their only value is timing — bet earlier, not differently.
On September 5 we scored two baseball signals the same way we score everything: on every graded game, against the market. Both came back with two results that look like they can't both be true.
| Signal | Games | Led the closing line by | z | Applied to the closing probability | z |
|---|---|---|---|---|---|
| Line movement | 338 | +0.47 points | +3.68 | Brier +0.00136 (worse) | +2.01 |
| Recent form | 341 | +0.39 points | +3.13 | Brier +0.00144 (worse) | +1.14 |
Each signal predicted where the price was going, by more than three standard errors. And each one, added to our published probability, made the forecast less accurate — the line signal by about two standard errors, the form signal in the same direction, short of two.
Both are true. The reason is the most useful idea in this whole subject.
What "leads the closing line" means
When a signal fires in the morning — the line has started moving, or a team's last five games lean one way — we record which way it points. Then we watch the market's own probability from that moment to the start of the game. On average, the market moved the signal's way: about half a point of win probability by the close.
That's a real fact about prices. The market drifts toward what the signal saw. (The line movement test, in full.) The form signal knew something the morning price didn't, too — which is more than our first look at it could say.
Why it makes the forecast worse
Our board's probability isn't set once in the morning. It follows the market all day, so the number graded at the end is the closing price, fee removed.
By then the market has already moved. Whatever the signal knew at 9 AM, the price has absorbed by first pitch. Adding the signal's nudge on top of a closing price counts the same information twice — momentum extrapolated past the point where it already played out.
Think of a forecast that says "it's been warming all morning, so add two degrees." At 8 AM, that's useful. At 3 PM, when the forecast already reflects the warm morning, adding two degrees just makes it wrong.
That's the whole puzzle: the lead is measured from the signal's first firing; the Brier score is measured on the latest probability. The signal is early, not smart. Once the price catches up, it has nothing left to add.
What a line-predicting signal is worth
Timing — and only timing. If you were going to bet a side anyway, and the line is moving toward it, betting earlier gets you a slightly better price than waiting: about half a point of win probability, on average. By September 20, on 862 baseball games, the line signal's lead held at +0.36 points, z +4.96.
Half a point is real, and it doesn't pay for itself. The fee on a two-way market runs about 1.5 to 2.25 points per side; the arithmetic leaves a follower about 2.6% behind on every stake. A later test on far more games found the same thing: the side the line moved toward did win a little more often than the morning price said, and the return at the best price roughly paid back the fee and no more. (Seven ways we tried to beat the market, idea 5.)
So the advice is narrow: a moving line is a reason to bet sooner, never a reason to bet.
What we do with them
Both signals stay in shadow: computed on every game, scored on every game, applied to nothing you see. A signal becomes a candidate to change a published probability only when it makes that probability more accurate — a better Brier score, paired game by game, by more than two standard errors, on games scored as they happen. Leading the line doesn't qualify. Neither of these comes close: both point the wrong way.
For the numbers people: the two measurements
Lead. For each game where the signal fired, take the direction it pointed (+1 toward the home side, −1 away) times the change in the market's fee-free home probability from the first firing to the close, in points. The mean is the lead; its standard error is the sample standard deviation over √n. A control splits the games by direction: if the market simply drifted one way all day, both groups would move together. For the line signal they pulled apart at z +3.70; for form, z +2.96.
Accuracy. Score the latest probability with the signal's adjustment applied and without it, game by game: dᵢ = (pᵢ + δᵢ − yᵢ)² − (pᵢ − yᵢ)². The mean of d is the Brier change (positive = worse), with standard error sd(d)/√n. If the market has already moved by the information the signal carries, δ is pure displacement, and a displacement away from a calibrated probability can only add error on average — which is what both signals show. (How Brier works.)
The two numbers answer different questions — "where does the price go?" and "is the final price wrong?" — and a signal can say yes to the first and no to the second.
What this doesn't tell you
That no signal can beat the close. It tells you these two don't, and why a signal built from public prices and public results usually won't: the closing price has seen them too. A signal with an input the market lacks would be a different test.
That the closing price is the truth. It's the best forecast we've found, not a perfect one. In four sports and twenty thousand games, it lands within a point or two of what happens — close enough that beating it takes evidence we haven't found.
Anything outside baseball. Both measurements are one sport, one season. The same idea runs through everything we test, but the numbers above are baseball's.
Model output is informational and entertainment content, not betting or financial advice. If you bet, bet what you can afford to lose.
Log what you actually bet — units and dollars — and see it graded beside a model that publishes every call it makes, wins and losses alike. Free, no card.
Start a free tracker