TRAILING_LATE at 10.8% over 344 bets: live markets reprice but don't read

The argument in this series is easy to state and hard to prove. In-game baseball markets are repriced by machines that update on the score, the inning and the outs. Almost nobody reads what is happening on the mound. Our proven signals are the case for that view. Our own record shows where the case still has gaps, and we say where below.

First, game state. Every price in this piece is a pregame number from the 2026-10-06 slate. None of them is a live in-game price, and none should be read as one.

What we found

The headline result is TRAILING_LATE: 10.8% ROI over 344 graded bets. It is the largest sample in our proven-signal book, and it is holding out-of-sample. Next comes VELO_DROP at 9.9% ROI over 201 graded bets, then SPIN_VS_AVG at 7.5% over 240 and THIRD_TIME_THROUGH at 6.2% over 70.

Most of these signals have something in common. VELO_DROP, SPIN_VS_AVG and THIRD_TIME_THROUGH are all reads on the pitcher's condition. A live pricing engine sees the score, the inning and the base-out state. It does not obviously see that a starter's velocity is down or that the spin is off its average. That gap, between what the repricer takes in and what you can see on the field, is where we think the remaining edge is.

TRAILING_LATE is the exception, and the most interesting of them. It is a game-state signal, the exact thing machines are supposed to price well. Our read is that late-game prices on the trailing side carry a behavioural tilt that the automated update does not remove. That is a hypothesis about the cause. The return is measured. The reason for it is not.

The trends

The pregame market on 2026-10-06 shows the opposite picture. Before first pitch, money and attention are concentrated, loud and one-sided.

- Concentrated money landed pregame. The top campaign was $124,770 on SD ML in 130MILSD, at 99% one-sided. - The same game's total drew a crowd. $79,248 went on Over 8, at 78% one-sided. - In 800LADATL the money leaned Under. $101,015 went on Under 7, at 84% one-sided. - The LAD line drifted. It moved from 56.0c to 57.5c (+1.5) over the last 20h. - Dustin May (MIL) is trending the wrong way. 4.97 ERA over the last 5 and 8.87 over the last 3, against a season mark of 4.70. - The opposing arm is running hot. Nick Pivetta (SD) has a 1.27 ERA over the last 5, with 9.3 K/9.

Across the 2 games on the board, tracked whale money totalled $418,679 across 5 campaigns. The biggest single ticket on the board was $32,064. This is the pregame market being read, and read hard.

The biggest total edge we priced was LAD @ ATL: model fair 6.2 vs line 6.5. That sits on the same side as the 84% one-sided Under money. Pregame, the model and the concentrated money point the same way.

What it means

The contrast is the argument. Pregame, a reader can see $418,679 of tracked money and an arm like May's that has gone from a 4.70 season ERA to 8.87 over the last 3. Once the game starts, that same arm's velocity and spin become the live information. The market's update, though, is built on the scoreboard. If May's slide shows up as a velocity dip in the middle innings, VELO_DROP is the signal built for that moment. Its 9.9% over 201 graded bets says that reading the arm has paid.

Now the part our record does not yet support. "Read by nobody" is a claim about the market, and nothing in this facts block measures it directly. What we have are signal returns. We do not have a study of how fast live prices absorb a velocity drop. THIRD_TIME_THROUGH at n=70 is also the thinnest sample in the book. Its 6.2% is promising, but it carries the widest uncertainty of the group.

The next tests target those gaps:

- Speed of repricing. We are timing how quickly live prices move after a VELO_DROP or SPIN_VS_AVG trigger fires. - Sample size. We are growing THIRD_TIME_THROUGH's graded sample until it can stand next to TRAILING_LATE's 344.

Our pipeline indexes 428 data points across 9 agents to feed that work. Every signal is judged against a 52.38% breakeven win rate. We publish probabilities and their track record, not predictions about what a pitcher or team will do. Nothing here is a guarantee. 21+ where legal.

What would change our mind: evidence that live prices move on velocity and spin before our triggers fire would mean the pitcher-state edge is being read after all.