---
name: project-mm-inplay-verdict
description: Quoting IN PLAY needs ~89% completion to break even — measured, not asserted. In-play jumps average 15.6c and persist 63% of the time, against a completed book worth ~2c. The thin in-play queue is thin because other makers pulled, not because it is unexploited.
metadata:
  type: project
---

Asked 2026-08-22: open a NEW book every 10-15 game minutes after kickoff, so a
game gets several independent shots instead of one. Measured with
`research/mm_inplay_swings.py` (207,425 pre-match and 118,657 in-play moves).

## The size difference

| | median | p90 | p99 | moves >=5c |
|---|---|---|---|---|
| pre-match | 0.00c | 0.00c | 1.50c | **0.10%** |
| in-play | 0.50c | 3.50c | 25.00c | **7.12%** |

In-play p99 is **16.7x** pre-match, and a >=5c move is **70x** more frequent.

## The poison, on 3,373 in-play jumps of >=5c

- mean jump **15.6c**
- price **held or kept rising 63%** of the time after the jump
- a completed book nets **~2c** ([[project-mm-completion-arithmetic]])

A resting ask is lifted BY the jump, at the pre-jump price. You do not fill
despite the goal, you fill BECAUSE of it — the definition of adverse selection.

## The break-even

    p * 2 = (1 - p) * 15.6   ->   p ~= 89% completion required in play

Best completion observed ANYWHERE is 82.6%, at a heavy volume gate, PRE-MATCH.
With 7% of in-play ticks moving >=5c it would be lower, not higher. So the idea
fails by a wide margin, and it fails on a measured number rather than the
assertion previously sitting in `mm_maker_live.books_in_window`'s docstring.

## The trap that makes it look attractive

Fill rate by posting window (n=1080 resting episodes) makes in-play look good:

| window | legs | fill rate | median queue ahead |
|---|---|---|---|
| in play | 87 | 18.4% | **886** |
| 0-0.5h | 91 | 25.3% | 8,233 |
| 24h+ | 324 | 6.8% | 2,667 |

The in-play queue is 5-10x THINNER than anywhere else. That is not unexploited
opportunity — **the queue is thin because the other makers withdrew.** Being
alone in front of goal risk is the compensation, not an edge.

## What would survive

A COMPLETED book is outcome-independent, so it does not care who scores. The
only in-play version the arithmetic allows is filling all three legs
near-simultaneously — which means crossing the spread, paying the taker fee
(4.4% of stake, [[project-xlow-spread-verdict]]) that making currently avoids
entirely. That is a different strategy, not this one.

## Byproduct worth acting on separately

30% of resting episodes (324 of 1080) sit in the **24h+** window at a 6.8% fill
rate — the worst of any window — because `WINDOW_H` was raised 12->24->48->72 on
2026-08-17. Concentrating nearer kickoff is a one-line change and untested.

Re-run: `python research\mm_inplay_swings.py`.