---
name: Raw Kalshi-vs-Poly CLV gap size is NOT predictive without the band gate
description: Tested whether the magnitude of disagreement between Kalshi and Poly (both measured vs Pinnacle fair) predicts profit on its own, dropping the [5,12) band requirement entirely. Non-monotonic buckets, OOS sign-flips at every threshold, and the one "significant" cell is 8 longshot bets. Confirms the band conditions do real selection work.
type: project
---

User asked 2026-07-16 (prompted by seeing an unusually wide Kalshi-vs-Poly price gap on Necaxa v Atlante, neither side actually cheap vs Pinnacle): does the SIZE of the K/P CLV gap predict profit, regardless of whether either leg clears the [5,12) band? Tested via `E:\soccer\research\kalshi_poly_gap_size.py` on poly_backfill (643 fixtures with a resolvable gap). Bet the cheaper-vs-Pinnacle leg unconditionally, sorted/bucketed by gap = clv(cheap) - clv(other).

**Verdict: NO edge from gap size alone.**
- **Disjoint buckets are NOT monotonic** (the tell): [0,2)% -4.0%, [2,5)% +4.2%, [5,8)% **-24.1%** (breaks the trend), [8,12)% -8.6%, [12,20)% +77.5% (n=11 only).
- **Cumulative thresholds collapse to longshot lottery cells** as T rises: n 643->83->33->8 while implied odds drop toward ~7%. The one "significant" result (gap>=15%, n=8, ROI+144%, z+2.06, p=.02) is 8 bets at ~14:1 payout — one win explains it. Same shape as the round-robin/sub-band traps already catalogued.
- **OOS 60/40 split kills every threshold**: gap>=0% TRAIN -8.2%->TEST +9.6% (sign flip); gap>=5% TRAIN +5.6%->TEST -23.9% (sign flip); gap>=8% TRAIN -12.9%->TEST +64.9% (sign flip). No threshold holds sign across the split.

**Why this matters:** this sits between two other negative results and closes the loop. [[project_pure_xbook_deviation]] found dropping Pinnacle entirely (fair = mean of the two books) is dead. This test keeps Pinnacle as each leg's reference but drops the frozen strict-xlow rule's band gate (cheap leg IN [5,12), other leg OUT of it) — also dead. So the band isn't an arbitrary cutoff riding on the coattails of a real continuous gap-size effect; it's doing real selection work on its own. A wide K-vs-P disagreement (like the Necaxa/Atlante case that prompted this) is NOT itself a signal unless the cheap leg's own CLV vs Pinnacle happens to land in band.

**How to apply:** Don't loosen or replace the xlow band conditions with a raw gap-size threshold. When a user flags a wide cross-book price gap, check whether either leg is actually cheap vs Pinnacle (not just whether the books disagree) before treating it as interesting.