---
name: Draws tested — no CLV edge; stay home/away only (2026-06-06)
description: Reconstructed draw signals (clv_kalshi_draw, same [5,12)% band/dedupe rules) and scored vs ftr=='D'. Draws are flat (ROI -1.8%, p=0.56) and the clean TEST slice is NEGATIVE. Edge is home/away-specific. Don't add draws.
type: project
---

User asked 2026-06-06 "is there any circumstance where we can bet on draws?" The live strategy is home/away only. Tested via `E:\soccer\_draw_check.py` (reconstructs draw signals with the SAME rules: clv_kalshi_draw in [5,12)%, first-band-entry, per-game dedupe, settle vs ftr=='D', odds=1/kalshi_draw_ask).

**Verdict: NO edge on draws. Stay home/away only.**
- Draws [5,12)%: n=242, WR 21.5%, ROI **-1.8%**, p=0.558 — coin-flip. (vs home/away same data/band: n=399, +14.4%, p=0.093.)
- **Clean train/test (first 60/last 40):** draw TEST slice = n=97, ROI **-9.6%**, p=0.69 — NEGATIVE out of sample. This is the decisive read.
- Tempting subsets are the usual in-sample traps: [8,12)% sub-band looks +36%/p=0.106 BUT [5,8)% is -10% (sub-band cherry-pick, same noise the probes flagged); league B1 shows +147%/p=0.023 but n=11 (would not survive the FDR that only E0 cleared on home/away).

**Why structurally:** the edge is a Kalshi UNDERPRICING of home/away longshots vs Pinnacle fair; the draw is a different, narrower-probability market (mean odds ~4.9) and is typically the tightest/most-efficient of the three outcomes — no mispricing to harvest.

**How to apply:** Don't add draws to SIDES. There's no conditional in current data that flips this — the only positive-looking draw cuts are sub-band/single-league noise and the OOS draw result is losing. Consistent with the standing theme (everything that isn't the core home/away longshot signal tests as noise). Re-runnable: `_draw_check.py`.

**Re-checked under STRICT XLOW 2026-07-13 (user request) — draws fail HARDER:** draw-as-exclusive-low (cheaper book's draw in [5,12) vs pin fair, other book's draw not in band) on poly_backfill: n=180, WR 18.9% vs 21.5% IMPLIED (below-market in TRAIN −9.6% AND TEST −8.9%), ROI −9.4%. Adding draw as a third candidate outcome dilutes the rule +24.7%→+21.7% (its 138 draw picks run −10.6%). So even the strongest selector picks losing draws — WR below implied means cheap exchange draws are cheap because the EXCHANGE is right and pin fair is stale, i.e. the draw permanently lives in xlow's "joint deviation" failure mode (no retail team-flow chases the draw; it's the professionally-priced leg). Closed under both selectors; don't re-open for a third.