---
name: Five strategy-improvement probes (2026-06-06) — two non-obvious findings, rest noise
description: Tested CLV sub-band, cross-book agreement, spread filter, favorite/longshot split, and sizing/shrinkage on the live Kalshi CLV strategy. Cross-book anti-confirms; Kelly under-bets longshots. Most else is in-sample noise at n=365/p=0.056.
type: project
---

Ran 2026-06-06 via `E:\soccer\_ideas_study.py` (reconstructs the 365-signal set from live_data keeping b365 CLV + kalshi bid; matches baseline 365/26.3% WR/+18.6% flat ROI/+67.8u).

**#2 CROSS-BOOK AGREEMENT — the standout, counterintuitive finding (act on this):** Requiring B365 CLV to ALSO be in [5,12)% leaves only 7 bets. Profit actually CONCENTRATES where B365 does NOT confirm — B365 CLV<=0 subset = n=340, +20.9% ROI; B365 CLV>0 subset is NEGATIVE. So the edge is a **Kalshi-specific mispricing, NOT a cross-book consensus.** Do NOT add B365 confirmation to the signal — it's anti-correlated and would gut the strategy.

**#5 SIZING — also non-obvious:** Current quarter-Kelly is very CONSERVATIVE (max drawdown only 4.7%, final +38.9%) and structurally UNDER-bets the longshots where the profit lives (Kelly off pin_fair sizes a ~20%-win bet near zero). Flat staking harvests them far better in-sample: flat 1% = +80%/DD 14.3%, flat 2% = +175%/DD 27.6%. BUT that is leverage on an edge that's only p=0.056 — if illusory it bleeds out. Verdict: keep the conservative sizing as insurance; the lever exists but don't pull it until the edge is better established. eighth-Kelly and 50%-edge-shrink both ≈ +17%/DD 2.6%.

**Noise / no-action (in-sample mining at n=365):**
- #1 CLV sub-band: [8,12)% looks much stronger (+46%, n=114) than [5,8)% (+6%, n=251), and [6,7)% even negative — a real-looking gradient but tiny per-bucket n; don't narrow the band without OOS/more data.
- #3 Spread filter: Kalshi spreads are tiny (median 1c, mean 1.3c). Only ~6 bets ever >4c and all lost. A light "skip spread>3-4c" filter nudges +18.6%->+19.9% — harmless, execution-sensible, marginal. Spread is NOT eating the edge (consistent with _spread_check.py).
- #4 Favorite/longshot: signal almost never fires on favorites (2 bets odds<2.0; 22 at p>=40%). Profit lives in 15-25% p_win / 4-8 odds. CANNOT de-risk by shifting to favorites — inherently a longshot strategy.

**How to apply:** The only finding worth wiring is the NEGATIVE one (#2: never require cross-book agreement). Everything else reaffirms the standing theme (round-robin + league-filter studies): at n=365/p=0.056 most tweaks are noise; collecting more data beats adding rules. Script kept at E:\soccer\_ideas_study.py.