""" Does the RAW SIZE of the Kalshi-vs-Poly CLV gap predict profit, with NO band gate at all (i.e. bet the cheaper book regardless of whether either leg clears [5,12) vs Pinnacle)? NO. Distinct from two prior negative results this complements: - kalshi_poly_pure_deviation.py: fair = mean(kalshi,poly), no Pinnacle at all -> dead. - forward_test_unique_book.py fire_xlow(): the LIVE rule, which REQUIRES the cheap leg's clv-vs-Pinnacle to land in [5,12) AND the other leg to sit OUTSIDE that band. This script asks the question in between: keep Pinnacle as the reference for each leg's CLV (so it's not "pure" deviation), but drop the band gate entirely and ask whether the raw GAP SIZE between clv_kalshi and clv_poly (both vs pin fair) is itself predictive, continuously, regardless of where either value falls. METHOD: first pre-KO snapshot per fixture; for each side, cheap leg = lower- priced book, gap = clv(cheap leg) - clv(other leg) [always >= 0]; pick the side with the single largest gap; settle at the cheap book's ask (Kalshi ask / Poly est-ask). Bet UNCONDITIONALLY -- no band requirement anywhere. RESULT on poly_backfill (643 fixtures with a resolvable K/P gap, 2026-07-16): Disjoint gap buckets are NOT monotonic (the tell something's wrong): [0,2)% n=305 ROI -4.0% z-1.05 [2,5)% n=255 ROI +4.2% z-0.01 [5,8)% n= 50 ROI -24.1% z-0.86 <- dips hard, breaks monotonicity [8,12)% n= 22 ROI -8.6% z+0.86 [12,20)% n= 11 ROI +77.5% z+1.37 <- only n=11, ~7% implied (longshot) Cumulative thresholds all collapse to a handful of ultra-longshot bets as T rises (n: 643->83->33->8 as T: 0%->5%->8%->15%); the ONE "significant" cell (gap>=15%, n=8, ROI+144%, z+2.06, p=.02) is 8 bets at ~7% implied odds -- one win at ~14:1 explains the whole result. Classic small-cell longshot mirage, same shape as the round-robin/sub-band traps already catalogued. OOS 60/40 split KILLS every threshold tried: gap>=0%: TRAIN -8.2% -> TEST +9.6% (sign flip, both near zero -- just noise) gap>=5%: TRAIN +5.6% -> TEST -23.9% (sign flip) gap>=8%: TRAIN -12.9% -> TEST +64.9% (sign flip, both small n) No threshold holds sign across the split. VERDICT: raw cross-book CLV-gap magnitude carries NO exploitable edge without the band gate. This is not a contradiction of the frozen strict-xlow rule -- it's confirmation that the band conditions (cheap leg IN [5,12), other leg OUT of [5,12)) are doing real selection work, not just being an arbitrary cutoff. Don't drop or loosen the band based on gap size alone. Read-only, backtest only. Mirrors kalshi_poly_pure_deviation.py conventions. """ import csv, glob, math from datetime import datetime from collections import defaultdict DIR = r"E:\soccer\poly_backfill" _FMTS = ['%Y-%m-%d %H:%M:%S', '%m/%d/%Y %H:%M', '%m/%d/%Y %H:%M:%S', '%Y-%m-%d %H:%M'] def _pts(s): for f in _FMTS: try: return datetime.strptime(s, f) except (ValueError, TypeError): pass return None def _f(x): try: return float(x) except (TypeError, ValueError): return None def _result(rows): for r in reversed(rows): v = (r.get('ftr') or '').upper().strip() if v in ('H', 'D', 'A'): return v return None def fair_mult(o): r = [1.0/x for x in o]; s = sum(r); return [ri/s for ri in r] def half_spread(mid): if mid < 0.05: return 0.003 elif mid < 0.25: return 0.005 elif mid < 0.35: return 0.025 return 0.020 def load(): fixtures = defaultdict(list) for path in glob.glob(f"{DIR}\\*_polybackfill.csv"): with open(path, newline='', encoding='utf-8') as fh: for r in csv.DictReader(fh): fixtures[(r.get('game_date'), r.get('home'), r.get('away'))].append(r) out = [] for key, rows in fixtures.items(): gd = key[0] if not gd: continue pre = sorted([(_pts(r['timestamp']), r) for r in rows if (r.get('status') or '').upper() == 'PRE' and _pts(r['timestamp'])], key=lambda t: t[0]) ftr = _result(rows) if pre and ftr: out.append(dict(gd=gd, ftr=ftr, pre=[r for _, r in pre])) out.sort(key=lambda f: f['gd']) return out SIDX = {'home': 0, 'away': 2} def pick_no_band(fx): """First pre-KO snapshot: pick the side/book with the LARGEST clv-vs-Pinnacle gap between Kalshi and Poly. No band requirement anywhere. Bet unconditionally.""" for r in fx['pre']: ph, pd_, pa = _f(r.get('pin_home')), _f(r.get('pin_draw')), _f(r.get('pin_away')) if None in (ph, pd_, pa) or not (ph > 1 and pd_ > 1 and pa > 1): continue pf = fair_mult((ph, pd_, pa)); cands = [] for side, si in SIDX.items(): km = _f(r.get(f'kalshi_{side}_mid')); ka = _f(r.get(f'kalshi_{side}_ask')) pm = _f(r.get(f'poly_{side}_price')) if km is None or pm is None or not (0 < km < 1 and 0 < pm < 1): continue if km <= pm: mid, px, other = km, (ka if ka and 0 < ka < 1 else km), pm else: mid, px, other = pm, pm + half_spread(pm), km clv = (pf[si] - mid) / pf[si]; oclv = (pf[si] - other) / pf[si] gap = clv - oclv won = (fx['ftr'] == 'H' and side == 'home') or (fx['ftr'] == 'A' and side == 'away') cands.append(dict(clv=clv, gap=gap, side=side, px=px, won=won, gd=fx['gd'])) if cands: return max(cands, key=lambda d: d['gap']) return None def stat(bets): n = len(bets) if not n: return None w = sum(b['won'] for b in bets) net = sum(((1-b['px'])/b['px']) if b['won'] else -1.0 for b in bets) mu = sum(b['px'] for b in bets); var = sum(b['px']*(1-b['px']) for b in bets) z = (w-mu)/math.sqrt(var) if var > 0 else 0.0 return dict(n=n, w=w, wr=w/n*100, iwr=mu/n*100, roi=net/n*100, z=z, p=0.5*math.erfc(z/math.sqrt(2))) def line(tag, s): if not s: print(f" {tag:26s} (none)"); return print(f" {tag:26s} n={s['n']:3d} {s['w']}W-{s['n']-s['w']}L WR={s['wr']:4.1f}% " f"(impl {s['iwr']:4.1f}%) ROI={s['roi']:+7.1f}% z={s['z']:+.2f} p={s['p']:.4f}") def main(): F = load() bets = [b for b in (pick_no_band(f) for f in F) if b] bets.sort(key=lambda b: b['gd']) print(f"Total fixtures with a K/P gap on at least one side: {len(bets)}\n") print("=== Disjoint buckets by GAP SIZE ===") for lo, hi in [(0, 0.02), (0.02, 0.05), (0.05, 0.08), (0.08, 0.12), (0.12, 0.20), (0.20, 1.0)]: sub = [b for b in bets if lo <= b['gap'] < hi] line(f"gap [{lo*100:.0f},{hi*100:.0f})%", stat(sub)) print("\n=== Cumulative gap thresholds (bet if gap >= T, no cap) ===") for T in (0.0, 0.02, 0.03, 0.05, 0.08, 0.10, 0.12, 0.15, 0.20): sub = [b for b in bets if b['gap'] >= T] line(f"gap >= {T*100:.0f}%", stat(sub)) print("\n=== OOS 60/40 check ===") for T in (0.0, 0.05, 0.08): sub = sorted([b for b in bets if b['gap'] >= T], key=lambda b: b['gd']) cut = int(len(sub) * 0.6) line(f"T={T*100:.0f}% TRAIN", stat(sub[:cut])) line(f"T={T*100:.0f}% TEST", stat(sub[cut:])) if __name__ == '__main__': main()