""" Is Pinnacle structurally mis-calibrated in SOUTH AMERICA? (read-only, trades nothing) Built 2026-07-25 from the user's hypothesis: "there's something structurally wrong with South American leagues." Pooling regions is the point -- no single SA league has the sample to say anything (BSA n=32, PE1 n=12, CL1 n=8), but pooled they do. Same engine as pin_calibration_by_league.py: EVERY fixture with Pinnacle odds and a result, not just the ones we bet. THE THREE CONTROLS THAT MAKE THIS MORE THAN A SCAN: 1. DROP THE MOTIVATING LEAGUE. The hypothesis came from BSA's 0-for-12, and BSA is 44% of the SA sample. If the pattern is really just BSA, removing it kills the effect. (It doesn't -- the effect gets STRONGER.) 2. SIGN CONSISTENCY across independent leagues. pin_calibration_by_league's own standing rule: randomly-signed deviations at small n = noise; a real structural bias is consistent across comparable leagues. Reported as a sign test. 3. WHICH SIDE WE ACTUALLY BET. A calibration error only costs us money if it lands on the side we take. This is the step that stops a real finding from being mis-read as an explanation for a bad record. RESULT ON FIRST RUN (2026-07-25): SA n=73: homes implied 48.2% -> actual 60.3% (z+2.06); draws 27.4% -> 16.4% (z-2.10); aways 24.4% -> 23.3% (z-0.21). Excluding BSA (n=41) the home/draw effect PERSISTS (z+1.87 / -1.55). Home z>0 in 5/5 SA leagues (sign test p=.062). Europe/N.America/Asia all flat (|z| <= 1.3). BUT: 25 of our 28 SA bets are AWAY, and away is the one side calibrated fine. So the finding does NOT explain the -35.3% SA record (which is z=-0.46 on win count = noise). If anything the home under-pricing is an OPPORTUNITY. Usage: python research\\region_calibration_test.py """ import sys import math sys.path.insert(0, r'E:\soccer\research') import pin_calibration_by_league as P import forward_test_unique_book as F SA = {'BSA', 'BR2', 'AR1', 'CL1', 'CO1', 'PE1'} NA = {'MLS', 'USL', 'MX1', 'CAN', 'NWSL'} ASIA = {'JP1', 'KO1', 'CN1'} MOTIVATING = 'BSA' # the league that generated the hypothesis -- control for it def region(lg): if lg in SA: return 'S.America' if lg in NA: return 'N.America' if lg in ASIA: return 'Asia' return 'Europe' def z_for(n, p, actual): sd = math.sqrt(n * p * (1 - p)) return (actual - p * n) / sd if sd > 0 else float('nan') def detectable(n, base=0.45): """Smallest calibration error (pp) detectable at 95% conf / 80% power.""" return 2.8 * math.sqrt(base * (1 - base) / n) * 100 def line(s, label): n = len(s) if n < 5: print(f' {label:26s} n={n:4d} -- too few') return out = [f' {label:26s} n={n:4d}'] for key, letter in (('pfh', 'H'), ('pfd', 'D'), ('pfa', 'A')): imp = sum(x[key] for x in s) / n cnt = sum(x['ftr'] == letter for x in s) out.append(f'{letter} {imp*100:5.1f}->{cnt/n*100:5.1f} ' f'(z{z_for(n, imp, cnt):+5.2f})') out.append(f'| detects {detectable(n):.0f}pp') print(' '.join(out)) def main(): fx = P.load_fixtures() print('A) PINNACLE CALIBRATION BY REGION (all fixtures with a result)\n') for rg in ('S.America', 'Europe', 'N.America', 'Asia'): line([f for f in fx if region(f['lg']) == rg], rg) print(f'\nB) CONTROL -- does SA survive dropping {MOTIVATING} ' f'(the league that raised the question)?\n') line([f for f in fx if f['lg'] in SA], 'SA all') line([f for f in fx if f['lg'] in SA and f['lg'] != MOTIVATING], f'SA excluding {MOTIVATING}') line([f for f in fx if f['lg'] == MOTIVATING], f'{MOTIVATING} alone') print('\nC) SIGN CONSISTENCY across SA leagues (noise is randomly signed)\n') rows = [] for lg in sorted(SA): s = [f for f in fx if f['lg'] == lg] n = len(s) if n < 5: continue iH = sum(x['pfh'] for x in s) / n cH = sum(x['ftr'] == 'H' for x in s) iD = sum(x['pfd'] for x in s) / n cD = sum(x['ftr'] == 'D' for x in s) rows.append((lg, n, z_for(n, iH, cH), z_for(n, iD, cD))) print(f' {lg:6s} n={n:4d} home z{rows[-1][2]:+5.2f} draw z{rows[-1][3]:+5.2f}') k = len(rows) ph = sum(1 for r in rows if r[2] > 0) pd_ = sum(1 for r in rows if r[3] < 0) print(f'\n home z>0 in {ph}/{k} leagues, draw z<0 in {pd_}/{k}') print(f' P(all {k} one way by chance) = {2 * 0.5 ** k:.3f} (two-tailed sign test)') print('\nD) DOES IT LAND ON THE SIDE WE BET? (forward bets, post-freeze)\n') bets = [] for f in F.load_fixtures(): b = fire_with_side(f) if b and f['gd'] > F.FREEZE_DATE: b['sa'] = f['league'] in SA bets.append(b) for lbl, sel in (('South America', lambda b: b['sa']), ('rest of world', lambda b: not b['sa'])): s = [b for b in bets if sel(b)] for side in ('home', 'away'): ss = [b for b in s if b['side'] == side] if not ss: continue n = len(ss) w = sum(b['won'] for b in ss) roi = sum((1 / b['px'] - 1) if b['won'] else -1 for b in ss) / n print(f' {lbl:14s} {side:5s} n={n:3d} W={w:2d} WR={w/n*100:5.1f}% ' f'implied={sum(b["px"] for b in ss)/n*100:5.1f}% ROI={roi*100:+7.1f}%') print() print(' A calibration error only costs money on the side we actually take.') def fire_with_side(fx): """fire_xlow, but keeping which side qualified (the base version drops it).""" for r in fx['pre']: ph, pd_, pa = (F._f(r.get('pin_home')), F._f(r.get('pin_draw')), F._f(r.get('pin_away'))) if None in (ph, pd_, pa) or not (ph > 1 and pd_ > 1 and pa > 1): continue pf = F.fair_mult((ph, pd_, pa)) cands = [] for side, si in F.SIDX.items(): km, ka = F.book_quote(r, 'kalshi', side) pm, pa_ = F.book_quote(r, 'poly', side) 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, pm else: mid, px, other = pm, pa_, km if px is None or not (0 < px < 1): continue clv = (pf[si] - mid) / pf[si] oc = (pf[si] - other) / pf[si] if not (mid < other and F.BAND[0] <= clv < F.BAND[1]): continue if F.BAND[0] <= oc < F.BAND[1]: continue won = ((fx['ftr'] == 'H' and side == 'home') or (fx['ftr'] == 'A' and side == 'away')) cands.append(dict(clv=clv, px=px, won=won, side=side)) if cands: return max(cands, key=lambda d: d['clv']) return None if __name__ == '__main__': main()