""" Pressure-test the surviving "unique-bet" OOS result before trusting it: 1. Split-point robustness: re-run the train/test at split = 0.4,0.5,0.6,0.7 and also a LAST-40%-only test, to see if the held-out edge depends on where we cut. 2. Concentration: in the TEST half, how many distinct game-DATES and LEAGUES do the unique WINS come from? (a +81% driven by 3 wins on one day = noise, per prior round-robin/BSA lessons.) 3. Leave-one-day-out on the full unique set: drop each game-date, recompute z, report the biggest single-day swing (fragility check). Multiplicative de-vig, [0.05,0.12), shared-price fixtures, Kalshi@ask / Poly@mid. Read-only. """ import csv, glob, os, math from datetime import datetime from collections import defaultdict DIR = r"E:\soccer\poly_backfill" BAND = (0.05, 0.12) _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 load(): fixtures = defaultdict(list) for path in glob.glob(os.path.join(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('league'), r.get('home'), r.get('away'))].append(r) fx = [] for key, rows in fixtures.items(): gd, league, home, away = key 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]) if not pre: continue ftr = _result(rows) if ftr is None: continue fx.append(dict(gd=gd, league=league, ftr=ftr, pre=[r for _, r in pre])) fx.sort(key=lambda f: f['gd']) return fx def fire(fx, book): cands = []; have = False for side, sidx in (('home', 0), ('away', 2)): 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 if book == 'kalshi': mid = _f(r.get(f'kalshi_{side}_mid')); px = _f(r.get(f'kalshi_{side}_ask')) else: mid = _f(r.get(f'poly_{side}_price')); px = mid if mid is None or px is None or not (0 < mid < 1) or not (0 < px < 1): continue have = True pf = fair_mult((ph, pd_, pa))[sidx] clv = (pf - mid) / pf if pf else None if clv is None or not (BAND[0] <= clv < BAND[1]): continue won = (fx['ftr'] == 'H' and side == 'home') or (fx['ftr'] == 'A' and side == 'away') cands.append((clv, side, px, won)); break if not cands: return (None, have) clv, side, px, won = max(cands, key=lambda t: t[0]) return (dict(side=side, px=px, won=won, clv=clv), have) def stat(bets): n = len(bets) if not n: return None wins = 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 = (wins-mu)/math.sqrt(var) if var > 0 else 0.0 return dict(n=n, wr=wins/n*100, roi=net/n*100, z=z, p=0.5*math.erfc(z/math.sqrt(2))) def main(): fx = load() uniq = [] # unique (one-book-only) bets, each tagged gd/league for f in fx: k, kh = fire(f, 'kalshi'); p, ph = fire(f, 'poly') if not (kh and ph): continue if k and not p: uniq.append(dict(gd=f['gd'], league=f['league'], book='K', **k)) elif p and not k: uniq.append(dict(gd=f['gd'], league=f['league'], book='P', **p)) uniq.sort(key=lambda b: b['gd']) print(f"total unique (one-book-only) bets = {len(uniq)}\n") # 1) split-point robustness print("1) Split-point robustness (TEST = held-out tail):") print(f" {'split':>6} {'trN':>4} {'trROI':>7} {'trz':>6} | {'teN':>4} {'teROI':>7} {'tez':>6} {'tep':>6}") for sp in (0.40, 0.50, 0.60, 0.70): cut = int(len(uniq)*sp); tr, te = uniq[:cut], uniq[cut:] st, se = stat(tr), stat(te) print(f" {sp:6.0%} {st['n']:4d} {st['roi']:+7.1f} {st['z']:+6.2f} | " f"{se['n']:4d} {se['roi']:+7.1f} {se['z']:+6.2f} {se['p']:6.3f}") # 2) concentration in the 60/40 test half cut = int(len(uniq)*0.60); te = uniq[cut:] print(f"\n2) Concentration of the 60/40 TEST half (n={len(te)}):") bydate = defaultdict(list); bylg = defaultdict(list) for b in te: bydate[b['gd']].append(b); bylg[b['league']].append(b) wins = [b for b in te if b['won']] wdates = defaultdict(int) for b in wins: wdates[b['gd']] += 1 print(f" wins={len(wins)} across {len(wdates)} distinct dates; " f"bets across {len(bydate)} dates, {len(bylg)} leagues") top = sorted(wdates.items(), key=lambda kv: -kv[1])[:5] print(f" top win-dates: {top}") print(f" net ROIturn if drop single best win-date:") for d, _ in top[:1]: rest = [b for b in te if b['gd'] != d] s = stat(rest) print(f" drop {d}: n={s['n']} ROIturn={s['roi']:+.1f}% z={s['z']:+.2f} p={s['p']:.3f}") # 3) leave-one-day-out on FULL unique set print(f"\n3) Leave-one-DAY-out fragility (full unique set, n={len(uniq)}):") base = stat(uniq) dates = sorted({b['gd'] for b in uniq}) swings = [] for d in dates: s = stat([b for b in uniq if b['gd'] != d]) swings.append((s['z'] - base['z'], d, s['z'], s['n'])) swings.sort() print(f" full-set: n={base['n']} ROIturn={base['roi']:+.1f}% z={base['z']:+.2f} p={base['p']:.3f}") worst = swings[0] # most negative => that day was propping it up print(f" biggest DROP in z from removing one day: {worst[0]:+.2f} " f"(remove {worst[1]} -> z {worst[2]:+.2f}, n={worst[3]})") n_pos = sum(1 for s in swings if s[2] > 1.64) print(f" days where z stays > 1.64 after removal: {n_pos}/{len(dates)}") if __name__ == '__main__': main()