"""Does pre-quote volatility predict fills? Clean version. The first pass (mm_volatility_fill.py) found +5pp for the wildest quartile but could not be trusted, because both halves of the test came from the same file. The paper log marks a leg filled when a trade printed at or above our ask, so a leg whose price moves more mechanically generates more qualifying prints -- the instrument was biased in exactly the direction of the hypothesis. This separates the two sides completely: MOVEMENT from live_data -- the snapshot feed the xlow/xhigh strategies run on. 93 days of it, written by a different process for a different purpose, and measured over the window BEFORE the maker ever quoted the leg, so no fill can be inside it. FILL from real Kalshi trade prints, queue-aware: filled only once cumulative volume at or above our price exceeds the depth that was already resting there when we posted. Not "a print happened". If the effect survives that, it is a fact about the market. If it vanishes, the first result was an artifact of the fill model and the hypothesis is unproven. """ import csv, collections, glob, json, os, statistics, sys, datetime as dt sys.path.insert(0, r'E:\soccer\main') sys.path.insert(0, r'E:\soccer\research') import mm_settle as S QUOTES = r'E:\soccer\research\mm_paper_quotes.csv' CACHE = r'E:\soccer\research\.trades_cache.json' LIVE_DIR = r'E:\soccer\live_data' SIDE_COL = {'home': 'kalshi_home_mid', 'draw': 'kalshi_draw_mid', 'away': 'kalshi_away_mid'} def _f(x): try: v = float(x) return v if 0 < v < 1 else None except (TypeError, ValueError): return None def load_legs(): """First quote per leg: price, queue depth, kickoff, fixture identity.""" first = {} with open(QUOTES, encoding='utf-8', errors='replace') as fh: for r in csv.DictReader(fh): t = r.get('leg_ticker') if not t or t in first: continue if not (r.get('ask_size') or '').strip(): continue # before depth logging a, ts, ko = _f(r.get('our_ask')), r.get('ts'), r.get('kickoff') if not (a and ts and ko): continue first[t] = {'our_ask': a, 'queue': float(r['ask_size']), 'ts': S._pt(ts), 'ko': S._pt(ko), 'home': r.get('home'), 'away': r.get('away'), 'side': (r.get('side') or '').lower(), 'gd': (r.get('kickoff') or '')[:10]} return {t: v for t, v in first.items() if v['ts'] and v['ko'] and v['side'] in SIDE_COL} def load_movement(legs): """{leg: stdev of the leg's mid BEFORE we first quoted it}, from live_data.""" want = collections.defaultdict(list) # (gd, home, away) -> [leg tickers] for t, v in legs.items(): want[(v['gd'], v['home'], v['away'])].append(t) dates = sorted({k[0] for k in want}) series = collections.defaultdict(list) for d in dates: path = os.path.join(LIVE_DIR, f'{d}.csv') if not os.path.exists(path): continue with open(path, encoding='utf-8', errors='replace') as fh: for r in csv.DictReader(fh): key = (r.get('game_date'), r.get('home'), r.get('away')) if key not in want: continue ts = S._pt(r.get('timestamp')) if not ts: continue for t in want[key]: col = SIDE_COL[legs[t]['side']] m = _f(r.get(col)) if m is not None: series[t].append((ts, m)) out = {} for t, pts in series.items(): cut = legs[t]['ts'] # STRICTLY BEFORE the first quote -- this is what makes it non-circular # live_data timestamps are naive local; the quote feed is tz-aware. # Comparing them raises, and stripping tzinfo off the quote side is the # SAME truncation bug that made me report an 8h collection ceiling # earlier today when the real one was 12h. Normalise by converting the # aware side to naive LOCAL time rather than by discarding the offset. cutn = cut.astimezone().replace(tzinfo=None) if cut.tzinfo else cut pre = [m for ts, m in sorted(pts) if (ts.replace(tzinfo=None) if ts.tzinfo else ts) < cutn] if len(pre) < 8: continue out[t] = {'sd': statistics.pstdev(pre), 'n': len(pre), 'range': max(pre) - min(pre), 'changes': sum(1 for a, b in zip(pre, pre[1:]) if abs(b - a) > 1e-9)} return out def main(): if not os.path.exists(CACHE): print('no trades cache -- run mm_ratchet_backtest.py first') return trades = json.load(open(CACHE, encoding='utf-8')) legs = load_legs() print(f'{len(legs)} legs with depth logged; {len(trades)} tickers in the ' f'trade cache') mv = load_movement(legs) print(f'{len(mv)} of those have pre-quote price history in live_data') rows = [] for t, v in legs.items(): if t not in mv: continue key = f"{t}|{v['ts'].isoformat()}" if key not in trades: continue cum, filled = 0.0, False for ts_s, p, c in trades[key]: ts = S._pt(ts_s) if not ts: break ko = v['ko'] if (ts.tzinfo is None) != (ko.tzinfo is None): ts = ts.astimezone() if ts.tzinfo else ts.replace( tzinfo=ko.tzinfo) if ts >= ko: break if p >= v['our_ask'] - 1e-9: cum += c if cum > v['queue']: filled = True break rows.append(dict(mv[t], filled=int(filled), ticker=t)) if len(rows) < 40: print(f'only {len(rows)} legs join across all three sources -- too few') return base = 100.0 * sum(r['filled'] for r in rows) / len(rows) print(f'\n{len(rows)} legs with movement AND queue-aware fills, ' f'base fill rate {base:.1f}%\n') for metric, label in (('sd', 'pre-quote stdev'), ('changes', 'price changes'), ('range', 'pre-quote range')): vals = sorted(r[metric] for r in rows) q = [vals[int(p * len(vals))] for p in (0.5, 0.75)] buckets = [(-1, q[0], 'calm half'), (q[0], q[1], '3rd quartile'), (q[1], 1e9, 'wildest 25%')] print(f' by {label}:') got = [] for lo, hi, name in buckets: sel = [r for r in rows if lo <= r[metric] < hi] if len(sel) < 12: continue f = sum(r['filled'] for r in sel) got.append((name, 100.0 * f / len(sel))) print(f' {name:>14} {len(sel):>5} legs {f:>4} filled ' f'{100.0*f/len(sel):>6.1f}%') if len(got) >= 2: print(f' -> {got[-1][0]} vs {got[0][0]}: ' f'{got[-1][1]-got[0][1]:+.1f}pp') print() if __name__ == '__main__': main()