"""Does gating on a book's GROSS ask-margin help, historically? Today's 7 paper books suggested: only post a 3-leg book if sum(asks) - 100 covers the ~2.4c it costs to complete the last leg (1c crossing + taker fee). That threshold was fitted to 7 points, so test the mechanism on the 641 settled events in kalshi_backfill.csv. THE TAUTOLOGY TRAP, stated up front: if all three legs fill, book P&L IS sum(asks) - 100. Gating on gross would then "work" by definition and prove nothing. The only real question is whether high-gross books still FILL -- a wide book is wide because it is illiquid, and gating toward wide books may select books nobody trades, leaving 2-of-3 partial books, which is the dangerous state. So the test here is fill behaviour vs gross, not P&L vs gross. KNOWN LOOK-AHEAD (do not paper over it): kalshi_backfill snaps the LAST candle at or before kickoff, but we would post hours earlier. Using that near-kickoff ask as our posting price and then counting trades over the whole pre-KO window counts fills at prices the book only reached later -- it inflates fill rates. Mitigated by reporting a NEAR window (last 3h) alongside the full window; the near window is the honest one. This is the same class of error that produced the 83%-vs-19% fill bug in the paper maker. MAKER-SELL SEMANTICS (verified against our own fills): taker_book_side='bid' means the taker BOUGHT, i.e. a maker SOLD. Those are our fills. """ import csv import datetime as dt import statistics from collections import defaultdict BACKFILL = r'E:\soccer\research\kalshi_backfill.csv' TRADES = r'E:\soccer\research\mm_trades.csv' NEAR_H = 3.0 # honest window: trades within 3h of the quote snapshot FEE = lambda p: 0.07 * p * (1 - p) * 100 # cents per contract # LIQUIDITY FILTER (added after first run, which was garbage without it). # 192 of 641 backfill books have gross >= 5c, and they are not opportunities -- # they are EMPTY books. Their mean 3-leg spread is 20.1c vs 4.2c for the rest; # the worst quotes asks 0.80/0.81/0.80 against bids 0.20/0.20/0.20 (gross # "+141c"). A resting sell at an ask nobody will ever lift does not fill, so # counting those as tradeable inflates everything. Our live quoter only ever # competes on books already at/near the 1c tick, so cap the total 3-leg spread. MAX_SPREAD_C = 6.0 def _pt(s): if not s: return None try: d = dt.datetime.fromisoformat(s.replace('Z', '+00:00')) except ValueError: return None return d if d.tzinfo else d.replace(tzinfo=dt.timezone.utc) def load_books(): books = {} with open(BACKFILL, encoding='utf-8') as fh: for r in csv.DictReader(fh): try: asks = {s: float(r[f'k_{s}_ask']) for s in ('home', 'draw', 'away')} bids = {s: float(r[f'k_{s}_bid']) for s in ('home', 'draw', 'away')} except (ValueError, KeyError): continue ko = _pt(r['kickoff']) if ko is None or not all(0 < a < 1 for a in asks.values()): continue spread = sum(asks[s] - bids[s] for s in asks) * 100 if spread > MAX_SPREAD_C or spread < 0: continue # COHERENCE FILTER (added after the spread filter still left a # 20c-gross tail). A real 1X2 book must satisfy sum(bids) < 100 < # sum(asks): buying all three costs less than the payout, selling # all three collects more. 66 of the 69 remaining gross>=5c books # had sum(bids) up to 155c, with one leg at 0.96 -- those are # IN-PLAY prices that landed in a "pre-kickoff" dataset because the # recorded kickoff is wrong, so the last-candle-before-kickoff rule # picked a mid-match candle. Not tradeable, and not pre-match. if sum(bids.values()) >= 1.0: continue books[r['event']] = {'div': r['div'], 'ko': ko, 'asks': asks, 'bids': bids, 'ftr': r['ftr'], 'spread': spread, 'gross': sum(asks.values()) * 100 - 100} return books def load_fills(books): """Per (event, leg-suffix): maker-sell trades, bounded pre-kickoff.""" want = {e: b['ko'] for e, b in books.items()} out = defaultdict(list) with open(TRADES, encoding='utf-8') as fh: for r in csv.DictReader(fh): tk = r['ticker'] i = tk.rfind('-') if i < 0: continue ev = tk[:i] ko = want.get(ev) if ko is None: continue if r['taker_book_side'] != 'bid': # taker bought => maker sold continue t = _pt(r['ts']) if t is None or t >= ko: continue try: out[(ev, tk[i + 1:])].append((t, float(r['price']), float(r['count']))) except ValueError: continue return out def leg_suffixes(books, fills): """Map each event's 3 sides to their ticker suffixes by matching ask prices. Suffixes are team codes of varying length, so string-slicing them is the bug that mis-assigned 24/1060 legs before. Instead: the draw leg's suffix is the one Kalshi tags TIE/DRAW; home/away are resolved by which observed trade prices sit closest to that side's quoted ask. """ res = {} for ev, b in books.items(): sufs = sorted({s for (e, s) in fills if e == ev}) if len(sufs) != 3: continue draw = [s for s in sufs if s.upper() in ('TIE', 'DRAW')] if len(draw) != 1: continue rest = [s for s in sufs if s != draw[0]] med = {} for s in rest: px = [p for (_, p, _) in fills[(ev, s)]] med[s] = statistics.median(px) if px else None if any(v is None for v in med.values()): continue # assign so total distance to the two quoted asks is minimised a, c = rest d1 = abs(med[a] - b['asks']['home']) + abs(med[c] - b['asks']['away']) d2 = abs(med[c] - b['asks']['home']) + abs(med[a] - b['asks']['away']) home, away = (a, c) if d1 <= d2 else (c, a) res[ev] = {'home': home, 'draw': draw[0], 'away': away} return res def filled(trades, our_ask, ko, near_only): """Would a resting sell at our_ask have been hit? Trade price >= our ask.""" lo = ko - dt.timedelta(hours=NEAR_H) for (t, p, _) in trades: if near_only and t < lo: continue if p >= our_ask - 1e-9: return True return False def main(): books = load_books() fills = load_fills(books) smap = leg_suffixes(books, fills) print(f'{len(books)} settled books; {len(smap)} with all 3 legs resolvable in trade data\n') for near in (True, False): lbl = f'NEAR window (last {NEAR_H:g}h, honest)' if near else 'FULL pre-KO window (look-ahead inflated)' rows = [] for ev, sfx in smap.items(): b = books[ev] nf = sum(filled(fills[(ev, sfx[s])], b['asks'][s], b['ko'], near) for s in ('home', 'draw', 'away')) rows.append({'gross': b['gross'], 'nf': nf, 'ev': ev, 'b': b, 'sfx': sfx}) print(f'--- {lbl}') print(f" {'gross bucket':>16s} {'books':>6s} {'3/3':>6s} {'2/3':>6s} {'<=1':>6s} " f"{'mean net c/contract':>21s}") buckets = [(-99, 1.0), (1.0, 2.0), (2.0, 3.0), (3.0, 5.0), (5.0, 999)] for lo, hi in buckets: g = [x for x in rows if lo <= x['gross'] < hi] if not g: continue n3 = sum(1 for x in g if x['nf'] == 3) n2 = sum(1 for x in g if x['nf'] == 2) n1 = sum(1 for x in g if x['nf'] <= 1) nets = [] for x in g: if x['nf'] == 3: nets.append(x['gross']) # all natural, no fee elif x['nf'] == 2: # complete the last leg by selling into the bid, pay taker fee b = x['b'] worst = min(b['bids'].values()) nets.append(x['gross'] - 1.0 - FEE(worst)) # <=1 filled: book never formed; no position taken, contributes 0 mn = statistics.mean(nets) if nets else float('nan') name = f'{lo:g} to {hi:g}c' if lo > -99 else f'< {hi:g}c' print(f' {name:>16s} {len(g):6d} {n3:6d} {n2:6d} {n1:6d} {mn:20.2f}c') print() # The decision the gate actually makes, on the honest window. print('GATE SIMULATION (near window): post only books with gross >= T') print(f" {'T':>6s} {'books posted':>13s} {'mean net':>10s} {'total':>10s} {'2/3 stranded':>13s}") rows = [] for ev, sfx in smap.items(): b = books[ev] nf = sum(filled(fills[(ev, sfx[s])], b['asks'][s], b['ko'], True) for s in ('home', 'draw', 'away')) if nf == 3: net = b['gross'] elif nf == 2: net = b['gross'] - 1.0 - FEE(min(b['bids'].values())) else: net = None rows.append({'gross': b['gross'], 'nf': nf, 'net': net}) for T in (-99, 0, 1, 2, 2.5, 3, 4): g = [x for x in rows if x['gross'] >= T] nets = [x['net'] for x in g if x['net'] is not None] if not nets: continue strand = sum(1 for x in g if x['nf'] == 2) / len(g) * 100 print(f' {T:6.1f} {len(g):13d} {statistics.mean(nets):9.2f}c ' f'{sum(nets):9.1f}c {strand:12.1f}%') if __name__ == '__main__': main()