"""Score the Poly-US paper market maker. WHY THIS EXISTS. mm_quoter_polyus.py has been quoting since 2026-07-29 and writing mm_polyus_quotes.csv, but nothing ever read that file back. The Kalshi book has mm_settle.py; this side had no scorer at all, so ~460k quote rows across 264 books sat unmeasured -- on the one venue that CREDITS the maker (MAKER_THETA = 0.0125 x p(1-p)). SCORE BOOKS, NOT LEGS. A leg is not the unit of risk here. The quoter takes two terminal actions that a leg-by-leg reading misses entirely, and an earlier version of this file missed both: COMPLETE_AT_BID -- two legs have sold, so cross into the bid on the third near kickoff. The book is then short all three outcomes, exactly one of which pays: the P&L is LOCKED and outcome-independent. Holding these to settlement scores a position that was already closed. FLATTEN_1OF3 -- only one leg sold, so buy it back rather than carry a naked position into the match. Also closed before kickoff, also not a settlement. Both write their realised per-contract economics into `reason` at the moment they fire, computed with the live fee schedule. This reads those numbers rather than recomputing them, for the same reason it reads `rebate_c` off the row: a second implementation of the fee model is a second thing to drift. WHAT A PASSIVE FILL MEANS, and why it is a CEILING. Poly-US exposes no trade tape; `last_px` is a LEVEL sampled every 20s. A level already above our ask when we start quoting is not evidence anyone traded with us, so a fill counts only when last_px CHANGES to a value at or above our resting ask. Still optimistic -- it assumes we held queue priority for that print. Same caveat mm_settle carries. Usage: python main\\mm_settle_polyus.py """ import csv import re import shutil import sys import tempfile from collections import defaultdict from pathlib import Path sys.path.insert(0, r'E:\soccer\main') sys.path.insert(0, r'E:\soccer\research') QUOTES = Path(r'E:\soccer\research\mm_polyus_quotes.csv') RESULTS = Path(r'E:\soccer\research\mm_polyus_results.csv') FIELDS = ['event_slug', 'div', 'match', 'kickoff', 'legs_sold', 'outcome', 'pnl_c_per_contract', 'contracts', 'pnl_dollars', 'rebate_c', 'regime'] SIDE_WIN = {'home': 'H', 'draw': 'D', 'away': 'A'} RE_LOCKED = re.compile(r'=>\s*([+-]?[\d.]+)c/contract locked') RE_COST = re.compile(r'=>\s*cost\s*([+-]?[\d.]+)c/contract') def _f(x, d=None): try: return float(x) except (TypeError, ValueError): return d def load_quotes(): """Rows grouped by leg, timestamp-ordered. Snapshotted first: the quoter holds the real file open and polls every 20s, and a partial read of a 96MB CSV mid-write is a silent wrong answer rather than an error.""" tmp = Path(tempfile.gettempdir()) / 'mm_polyus_quotes.snap.csv' shutil.copyfile(QUOTES, tmp) by_leg = defaultdict(list) with tmp.open(newline='', encoding='utf-8', errors='replace') as fh: for r in csv.DictReader(fh): if r.get('market_slug'): by_leg[r['market_slug']].append(r) for v in by_leg.values(): v.sort(key=lambda r: r.get('ts') or '') return by_leg def results_map(): import forward_test_unique_book as F out = {} for fx in F.load_fixtures(require_result=True): r0 = fx['pre'][0] if fx['pre'] else {} out[(fx['gd'], fx['league'], r0.get('home'), r0.get('away'))] = fx['ftr'] return out def leg_state(rows): """(state, price, per_contract_c, rebate_c) for one leg. state is 'completed' (crossed into the bid), 'flattened' (bought back), 'sold' (passive fill, still open at kickoff) or 'none'. """ sold_px = None reb = 0.0 prev = None for r in rows: reason = r.get('reason') or '' if 'COMPLETE_AT_BID' in reason: m = RE_LOCKED.search(reason) return 'completed', _f(r.get('our_ask')), (_f(m.group(1), 0.0) if m else 0.0), reb if 'FLATTEN_1OF3' in reason: m = RE_COST.search(reason) return 'flattened', _f(r.get('our_ask')), (_f(m.group(1), 0.0) if m else 0.0), reb ask, lp = _f(r.get('our_ask')), _f(r.get('last_px')) if sold_px is None and ask is not None and lp is not None and 0 < ask < 1: if prev is not None and lp != prev and lp >= ask: sold_px = ask reb = _f(r.get('rebate_c'), 0.0) or 0.0 if lp is not None: prev = lp return ('sold', sold_px, None, reb) if sold_px is not None else ('none', None, None, 0.0) def main(): by_leg = load_quotes() res = results_map() books = defaultdict(list) for slug, rows in by_leg.items(): books[rows[0].get('event_slug')].append((slug, rows)) out, unresolved = [], 0 for ev, legs in books.items(): r0 = legs[0][1][0] gd = (r0.get('kickoff') or '')[:10] ftr = res.get((gd, (r0.get('div') or '').upper(), r0.get('home'), r0.get('away'))) states = {} for slug, rows in legs: side = (rows[0].get('side') or '').lower() states[side] = leg_state(rows) sold = [s for s in states.values() if s[0] in ('sold', 'completed')] n_sold = len(sold) flat = [s for s in states.values() if s[0] == 'flattened'] completed = any(s[0] == 'completed' for s in states.values()) if not sold and not flat: continue # never traded if ftr not in ('H', 'D', 'A') and not (completed or flat): unresolved += 1 # unsettled AND still open continue size = _f(r0.get('size'), 10.0) or 10.0 reb = sum(s[3] for s in states.values()) # SIGN CONVENTIONS, and they are OPPOSITE -- this is worth stating because # reading them the same way silently flips the whole result: # COMPLETE_AT_BID net_c = (collected - 1 + rebates)*100 - fee # -> already P&L-signed, and ALREADY INCLUDES the rebates. Adding # `reb` again here would double-count them. # FLATTEN_1OF3 cost_c = (ask - sold)*100 + fee # -> a COST. P&L is its NEGATION. It excludes the rebate earned on # the original sale, so that is added back separately. flat_pnl = sum(-(s[2] or 0.0) for s in flat) if completed: per = next(s[2] for s in states.values() if s[0] == 'completed') outcome = 'complete' elif flat and n_sold == 0: per = flat_pnl + reb outcome = 'flattened' else: # Still open at kickoff: short n_sold outcomes, settle each. per = 0.0 for side, s in states.items(): if s[0] == 'sold' and s[1] is not None: per += (s[1] - (1 if SIDE_WIN.get(side) == ftr else 0)) * 100.0 per += flat_pnl + reb outcome = f'open {n_sold}/3' cts = size * max(n_sold + len(flat), 1) out.append({ 'event_slug': ev, 'div': r0.get('div'), 'match': f"{r0.get('home')} vs {r0.get('away')}", 'kickoff': r0.get('kickoff'), 'legs_sold': n_sold, 'outcome': outcome, 'pnl_c_per_contract': round(per, 3), 'contracts': cts, 'pnl_dollars': round(per * size / 100.0, 4), 'rebate_c': round(reb, 4), 'regime': r0.get('regime')}) with RESULTS.open('w', newline='', encoding='utf-8') as fh: w = csv.DictWriter(fh, fieldnames=FIELDS) w.writeheader() for o in out: w.writerow(o) print(summarise(out, len(books), unresolved)) def summarise(out, books_seen, unresolved): L = ['POLY-US PAPER MARKET MAKER (book-level)', '', f'{books_seen} books in the quote log, {len(out)} that traded and can ' f'be scored ({unresolved} traded but unsettled)', ''] if not out: return '\n'.join(L + ['nothing scoreable yet']) tot = sum(o['pnl_dollars'] for o in out) L += [f' books scored : {len(out)}', f' TOTAL P&L : {tot:+,.2f} dollars', ''] grp = defaultdict(lambda: [0, 0.0]) for o in out: k = o['outcome'] if o['outcome'] in ('complete', 'flattened') else o['outcome'] grp[k][0] += 1 grp[k][1] += o['pnl_dollars'] L += ['-' * 62, 'BY HOW THE BOOK ENDED', '', f'{"outcome":12} {"books":>6} {"P&L $":>10} {"per book":>10}'] for k in sorted(grp, key=lambda x: -grp[x][1]): n, p = grp[k] L.append(f'{k:12} {n:>6} {p:>10,.2f} {p/n:>10,.3f}') L.append('') by_div = defaultdict(lambda: [0, 0.0]) for o in out: by_div[o['div']][0] += 1 by_div[o['div']][1] += o['pnl_dollars'] L += ['-' * 62, 'BY LEAGUE', ''] for d, (n, p) in sorted(by_div.items(), key=lambda kv: -kv[1][1]): L.append(f' {d:6s} {n:>4} books {p:+8,.2f} $ {p/n:+7.3f} $/book') L += ['', 'Complete books are OUTCOME-INDEPENDENT -- locked at the moment we', 'crossed. Flattened books are closed positions. Only "open n/3" rows', 'carry match risk. Passive fills assume queue priority, so those are', 'a CEILING, same caveat mm_settle carries.'] return '\n'.join(L) if __name__ == '__main__': main()