{ "key": "f-fill-rate-size", "mech": "fill-rate adaptive size", "approach": "Overlay on N. Each step, own-fill notional (fill quantity times own best quote) enters a rolling 200-step sum; ratio = sum / (cumulative notional per step * 200) after a 400-step warmup. New quotes of at least 0.05 shares scale by clip(ratio, 0.6, 1.0); N's 0.01-share orders are never touched. Acts on s=2,3,4 only. A guard replays the book's collateral check from StepState and shrinks or drops orders that would raise, because resized resting orders made N's later orders fail without it. Screen result: null.", "tried": [ "Two variants first, down-only A (0.6,1.0) and up-only B (1.0,1.6), all spreads, on the tuning seeds. Those first runs failed 7 of 200 (A) and 1 of 200 (B) simulations with 'Insufficient free cash for order collateral'; N failed none.", "Added the collateral guard and reran A and B on the same tuning seeds; failures went to 0. Reported variant numbers are the guarded reruns. Unguarded runs: A -0.65 (se 0.81), B -0.91 (se 0.30).", "Tuning read: up-sizing in high-fill windows buys arbitrage loss (B retail +0.62, arb -1.42); down-sizing in low-fill windows cuts arbitrage loss but costs retail, worst at s=1 (A retail -1.92, arb +0.40 there).", "Third variant C is A restricted to s=2,3,4, frozen on tuning (+0.66, se 0.53). Holdout gave +0.01, se 0.32: the tuning gain did not hold, and the s=4 tuning gain of +2.28 (se 1.98) came back -0.72 (se 0.61).", "Diagnostics only, on seeds 310000..310007: action counts by quantity, which showed the 0.01-share orders, and constant-scale churn runs. One smoke run of N and A on tuning seed 320025 outside the sweep files." ], "next": "Do not promote; drop this signal as a size scaler. On the holdout it trades retail for arbitrage about one for one (retail -0.42, arb +0.43). Retail arrival rate is fixed within a game, so my untested guess is that high-fill windows mostly reflect price moves and ladder position rather than better flow. The collateral guard is reusable by any overlay that resizes N's orders. Untested: lo below 0.6, and a per-side ratio.", "spreads_acted": [ 2, 3, 4 ], "off_equals_n": true, "off_equals_n_detail": "ON=False: action lists equal to N's at all 2000 steps on each of seeds 310000..310009 (20000 steps), by repr of every action, with equal total edge.", "code": "from collections import deque as _deque\n\n\nclass Strategy(_N):\n \"\"\"Fill-rate adaptive size: new quotes scale by clip(rolling/avg, LO, HI).\"\"\"\n ON = True\n LO = 0.6\n HI = 1.0\n SPREADS = (2, 3, 4)\n WINDOW = 200\n WARMUP = 400\n MIN_Q = 0.05\n\n def __init__(self):\n super().__init__()\n self._fr_win = _deque()\n self._fr_sum = 0.0\n self._fr_tot = 0.0\n self._fr_bid = None\n self._fr_ask = None\n self._fr_touched = False\n\n def _fr_scale(self, state):\n n = 0.0\n if self._fr_bid is not None:\n n += state.buy_filled_quantity * self._fr_bid / 100.0\n if self._fr_ask is not None:\n n += state.sell_filled_quantity * self._fr_ask / 100.0\n self._fr_tot += n\n self._fr_sum += n\n self._fr_win.append(n)\n if len(self._fr_win) > self.WINDOW:\n self._fr_sum -= self._fr_win.popleft()\n t = state.step\n if t < self.WARMUP or self._fr_tot <= 0.0:\n return 1.0\n ratio = max(self._fr_sum, 0.0) / (self._fr_tot / t * self.WINDOW)\n return max(self.LO, min(self.HI, ratio))\n\n def on_step(self, state):\n actions = super().on_step(state)\n if not self.ON:\n return actions\n k = self._fr_scale(state)\n out = actions\n if getattr(self, \"s\", None) in self.SPREADS:\n if k != 1.0:\n self._fr_touched = True\n out = self._fr_resize(state, actions, k)\n if self._fr_touched:\n out = self._fr_guard(state, out)\n self._fr_track(state, out)\n return out\n\n def _fr_guard(self, state, actions):\n \"\"\"Replay the book's collateral check; shrink or drop what would raise.\"\"\"\n live = {o.order_id: o for o in state.own_orders}\n bids = sum(o.price_ticks / 100.0 * o.remaining_quantity\n for o in live.values() if o.side.value == \"BUY\")\n sell_cash = max(0.0, state.reserved_cash - bids)\n free = state.free_cash\n held = 0.0\n out = []\n for a in actions:\n name = type(a).__name__\n if name == \"CancelOrder\":\n o = live.pop(a.order_id, None)\n if o is None:\n continue\n if o.side.value == \"BUY\":\n free += o.price_ticks / 100.0 * o.remaining_quantity\n else:\n kept = sum((1.0 - x.price_ticks / 100.0) * x.remaining_quantity\n for x in live.values() if x.side.value == \"SELL\")\n back = max(0.0, sell_cash - kept)\n free += back\n sell_cash -= back\n out.append(a)\n elif name == \"PlaceOrder\":\n p = a.price_ticks / 100.0\n room = max(0.0, free - 1e-6)\n if a.side.value == \"BUY\":\n top = room / p\n avail = 0.0\n else:\n resting = sum(x.remaining_quantity for x in live.values()\n if x.side.value == \"SELL\")\n avail = max(0.0, state.yes_inventory - resting - held)\n top = avail + room / (1.0 - p)\n q = a.quantity\n if q > top:\n q = int(top * 100) / 100.0\n if q < 0.01:\n continue\n a = type(a)(side=a.side, price_ticks=a.price_ticks, quantity=q,\n client_order_id=a.client_order_id)\n if a.side.value == \"BUY\":\n free -= p * q\n else:\n cover = min(q, avail)\n held += cover\n free -= (1.0 - p) * (q - cover)\n out.append(a)\n else:\n out.append(a)\n return out\n\n def _fr_resize(self, state, actions, k):\n need = 0.0\n for a in actions:\n if type(a).__name__ == \"PlaceOrder\":\n need += a.quantity * self._fr_unit(a)\n room = max(0.0, state.free_cash - need)\n out = []\n for a in actions:\n if type(a).__name__ != \"PlaceOrder\" or a.quantity < self.MIN_Q:\n out.append(a)\n continue\n q = a.quantity * k\n if q > a.quantity:\n u = self._fr_unit(a)\n extra = min(q - a.quantity, room / u) if u > 0 else 0.0\n room -= extra * u\n q = a.quantity + extra\n q = int(q * 100 + 1e-9) / 100.0\n if q < 0.01:\n continue\n out.append(type(a)(side=a.side, price_ticks=a.price_ticks, quantity=q,\n client_order_id=a.client_order_id))\n return out\n\n @staticmethod\n def _fr_unit(a):\n p = a.price_ticks / 100.0\n return p if a.side.value == \"BUY\" else 1.0 - p\n\n def _fr_track(self, state, actions):\n gone = {a.order_id for a in actions if type(a).__name__ == \"CancelOrder\"}\n bid = ask = None\n for o in state.own_orders:\n if o.order_id in gone:\n continue\n if o.side.value == \"BUY\":\n bid = o.price_ticks if bid is None else max(bid, o.price_ticks)\n else:\n ask = o.price_ticks if ask is None else min(ask, o.price_ticks)\n for a in actions:\n if type(a).__name__ == \"PlaceOrder\":\n if a.side.value == \"BUY\":\n bid = a.price_ticks if bid is None else max(bid, a.price_ticks)\n else:\n ask = a.price_ticks if ask is None else min(ask, a.price_ticks)\n self._fr_bid = bid\n self._fr_ask = ask\n", "code_sha256": "2eed99754103e32ef765bb44910bc40c3635f990186de8cd700378ff3e03ac46", "assembled_sha256": "1a301bbc492851942816bde716632d97c882316b94f48b380e8eb1215a5406fd", "tuning": { "seeds": "320000..320199", "selected": "C", "selected_params": { "lo": 0.6, "hi": 1.0, "window": 200, "warmup": 400, "spreads": [ 2, 3, 4 ] }, "tuning_score": { "mean_diff": 0.6606159068884991, "se": 0.5298262415524924 }, "variants": [ { "name": "A", "lo": 0.6, "hi": 1.0, "spreads": [ 1, 2, 3, 4 ], "n": 200, "mean_diff": 0.27964863764899894, "se": 0.6009169915760337, "retail_diff": -0.41746170045600045, "arb_diff": 0.6971103381249999, "by_spread_mean_diff": { "1": -1.5238690769580012, "2": 0.1564470886679986, "3": 0.12173801114791466, "4": 2.2780269694076924 }, "failures": 0 }, { "name": "B", "lo": 1.0, "hi": 1.6, "spreads": [ 1, 2, 3, 4 ], "n": 200, "mean_diff": -0.7957588123885002, "se": 0.2996461512739173, "retail_diff": 0.6197203922480011, "arb_diff": -1.4154792046, "by_spread_mean_diff": { "1": -1.5184115720799998, "2": -0.4960142948660001, "3": -1.060583632112501, "4": -0.14466259209615362 }, "failures": 0 }, { "name": "C", "lo": 0.6, "hi": 1.0, "spreads": [ 2, 3, 4 ], "n": 200, "mean_diff": 0.6606159068884991, "se": 0.5298262415524924, "retail_diff": 0.06300114063149975, "arb_diff": 0.59761476627, "by_spread_mean_diff": { "1": 0.0, "2": 0.1564470886679986, "3": 0.12173801114791466, "4": 2.2780269694076924 }, "failures": 0 } ] }, "paired": { "seeds": "430000..430499", "n": 500, "complete": true, "missing_seeds": [], "overall": { "n": 500, "sum_d": 3.6039548935996937, "sum_d2": 25204.767812641334, "mean_diff": 0.0072079097871993875, "se": 0.31783802378154985, "incumbent_mean": 95.54201511346659, "candidate_mean": 95.5492230232538, "retail_diff": -0.4216279529919995, "arb_diff": 0.4288358627695998 }, "by_spread": { "1": { "n": 131, "sum_d": 0.0, "sum_d2": 0.0, "mean_diff": 0.0, "se": 0.0, "incumbent_mean": 38.277936657049615, "candidate_mean": 38.277936657049615, "retail_diff": 0.0, "arb_diff": 0.0 }, "2": { "n": 120, "sum_d": -9.219568894700073, "sum_d2": 593.1455418617153, "mean_diff": -0.07682974078916728, "se": 0.20368406715749726, "incumbent_mean": 77.8558643979175, "candidate_mean": 77.77903465712832, "retail_diff": -0.9636309508383315, "arb_diff": 0.8868012099974998 }, "3": { "n": 133, "sum_d": 95.91874633739992, "sum_d2": 19602.246329387068, "mean_diff": 0.72119358148421, "se": 1.0548057072466617, "incumbent_mean": 146.55140363646615, "candidate_mean": 147.27259721795036, "retail_diff": -0.6784684162353379, "arb_diff": 1.3996619977195484 }, "4": { "n": 116, "sum_d": -83.09522254910016, "sum_d2": 5009.375941392552, "mean_diff": -0.7163381254232772, "se": 0.6091414845780005, "incumbent_mean": 120.02204692465259, "candidate_mean": 119.30570879922931, "retail_diff": -0.04260312962155187, "arb_diff": -0.6737349957896556 } } }, "failures": { "holdout": 0, "holdout_detail": [], "tuning_frozen_variant": 0, "n_failures_tuning": 0, "note": "8 failures in the unguarded first tuning runs of A and B, fixed by the guard before freezing; see tried." }, "check_values": { "310000": "17.116002", "310001": "20.017111", "310002": "2.390311", "310003": "241.484752" }, "check_values_match": true, "python_version": "3.12.3", "seconds_per_simulation": 0.777, "n_sha256_verified": "bf0d179f3e47d28819f6fa7e0d25548c2ced43374c216d36460b00a26a26ffae", "screen_only": true }