{ "key": "f-order-count-cap", "mech": "order count cap: keep at most the best-priced K resting orders per side, cancelling deeper ones (K=2 and K=3 screened)", "approach": "Overlay on N: after N's on_step, rebuild the resting book from own_orders plus N's cancels/places, rank each side by price (ties keep older), cancel existing orders beyond rank K and drop N's new placements beyond K. A conservative collateral guard skips placements the account cannot fund after a drop. K=3 frozen on tuning.", "tried": "Tuning 320000..320199: K=2 all spreads (-40.63, se 4.19), K=3 all spreads (-10.32, se 1.50), K=2 spreads 3-4 only (-34.01, se 4.23). Every variant loses edge at every spread: the cap removes retail fills and N relies on deep orders. A first overlay without the collateral guard failed 1 of 200 sims (seed 320151, insufficient collateral); the guard fixed it and all tuning runs were redone.", "next": "Cap hurts; do not promote. If anything, test K>=4 or capping only orders beyond a price distance from the fair-value estimate, but the K=3 loss suggests N's depth is worth more than its arbitrage exposure.", "spreads_acted": [ 1, 2, 3, 4 ], "off_equals_n": { "seeds": "310000..310009", "equal": true, "note": "CAP=None overlay actions identical to N every step (2000 steps, repr compared) and equal total_edge" }, "code": "class Strategy(_N):\n \"\"\"Order-count cap: keep at most CAP best-priced resting orders per side.\"\"\"\n CAP = 3\n SPREADS = (1, 2, 3, 4)\n\n def on_step(self, state):\n actions = list(super().on_step(state))\n cap = self.CAP\n if cap is None or getattr(self, \"s\", None) not in self.SPREADS:\n return actions\n live = [[o.order_id, o.side, o.price_ticks, None, o.remaining_quantity] for o in state.own_orders]\n for a in actions:\n if isinstance(a, CancelAll):\n live = []\n elif isinstance(a, CancelOrder):\n live = [x for x in live if x[0] != a.order_id]\n elif isinstance(a, PlaceOrder):\n live.append([None, a.side, a.price_ticks, a, a.quantity])\n drop_ids = set()\n drop_new = set()\n for side in (Side.BUY, Side.SELL):\n lst = [x for x in live if x[1] == side]\n lst.sort(key=lambda x: -x[2] if side == Side.BUY else x[2])\n for x in lst[cap:]:\n if x[0] is None:\n drop_new.add(id(x[3]))\n else:\n drop_ids.add(x[0])\n if not drop_ids and not drop_new:\n return actions\n cancels = [CancelOrder(i) for i in sorted(drop_ids)]\n out = cancels + [a for a in actions if id(a) not in drop_new]\n # conservative collateral guard: skip a placement the account cannot fund\n free = state.free_cash\n yes = state.yes_inventory\n resting = {o.order_id: o for o in state.own_orders}\n asks = sum(o.remaining_quantity for o in state.own_orders if o.side == Side.SELL)\n final = []\n for a in out:\n if isinstance(a, CancelAll):\n free = state.cash\n asks = 0.0\n resting = {}\n elif isinstance(a, CancelOrder):\n o = resting.pop(a.order_id, None)\n if o is not None:\n if o.side == Side.BUY:\n free += o.price_ticks / 100.0 * o.remaining_quantity\n else:\n asks -= o.remaining_quantity\n elif isinstance(a, PlaceOrder):\n if a.side == Side.BUY:\n need = a.price_ticks / 100.0 * a.quantity\n else:\n avail = max(0.0, yes - asks)\n need = (1.0 - a.price_ticks / 100.0) * max(0.0, a.quantity - avail)\n asks_add = a.quantity\n if need > free + 1e-6:\n continue\n free -= need\n if a.side == Side.SELL:\n asks += asks_add\n final.append(a)\n return final\n", "code_sha256": "3a013516071b1dd153574048a64925fc7bf0630936e01c5a8cca36c53f1b186c", "assembled_sha256": "407026518f77f9b68f9cb959cda3ea26b13b4fb1145bc6575913ce8c2902e245", "tuning": { "seeds": "320000..320199", "selected": "k3", "variants": { "k2": { "desc": "cap 2 per side, spreads 1-4", "mean_diff": -40.633085183775, "se": 4.192214792815564, "n": 200, "candidate_mean": 57.8376331307905, "incumbent_mean": 98.47071831456549 }, "k3": { "desc": "cap 3 per side, spreads 1-4", "mean_diff": -10.31507285485, "se": 1.4976024376707837, "n": 200, "candidate_mean": 88.1556454597155, "incumbent_mean": 98.47071831456549 }, "k2s34": { "desc": "cap 2 per side, spreads 3-4 only", "mean_diff": -34.0082316142735, "se": 4.232662617740361, "n": 200, "candidate_mean": 64.462486700292, "incumbent_mean": 98.47071831456549 } } }, "paired": { "seeds": "430000..430499", "n": 500, "complete": true, "missing_seeds": [], "overall": { "n": 500, "sum_d": -5359.1246263210005, "sum_d2": 519219.504106475, "mean_diff": -10.718249252642002, "se": 1.3604476903500318, "incumbent_mean": 95.54201511346659, "candidate_mean": 84.8237658608246, "retail_diff": -10.4096759077388, "arb_diff": -0.30857334489800003 }, "by_spread": { "1": { "n": 131, "sum_d": -98.44607642220004, "sum_d2": 14954.686593408876, "mean_diff": -0.7514967665816796, "se": 0.9347692662100039, "incumbent_mean": 38.277936657049615, "candidate_mean": 37.526439890467934, "retail_diff": -1.0693465390442747, "arb_diff": 0.31784977249313034 }, "2": { "n": 120, "sum_d": -1316.5579468472001, "sum_d2": 245823.95714595116, "mean_diff": -10.971316223726667, "se": 4.025301488301815, "incumbent_mean": 77.8558643979175, "candidate_mean": 66.88454817419083, "retail_diff": -8.018236705425833, "arb_diff": -2.9530795182775003 }, "3": { "n": 133, "sum_d": -1795.4348896592, "sum_d2": 97715.24162699885, "mean_diff": -13.499510448565413, "se": 2.045809246628051, "incumbent_mean": 146.55140363646615, "candidate_mean": 133.05189318790076, "retail_diff": -14.220862833642855, "arb_diff": 0.7213523850436088 }, "4": { "n": 116, "sum_d": -2148.6857133924, "sum_d2": 160725.61874011607, "mean_diff": -18.52315270165862, "se": 3.0107900057960246, "incumbent_mean": 120.02204692465259, "candidate_mean": 101.49889422299397, "retail_diff": -19.06198617007759, "arb_diff": 0.5388334684215519 } } }, "failures": [], "check_values": [ 17.116002, 20.017111, 2.390311, 241.484752 ], "python_version": "3.12.3", "seconds_per_simulation": 3.38 }