{ "key": "f-drift-skew", "approach": "Blind latent-score trend shrinkage. Competitor midpoint p is inverse-normal transformed to x=Phi^-1(p)*sqrt(steps_remaining). EWMA(128) of clipped x increments estimates score slope; shrink slope by 90% toward zero. Predicted dp=phi(z)*(z/(2*T)+0.1*slope/sqrt(T)). Resize N placements by 1 +/- clip(gain*100*dp,-0.30,0.30): positive drift reduces sells and increases buys. Preserve prices, cancellations, and IDs; cap increases by available collateral. No added orders. No source inspection of N. Reference seed 310000 edge 17.1160017503 matches 17.116002 rounded. Current best hash bf0d179f3e47d28819f6fa7e0d25548c2ced43374c216d36460b00a26a26ffae.", "tried": { "selection": "For each gain retain spreads whose tuning mean exceeds one SE; if none, retain the single spread with largest mean. Choose gain with largest all-200-seed masked tuning mean. No holdout-based changes.", "variants": [ { "gain": 8.0, "spreads": [ 1 ], "tuning": { "n": 200, "mean_diff": -2.8112894160675, "se": 1.142947729493399, "sum_d": -562.2578832135, "sum_d2": 53572.58422787244 }, "by_spread": { "1": { "n": 47, "mean_diff": -11.962933685393617, "se": 4.654895588548315, "sum_d": -562.2578832135, "sum_d2": 53572.58422787244 }, "2": { "n": 57, "mean_diff": -52.989859775278944, "se": 12.659452878171267, "sum_d": -3020.4220071909, "sum_d2": 671607.2356046593 }, "3": { "n": 53, "mean_diff": -74.62404931025283, "se": 15.627017005064882, "sum_d": -3955.0746134434003, "sum_d2": 968168.9712528032 }, "4": { "n": 43, "mean_diff": -75.64866548901628, "se": 17.881436374816204, "sum_d": -3252.8926160277, "sum_d2": 823537.8402704104 } } }, { "gain": 24.0, "spreads": [ 1 ], "tuning": { "n": 200, "mean_diff": -3.2258301936660003, "se": 1.2631738468640987, "sum_d": -645.1660387332, "sum_d2": 65586.40115025098 }, "by_spread": { "1": { "n": 47, "mean_diff": -13.726936994323404, "se": 5.122470838027282, "sum_d": -645.1660387332, "sum_d2": 65586.40115025098 }, "2": { "n": 57, "mean_diff": -64.45035890801404, "se": 15.103334678091432, "sum_d": -3673.6704577568, "sum_d2": 964898.7926400121 }, "3": { "n": 53, "mean_diff": -97.85951668322075, "se": 23.039946852052367, "sum_d": -5186.5543842107, "sum_d2": 1970546.405295615 }, "4": { "n": 43, "mean_diff": -94.57217578933721, "se": 21.95019291868184, "sum_d": -4066.6035589415, "sum_d2": 1254738.1569579944 } } } ] }, "next": "Do not promote: this frozen overlay loses to N on the 500-seed paired holdout. Preserve N. A separate future experiment could test whether quantity changes disrupt incumbent order management; that explanation is a hypothesis, not established here. Finalist seeds 900000..903999 were untouched.", "spreads_acted": [ 1 ], "off_equals_n": true, "code": "import math as _ds_math\nfrom statistics import NormalDist as _DSNormal\nfrom dataclasses import replace as _ds_replace\nfrom orderbook_pm_challenge.types import PlaceOrder as _DSPlace, CancelAll as _DSAll, CancelOrder as _DSCancel, Side as _DSSide\n\nclass Strategy:\n # Latent normal-score slope is shrunk 90% toward its zero-trend prior.\n spreads = (1,)\n gain = 8.0\n enabled = True\n\n def __init__(self):\n self.n = _N()\n self.s = None\n self.last_x = None\n self.slope = 0.0\n self.normal = _DSNormal()\n\n def on_step(self, state):\n actions = self.n.on_step(state)\n if self.s is None:\n b, a = state.competitor_best_bid_ticks, state.competitor_best_ask_ticks\n self.s = max(1, (a-b+1)//2) if a is not None and b is not None else 0\n if not self.enabled or self.s not in self.spreads:\n return actions\n b, a = state.competitor_best_bid_ticks, state.competitor_best_ask_ticks\n if a is None or b is None:\n return actions\n p = min(0.995, max(0.005, (a+b)/200.0))\n t = max(1, state.steps_remaining)\n z = self.normal.inv_cdf(p)\n x = z * _ds_math.sqrt(t)\n if self.last_x is not None:\n delta = min(1.0, max(-1.0, x-self.last_x))\n self.slope += (delta-self.slope)/128.0\n self.last_x = x\n drift = _ds_math.exp(-0.5*z*z)/_ds_math.sqrt(2*_ds_math.pi) * (z/(2*t) + 0.1*self.slope/_ds_math.sqrt(t))\n skew = min(0.30, max(-0.30, self.gain*100*drift))\n # Resize only N's placements. No added orders need hiding from N.\n # Maintain a collateral ledger to keep increases valid after cancellations.\n orders = {o.order_id: (o.side, o.price_ticks, o.remaining_quantity) for o in state.own_orders}\n def reserves():\n yes = state.yes_inventory\n cash = 0.0\n for side, tick, qty in orders.values():\n if side == _DSSide.BUY:\n cash += qty*tick/100.0\n else:\n covered = min(yes, qty)\n yes -= covered\n cash += (qty-covered)*(1-tick/100.0)\n return cash, yes\n out = []\n for i, action in enumerate(actions):\n if isinstance(action, _DSAll):\n orders.clear()\n elif isinstance(action, _DSCancel):\n orders.pop(action.order_id, None)\n elif isinstance(action, _DSPlace):\n used, yes = reserves()\n free = max(0.0, state.cash-used-1e-7)\n price = action.price_ticks/100.0\n capacity = free/price if action.side == _DSSide.BUY else yes+free/(1-price)\n factor = 1+skew if action.side == _DSSide.BUY else 1-skew\n qty = _ds_math.floor(min(action.quantity*factor, capacity)*100+1e-9)/100.0\n if qty < 0.01:\n continue\n action = _ds_replace(action, quantity=qty)\n orders[action.client_order_id or ('__drift', i)] = (action.side, action.price_ticks, qty)\n out.append(action)\n return out\n", "code_sha256": "ef39c3c81b0a1881e544e3c2f40faf9d3401143ca906431d044ada27c53cd834", "assembled_sha256": "08cb7e9758be0000f74cc90dc9166e92645b6e599ef042abeab699afb4623f2c", "tuning": { "n": 200, "mean_diff": -2.8112894160675, "se": 1.142947729493399 }, "paired": { "n": 500, "mean_diff": -2.0238095518382, "se": 0.6009124229704581, "sum_d": -1011.9047759191, "sum_d2": 92141.28970107235, "first_seed": 410000, "last_seed": 410499, "complete": true, "incumbent_mean": 104.4137531881468, "retail_diff": -1.8872337652555997, "arb_diff": -0.1365757865886, "failures": [] }, "by_spread": { "1": { "mean_diff": -8.799171964513913, "se": 2.5197521091442283, "sum_d": -1011.9047759191, "sum_d2": 92141.28970107235, "runs": 115, "incumbent_mean": 28.495154459731303, "retail_diff": -8.205364196763476, "arb_diff": -0.5938077677765218 }, "2": { "mean_diff": 0.0, "se": 0.0, "sum_d": 0.0, "sum_d2": 0.0, "runs": 116, "incumbent_mean": 97.76626247947759, "retail_diff": 0.0, "arb_diff": 0.0 }, "3": { "mean_diff": 0.0, "se": 0.0, "sum_d": 0.0, "sum_d2": 0.0, "runs": 138, "incumbent_mean": 128.4054514027384, "retail_diff": 0.0, "arb_diff": 0.0 }, "4": { "mean_diff": 0.0, "se": 0.0, "sum_d": 0.0, "sum_d2": 0.0, "runs": 131, "incumbent_mean": 151.67248160310686, "retail_diff": 0.0, "arb_diff": 0.0 } }, "seconds_per_simulation": 0.32827966268733144 }