{ "key": "f-probe-stopping", "approach": "Sequential value-of-information stopping over a 201-point probability posterior. Probe survival/sweeps use a noisy Bernoulli likelihood (0.12 retail sweep hazard; 0.995 toxic sweep likelihood), approximate diffusion and jump prediction, and last-step competitor bounds. Add 0.01 shares one tick inside N's best same-side quote only when estimated interval-variance reduction, valued by nearby backing exposure, exceeds expected arbitrage loss. Stop each side below the value threshold and restart after competitor quote movement or a one-tick posterior-mean shift. A just-swept overlay probe shrinks retained/new orders behind it to 35% for that callback. Added orders are hidden from N via dataclasses.replace; N is called every step. These posterior/value models are heuristic, not calibrated guarantees.", "tried": [ { "variant": 1, "status": "aborted diagnostic, not a measurement", "n_completed": 19, "description": "An extra self-crossing guard prevented all additions in these initial tuning seeds. Removed the unnecessary guard after examining callback actions, not incumbent source. No complete score claimed." }, { "variant": 2, "status": "complete", "description": "Corrected guard; remaining parameters unchanged. Tuned on 310000..310199, selecting spreads with positive per-spread tuning mean (fallback: highest mean if none positive). Frozen before holdout. No other variants.", "all_spreads_tuning": { "1": { "n": 47, "mean_diff": -1.0669208764148936, "se": 0.6189473385318448 }, "2": { "n": 57, "mean_diff": 0.048753446861402004, "se": 0.12634558802317827 }, "3": { "n": 53, "mean_diff": 0.012579087616979442, "se": 0.014121497390995235 }, "4": { "n": 43, "mean_diff": -0.04226889511628031, "se": 0.034103838704347994 } } }, { "incumbent_sha256": "bf0d179f3e47d28819f6fa7e0d25548c2ced43374c216d36460b00a26a26ffae", "seed_310000_edge": 17.1160017503, "expected_rounded_edge": 17.116002, "matches": true }, { "off_equality_seeds": "310000..310009", "assembled_equality_seeds": "310000..310004", "assembled_equals_s1": true, "off_callbacks_checked": 20000, "synthetic_checks": "informative probe placement, swept backing shrink, order hiding, excluded-spread identity, sequential survival stopping and move reset", "blindness": "No incumbent source, other strategy source, earlier network result or leaderboard was viewed, printed, or searched; only opaque hash/loading/required assembly transformation and callback actions were used." } ], "next": "Retain N unless a later authorized finalist measurement establishes a gain. N already emits tiny inside orders, so incremental probing often fails this value test; do not interpret this single overlay holdout as a leaderboard or finalist promotion.", "spreads_acted": [ 2, 3 ], "off_equals_n": true, "code": "from dataclasses import replace as _probe_replace\nfrom math import sqrt as _probe_sqrt\nfrom orderbook_pm_challenge.types import PlaceOrder as _PP, CancelOrder as _PC, CancelAll as _PA, Side as _PS\n\nclass Strategy(BaseStrategy):\n # The incumbent is opaque; only callback inputs and returned actions are used.\n SPREADS = (2, 3)\n ENABLED = True\n SHRINK = 0.35\n\n def __init__(self):\n self.n = _N()\n self.spread = None\n self.probes = {}\n self.weights = None\n self.last_quotes = None\n self.anchor = None\n self.stopped = set()\n self.serial = 0\n\n def on_step(self, state):\n if self.spread is None:\n b, a = state.competitor_best_bid_ticks, state.competitor_best_ask_ticks\n self.spread = max(1, (a-b+1)//2) if b is not None and a is not None else 1\n if not self.ENABLED or self.spread not in self.SPREADS:\n return self.n.on_step(state)\n own = {o.order_id:o for o in state.own_orders}\n visible = tuple(o for o in state.own_orders if o.order_id not in self.probes)\n clean = _probe_replace(state, own_orders=visible)\n actions = list(self.n.on_step(clean))\n b = state.competitor_best_bid_ticks\n a = state.competitor_best_ask_ticks\n lo, hi = (b if b is not None else 0), (a if a is not None else 100)\n if self.weights is None:\n self.weights = [1.0 if lo <= i*.5 <= hi else 0.000001 for i in range(201)]\n w = self.weights\n # Diffusion plus a small jump hazard, followed by the last-step bracket.\n mass = sum(w)\n w = [(.84*w[i]+.08*w[max(0,i-1)]+.08*w[min(200,i+1)])*.998 + mass*.002/201 for i in range(201)]\n w = [v*(1 if lo <= i*.5 <= hi else .001) for i,v in enumerate(w)]\n swept = {}\n for oid,(side, tick, quantity) in self.probes.items():\n filled = oid not in own or own[oid].remaining_quantity < quantity-.005\n # A retail arrival can also sweep a tiny probe: use a noisy likelihood.\n for i in range(201):\n toxic = i*.5 < tick if side == _PS.BUY else i*.5 > tick\n pf = .995 if toxic else .12\n w[i] *= pf if filled else 1-pf\n if filled:\n swept[side] = tick\n self.stopped.discard(side)\n z = sum(w)\n w = [v/z for v in w]\n self.weights = w\n mean = sum(i*.5*v for i,v in enumerate(w))\n variance = sum((i*.5-mean)**2*v for i,v in enumerate(w))\n quotes = (b,a)\n if self.last_quotes != quotes or (self.anchor is not None and abs(mean-self.anchor)>1):\n self.stopped.clear()\n self.anchor = mean\n if self.anchor is None:\n self.anchor = mean\n self.last_quotes = quotes\n # Release surviving probes before the incumbent's action list.\n out = [_PC(oid) for oid in self.probes if oid in own]\n self.probes = {}\n # Track the book resulting from N's actions, without inspecting N itself.\n live = dict((o.order_id,o) for o in visible)\n revised = []\n for action in actions:\n if isinstance(action,_PA):\n live.clear()\n elif isinstance(action,_PC):\n live.pop(action.order_id,None)\n elif isinstance(action,_PP):\n tick = swept.get(action.side)\n behind = tick is not None and (action.price_tickstick)\n if behind:\n q = max(.01, int(action.quantity*self.SHRINK*100)/100)\n action = _probe_replace(action,quantity=q)\n revised.append(action)\n # Existing resting orders also shrink if N chose to retain them.\n for oid,o in list(live.items()):\n tick = swept.get(o.side)\n behind = tick is not None and (o.price_tickstick)\n if behind:\n revised.extend([_PC(oid),_PP(o.side,o.price_ticks,max(.01,int(o.remaining_quantity*self.SHRINK*100)/100),oid)])\n out.extend(revised)\n # Build a reservation-conservative view of final orders for probe budgeting.\n book = {oid:(o.side,o.price_ticks,o.remaining_quantity) for oid,o in visible_dict(visible).items()}\n for act in revised:\n if isinstance(act,_PA): book.clear()\n elif isinstance(act,_PC): book.pop(act.order_id,None)\n elif isinstance(act,_PP): book[act.client_order_id or ('new',len(book))]=(act.side,act.price_ticks,act.quantity)\n reserve = sum(q*(p/100 if side==_PS.BUY else 1-p/100) for side,p,q in book.values())\n budget = state.cash-reserve\n for side in (_PS.BUY,_PS.SELL):\n candidates = [(p,q) for s,p,q in book.values() if s==side]\n if not candidates or side in self.stopped: continue\n best = max(p for p,q in candidates) if side==_PS.BUY else min(p for p,q in candidates)\n tick = best+(1 if side==_PS.BUY else -1)\n other = [p for s,p,q in book.values() if s!=side]\n if not 1<=tick<=99: continue\n exposure = sum(q for p,q in candidates if abs(p-best)<=1)\n pf = [(.995 if (i*.5tick) else .12) for i in range(201)]\n prob = sum(v*f for v,f in zip(w,pf))\n mfill = sum(i*.5*v*f for i,(v,f) in enumerate(zip(w,pf)))/max(prob,1e-12)\n msurvive = (mean-prob*mfill)/max(1-prob,1e-12)\n reduction = prob*(mfill-mean)**2+(1-prob)*(msurvive-mean)**2\n loss = .01*.01*sum(v*max(0,tick-i*.5 if side==_PS.BUY else i*.5-tick) for i,v in enumerate(w))\n value = .15*.01*min(exposure,20)*reduction/max(1,_probe_sqrt(variance))\n if value <= loss+.00001 or reduction < .002:\n self.stopped.add(side)\n continue\n if budget < .02: continue\n self.serial += 1\n oid = '__sequential_probe_%d'%self.serial\n out.append(_PP(side,tick,.01,oid))\n self.probes[oid]=(side,tick,.01)\n budget -= .01\n return out\n\ndef visible_dict(orders):\n return {o.order_id:o for o in orders}\n", "code_sha256": "7fa10543f5349a27a9f5e739cf38e7fb62bfe02144cbad8a44a6202030a7745a", "assembled_sha256": "46434cb8ca37b6f4fd6b7e7e7d5a4ebf60dde34edfb20076eca024ee4ba77c8c", "tuning": { "n": 200, "mean_diff": 0.017228190573999123, "se": 0.036002742279609735 }, "paired": { "first_seed": 410000, "last_seed": 410499, "n": 500, "complete": true, "incumbent_mean": 104.4137531881468, "mean_diff": 0.016767629280000168, "se": 0.030340765993081643, "sum_d": 8.383814640000084, "sum_d2": 229.82081591704718, "retail_diff": -0.053693320717199906, "arb_diff": 0.07046094999859989, "failures": [] }, "by_spread": { "1": { "runs": 115, "incumbent_mean": 28.495154459731303, "mean_diff": 0.0, "se": 0.0, "sum_d": 0.0, "sum_d2": 0.0, "retail_diff": 0.0, "arb_diff": 0.0 }, "2": { "runs": 116, "incumbent_mean": 97.76626247947759, "mean_diff": 0.04460301534482666, "se": 0.12894293783950417, "sum_d": 5.173949779999893, "sum_d2": 222.0253652186517, "retail_diff": -0.261482267302586, "arb_diff": 0.3060852826525858 }, "3": { "runs": 138, "incumbent_mean": 128.4054514027384, "mean_diff": 0.02325989028985646, "se": 0.020208358226073298, "sum_d": 3.2098648600001916, "sum_d2": 7.7954506983954825, "retail_diff": 0.02525567136594221, "arb_diff": -0.0019957810753623525 }, "4": { "runs": 131, "incumbent_mean": 151.67248160310686, "mean_diff": 0.0, "se": 0.0, "sum_d": 0.0, "sum_d2": 0.0, "retail_diff": 0.0, "arb_diff": 0.0 } }, "seconds_per_simulation": 0.8543097922010348 }