{ "key": "f-late-widen", "mech": "f-late-widen", "approach": "Overlay on N: when steps_remaining <= W (W frozen at 100), every N order is placed one tick further from the rival midpoint (bids -1, asks +1, clamped to 1..99), sizes unchanged; orders already resting when the window opens are cancelled and re-placed one tick out under the same id. N sees its own orders at the prices it asked for, so its logic runs unchanged. Result: a large loss, almost all retail edge, because late fills at N's prices are worth a lot and one tick back gives them to the rival.", "tried": [ "W=100 tuning d=-15.56 se 1.66", "W=200 tuning d=-28.06 se 2.95", "W=400 tuning d=-46.75 se 5.11", "froze W=100 (least negative), all spreads; holdout d=-14.15 se 1.15" ], "next": "Drop late widening: it loses on every spread and the loss grows with W, so late retail fills are valuable and N should keep them. The opposite direction (late tightening, or a larger late size where the arb risk is small) is untested here. A rerun would also need a fix: place the re-placed orders before N's new orders, since 4 holdout runs failed on collateral.", "spreads_acted": [ 1, 2, 3, 4 ], "off_equals_n": { "seeds": "310000..310009", "steps_compared": 20000, "mismatched_steps": 0, "equal": true, "how": "LATE_W=0; N and overlay fed the identical StepState every step, repr of action lists compared" }, "code": "# f-late-widen overlay: in the last W steps every N order sits one tick\n# further from the rival midpoint (bids -1, asks +1); sizes unchanged.\n# N is shown its own orders at the prices it asked for, so its logic is unchanged.\nfrom dataclasses import replace as _dc_replace\nfrom orderbook_pm_challenge.types import CancelAll as _CA, CancelOrder as _CO, PlaceOrder as _PO, Side as _Side\n\nLATE_W = 100\nACT_SPREADS = (1, 2, 3, 4)\n\n\nclass Strategy(_N):\n def __init__(self, *a, **k):\n super().__init__(*a, **k)\n self._lw_shift = {}\n self._lw_n = 0\n\n def _lw_tick(self, side, px):\n q = px - 1 if side == _Side.BUY else px + 1\n return q if 1 <= q <= 99 else px\n\n def on_step(self, state):\n sh = self._lw_shift\n if sh:\n live = {o.order_id for o in state.own_orders}\n for oid in list(sh):\n if oid not in live:\n del sh[oid]\n view = tuple(_dc_replace(o, price_ticks=o.price_ticks - sh[o.order_id]) if o.order_id in sh else o\n for o in state.own_orders)\n state = _dc_replace(state, own_orders=view)\n acts = super().on_step(state)\n if LATE_W <= 0 or state.steps_remaining > LATE_W or getattr(self, \"s\", None) not in ACT_SPREADS:\n return acts\n acts = list(acts or [])\n out = []\n cancelled = set()\n wiped = False\n for a in acts:\n if isinstance(a, _PO):\n oid = a.client_order_id\n if oid is None:\n self._lw_n += 1\n oid = \"lw%d\" % self._lw_n\n px = self._lw_tick(a.side, a.price_ticks)\n sh[oid] = px - a.price_ticks\n out.append(_PO(side=a.side, price_ticks=px, quantity=a.quantity, client_order_id=oid))\n else:\n if isinstance(a, _CA):\n wiped = True\n elif isinstance(a, _CO):\n cancelled.add(a.order_id)\n out.append(a)\n pre, post = [], []\n if not wiped:\n for o in state.own_orders:\n if o.order_id in sh or o.order_id in cancelled:\n continue\n px = self._lw_tick(o.side, o.price_ticks)\n if px == o.price_ticks:\n continue\n pre.append(_CO(o.order_id))\n post.append(_PO(side=o.side, price_ticks=px, quantity=o.remaining_quantity, client_order_id=o.order_id))\n sh[o.order_id] = px - o.price_ticks\n return pre + out + post\n", "code_sha256": "6d2f94dae65d158bf0b26183a5f70750229109979d6054769d5127bc0c2974c6", "assembled_sha256": "65be48cbb1500035cc0ab5ede6aa060fa3f67e1c02a1b99ced01be5ed6173404", "n_sha256": "bf0d179f3e47d28819f6fa7e0d25548c2ced43374c216d36460b00a26a26ffae", "tuning": { "seeds": "320000..320199", "selected": "W100", "variants": { "W100": { "W": 100, "overall": { "n": 200, "mean_diff": -15.557867378893507, "se": 1.6582036835384026, "retail_diff": -16.231554870233467, "arb_diff": 0.6736874913230118 }, "by_spread_mean_diff": { "1": [ -10.127, 2.607 ], "2": [ -14.43, 3.445 ], "3": [ -16.691, 3.457 ], "4": [ -20.818, 3.565 ] } }, "W200": { "W": 200, "overall": { "n": 200, "mean_diff": -28.055027143984994, "se": 2.949801074746473, "retail_diff": -29.2322746893921, "arb_diff": 1.17724754539449 }, "by_spread_mean_diff": { "1": [ -18.297, 5.334 ], "2": [ -22.636, 5.413 ], "3": [ -30.985, 6.261 ], "4": [ -39.943, 6.214 ] } }, "W400": { "W": 400, "overall": { "n": 200, "mean_diff": -46.749935029562984, "se": 5.112054188704644, "retail_diff": -49.54590128003308, "arb_diff": 2.795966250465007 }, "by_spread_mean_diff": { "1": [ -22.518, 6.629 ], "2": [ -35.087, 8.803 ], "3": [ -54.596, 11.034 ], "4": [ -74.021, 12.129 ] } } } }, "paired": { "seeds": "430000..430499", "n": 500, "complete": true, "missing_seeds": [], "overall": { "n": 500, "sum_d": -7074.6297460202, "sum_d2": 432628.473255611, "mean_diff": -14.1492594920404, "se": 1.1544593355294224, "incumbent_mean": 95.54587347343701, "candidate_mean": 81.3966139813965, "retail_diff": -15.234180397404089, "arb_diff": 1.0849209053983841 }, "by_spread": { "1": { "n": 131, "sum_d": -1307.5303462037, "sum_d2": 61055.9873498945, "mean_diff": -9.981147680944275, "se": 1.6789486872704973, "incumbent_mean": 38.27342786619541, "candidate_mean": 28.29228018525116, "retail_diff": -10.199723174550428, "arb_diff": 0.21857549363282658 }, "2": { "n": 120, "sum_d": -1712.8634670041, "sum_d2": 70081.8273283534, "mean_diff": -14.273862225034167, "se": 1.7876139792467771, "incumbent_mean": 77.87680751697249, "candidate_mean": 63.602945291938354, "retail_diff": -15.459700668155065, "arb_diff": 1.1858384431316722 }, "3": { "n": 133, "sum_d": -2346.811966464599, "sum_d2": 219282.29516225538, "mean_diff": -17.645202755372928, "se": 3.183034907190754, "incumbent_mean": 146.5514036364661, "candidate_mean": 128.9062008810932, "retail_diff": -20.566060521201507, "arb_diff": 2.920857765877482 }, "4": { "n": 116, "sum_d": -1707.423966347799, "sum_d2": 82208.36341510764, "mean_diff": -14.719172123687922, "se": 2.068477196760355, "incumbent_mean": 120.02210431517413, "candidate_mean": 105.30293219148615, "retail_diff": -14.573071839116295, "arb_diff": -0.14610028451897605 } }, "overall_excluding_failed_seeds": { "n": 496, "mean_diff": -14.117619259839715, "se": 1.156910229156101, "retail_diff": -15.166810167916239, "arb_diff": 1.0491909081112851 } }, "failures": { "count": 4, "note": "Failed runs counted with the edge the engine reported at failure (SimulationResult with failed=True); N had no failures. Probable cause: re-placed resting orders were emitted after N's new orders, which could have used the collateral the cancels freed.", "runs": [ { "seed": 430009, "strategy": "candidate W=100", "error": "orderbook_pm_challenge.market.OrderBookError: Insufficient free cash for order collateral", "spread": 1, "total_edge_at_failure": 4.260062 }, { "seed": 430171, "strategy": "candidate W=100", "error": "orderbook_pm_challenge.market.OrderBookError: Insufficient free cash for order collateral", "spread": 1, "total_edge_at_failure": -11.363443 }, { "seed": 430185, "strategy": "candidate W=100", "error": "orderbook_pm_challenge.market.OrderBookError: Insufficient free cash for order collateral", "spread": 1, "total_edge_at_failure": -6.619581 }, { "seed": 430212, "strategy": "candidate W=100", "error": "orderbook_pm_challenge.market.OrderBookError: Insufficient free cash for order collateral", "spread": 3, "total_edge_at_failure": 343.390495 } ] }, "check_values": { "310000": 17.116002, "310001": 20.017111, "310002": 2.390311, "310003": 241.484752 }, "python_version": "3.10.12", "seconds_per_simulation": 0.787, "simulator": "danrobinson/prediction-market-challenge @ b244530ee2a6247b5f930830eeca9dd950da0e7e, unmodified, serial, default ChallengeConfig/ParameterVariance, 2000 steps", "screen_only": true }