{ "key": "f-exposure-ratchet", "mech": "Exposure ratchet: a reported own buy (sell) followed by a visible rival midpoint decline (rise) of >=2 ticks within 3 steps resets a 0.5 size multiplier. Hold 50 steps, then recover linearly over 50 steps. Both sides are scaled; repeat triggers restart the clock.", "approach": "Blind subclass overlay; N source was downloaded, hash-checked and assembled by script without viewing it. Anchor fills at the preceding callback midpoint; require both rival quotes. Apply on spreads 1-4. Scale new orders; reduce surviving orders on multiplier drops, never enlarge retained orders during recovery. Preserve N prices and cancellations. Tune 50/100-step hold/recovery plus a 50-step variant with conservative collateral capping after an observed failure; freeze the best failure-free paired mean. Screen only; no promotion.", "tried": [ { "name": "h50", "hold_steps": 50, "linear_recovery_steps": 50 }, { "name": "h100", "hold_steps": 100, "linear_recovery_steps": 100 }, { "name": "h50safe", "hold_steps": 50, "linear_recovery_steps": 50, "collateral_cap": true } ], "next": "Reject this overlay: the frozen screen loses edge and has holdout runtime failures. Keep N; no promotion. Do not tune further on this holdout. Statistics include partial engine edge on failed runs, which are explicitly listed.", "spreads_acted": [ 1, 2, 3, 4 ], "off_equals_n": true, "code": "from dataclasses import replace as _ratchet_replace\nfrom orderbook_pm_challenge.types import PlaceOrder as _RP, CancelAll as _RCA, CancelOrder as _RC\n\nclass Strategy(_N):\n ratchet_enabled = True\n ratchet_hold = 50\n ratchet_spreads = (1, 2, 3, 4)\n\n def on_step(self, state):\n actions = super().on_step(state)\n if not self.ratchet_enabled or self.s not in self.ratchet_spreads:\n return actions\n t = state.step\n if t == 0:\n self._er_pending = []\n self._er_mid = None\n self._er_last_trigger = None\n self._er_scales = {}\n bid = state.competitor_best_bid_ticks\n ask = state.competitor_best_ask_ticks\n mid = (bid + ask) / 2.0 if bid is not None and ask is not None else None\n # Fills reported at t happened at t-1; use its pre-fill midpoint.\n if self._er_mid is not None and t > 0:\n if state.buy_filled_quantity > 0:\n self._er_pending.append((t-1, self._er_mid, 1))\n if state.sell_filled_quantity > 0:\n self._er_pending.append((t-1, self._er_mid, -1))\n pending = []\n hit = False\n for filled_at, ref, direction in self._er_pending:\n if t - filled_at <= 3:\n if mid is not None and direction * (mid - ref) <= -2.0:\n hit = True\n else:\n pending.append((filled_at, ref, direction))\n self._er_pending = pending\n self._er_mid = mid\n if hit:\n self._er_last_trigger = t\n scale = 1.0\n if self._er_last_trigger is not None:\n age = t - self._er_last_trigger\n hold = self.ratchet_hold\n if age < hold:\n scale = 0.5\n elif age < 2 * hold:\n scale = 0.5 + 0.5 * (age - hold) / hold\n self._er_scales[t] = scale\n # N's cancellations remain authoritative. Reduce retained orders only\n # when the multiplier falls; recovery applies to newly placed orders.\n actions = list(actions)\n all_cancelled = any(isinstance(a, _RCA) for a in actions)\n cancelled = {a.order_id for a in actions if isinstance(a, _RC)}\n retained = []\n extra_cancels = []\n if not all_cancelled:\n for o in state.own_orders:\n old = self._er_scales.get(o.submitted_step, 1.0)\n if o.order_id not in cancelled and scale < old - 1e-12:\n extra_cancels.append(_RC(o.order_id))\n q = int((o.remaining_quantity * scale / old + 1e-9) * 100) / 100\n if q >= 0.01:\n retained.append(_RP(o.side, o.price_ticks, q, o.order_id))\n out = list(extra_cancels)\n for a in actions:\n if isinstance(a, _RP) and scale < 1.0:\n q = int((a.quantity * scale + 1e-9) * 100) / 100\n if q >= 0.01:\n out.append(_ratchet_replace(a, quantity=q))\n else:\n out.append(a)\n out.extend(retained)\n return self._er_cap(state, out) if scale < 1.0 else out\n\n def _er_cap(self, state, actions):\n # Lower bound on free collateral; no private simulator state is used.\n from orderbook_pm_challenge.types import Side as _RS\n cash = max(0.0, state.free_cash)\n yes = max(0.0, state.yes_inventory)\n live = {o.order_id: (o.side, o.price_ticks, o.remaining_quantity)\n for o in state.own_orders}\n out = []\n for i, a in enumerate(actions):\n if isinstance(a, _RCA):\n live.clear()\n cash = max(0.0, state.cash)\n out.append(a)\n elif isinstance(a, _RC):\n o = live.pop(a.order_id, None)\n if o is not None:\n side, tick, qty = o\n cash += qty * tick / 100 if side == _RS.BUY else max(0.0, qty - yes) * (1 - tick / 100)\n out.append(a)\n elif isinstance(a, _RP):\n price = a.price_ticks / 100\n free_yes = max(0.0, yes - sum(q for side, tick, q in live.values() if side == _RS.SELL))\n limit = cash / price if a.side == _RS.BUY else free_yes + cash / (1 - price)\n qty = int((min(a.quantity, max(0.0, limit)) + 1e-10) * 100) / 100\n if qty < 0.01:\n continue\n cost = qty * price if a.side == _RS.BUY else max(0.0, qty - free_yes) * (1 - price)\n cash = max(0.0, cash - cost)\n live[a.client_order_id if a.client_order_id else ('new', i)] = (a.side, a.price_ticks, qty)\n out.append(_ratchet_replace(a, quantity=qty))\n else:\n out.append(a)\n return out\n", "code_sha256": "4cc20e7b283dd38b42589255fdc0a04877b2588568d66146e6964167be8af1da", "assembled_sha256": "4e51e7fe7f19e8334bcece9f5e250d8781aedf9fadd6ee54a2435651d4fe8798", "tuning": { "seeds": { "start": 320000, "end_inclusive": 320199 }, "selected": "h50safe", "variants": [ { "name": "h50", "failed_seeds": [ 320151 ], "partial_edges_included_for_failed_runs": true, "collateral_cap": false, "hold_steps": 50, "linear_recovery_steps": 50, "spreads_acted": [ 1, 2, 3, 4 ], "code_sha256": "afb8e6e4ca62e2c14e7c4e52d0ea33cdb5b1dc2951ac38b1e97e0e26cbfe502c", "assembled_sha256": "98d376d9d085b7b0fe3ecbabfb6c29249d4f33a873f06dc5df705e6e7f4f3a85", "overall": { "n": 200, "sum_d": -6562.7751194511, "sum_d2": 612001.2257423342, "mean_diff": -32.8138755972555, "se": 3.1569144793846244, "incumbent_mean": 98.47071831456549, "candidate_mean": 65.65684271731, "retail_diff": -14.3049098142845, "arb_diff": -18.5089657829865 }, "by_spread": { "1": { "n": 50, "sum_d": -282.8099905367, "sum_d2": 9882.367530470585, "mean_diff": -5.656199810734, "se": 1.8386705174234261, "incumbent_mean": 25.711261479664, "candidate_mean": 20.05506166893, "retail_diff": -4.287810675136001, "arb_diff": -1.368389135634 }, "2": { "n": 50, "sum_d": -956.2372542968, "sum_d2": 33572.43074258464, "mean_diff": -19.124745085936, "se": 2.4977244347673486, "incumbent_mean": 61.219908510538, "candidate_mean": 42.095163424602, "retail_diff": -13.875054605644, "arb_diff": -5.249690480283999 }, "3": { "n": 48, "sum_d": -2201.5129310697, "sum_d2": 167613.8884037241, "mean_diff": -45.86485273061874, "se": 5.435054252914001, "incumbent_mean": 130.8342429047854, "candidate_mean": 84.96939017416666, "retail_diff": -21.183445139122917, "arb_diff": -24.68140759157917 }, "4": { "n": 52, "sum_d": -3122.2149435479, "sum_d2": 400932.5390655548, "mean_diff": -60.04259506822885, "se": 8.97177263328206, "incumbent_mean": 174.37579815333268, "candidate_mean": 114.33320308510385, "retail_diff": -18.00064100269231, "arb_diff": -42.041954065492305 } } }, { "name": "h100", "failed_seeds": [ 320109, 320151, 320154 ], "partial_edges_included_for_failed_runs": true, "collateral_cap": false, "hold_steps": 100, "linear_recovery_steps": 100, "spreads_acted": [ 1, 2, 3, 4 ], "code_sha256": "c78cbe292f307ca5429b10d567cf7e18e8ccb92996a46aeedc47b6b69d261ea0", "assembled_sha256": "f9dbe4ac8f3a0a25e4b2037dd46d5b305a320244532092169ca22b89d329979c", "overall": { "n": 200, "sum_d": -8381.6510248307, "sum_d2": 1037304.8642880145, "mean_diff": -41.9082551241535, "se": 4.1517826127777315, "incumbent_mean": 98.47071831456549, "candidate_mean": 56.562463190412, "retail_diff": -17.088103982312, "arb_diff": -24.8201511418775 }, "by_spread": { "1": { "n": 50, "sum_d": -382.17369180290007, "sum_d2": 20038.351294346478, "mean_diff": -7.643473836058002, "se": 2.643221340174061, "incumbent_mean": 25.711261479664, "candidate_mean": 18.067787643606, "retail_diff": -5.4335362531980005, "arb_diff": -2.2099375828680006 }, "2": { "n": 50, "sum_d": -1321.6472198009, "sum_d2": 77600.43487869423, "mean_diff": -26.432944396018, "se": 4.173062664599171, "incumbent_mean": 61.219908510538, "candidate_mean": 34.78696411452, "retail_diff": -17.370972357935997, "arb_diff": -9.06197203808 }, "3": { "n": 48, "sum_d": -2593.6856555130003, "sum_d2": 226245.2131146214, "mean_diff": -54.035117823187505, "se": 6.177598482854461, "incumbent_mean": 130.8342429047854, "candidate_mean": 76.79912508159792, "retail_diff": -24.3362038190625, "arb_diff": -29.698914004316666 }, "4": { "n": 52, "sum_d": -4084.1444577139, "sum_d2": 713420.8650003524, "mean_diff": -78.54123957142116, "se": 12.167866477898423, "incumbent_mean": 174.37579815333268, "candidate_mean": 95.83455858191154, "retail_diff": -21.33187658828269, "arb_diff": -57.20936298309424 } } }, { "name": "h50safe", "failed_seeds": [], "partial_edges_included_for_failed_runs": true, "collateral_cap": true, "hold_steps": 50, "linear_recovery_steps": 50, "spreads_acted": [ 1, 2, 3, 4 ], "code_sha256": "4cc20e7b283dd38b42589255fdc0a04877b2588568d66146e6964167be8af1da", "assembled_sha256": "4e51e7fe7f19e8334bcece9f5e250d8781aedf9fadd6ee54a2435651d4fe8798", "overall": { "n": 200, "sum_d": -6557.5581757957, "sum_d2": 611928.0832141116, "mean_diff": -32.7877908789785, "se": 3.1579851611629355, "incumbent_mean": 98.47071831456549, "candidate_mean": 65.682927435587, "retail_diff": -14.2788250960075, "arb_diff": -18.5089657829865 }, "by_spread": { "1": { "n": 50, "sum_d": -277.5930468813, "sum_d2": 9809.225002248071, "mean_diff": -5.5518609376259995, "se": 1.8370414334046437, "incumbent_mean": 25.711261479664, "candidate_mean": 20.159400542038, "retail_diff": -4.183471802028, "arb_diff": -1.368389135634 }, "2": { "n": 50, "sum_d": -956.2372542968, "sum_d2": 33572.43074258464, "mean_diff": -19.124745085936, "se": 2.4977244347673486, "incumbent_mean": 61.219908510538, "candidate_mean": 42.095163424602, "retail_diff": -13.875054605644, "arb_diff": -5.249690480283999 }, "3": { "n": 48, "sum_d": -2201.5129310697, "sum_d2": 167613.8884037241, "mean_diff": -45.86485273061874, "se": 5.435054252914001, "incumbent_mean": 130.8342429047854, "candidate_mean": 84.96939017416666, "retail_diff": -21.183445139122917, "arb_diff": -24.68140759157917 }, "4": { "n": 52, "sum_d": -3122.2149435479, "sum_d2": 400932.5390655548, "mean_diff": -60.04259506822885, "se": 8.97177263328206, "incumbent_mean": 174.37579815333268, "candidate_mean": 114.33320308510385, "retail_diff": -18.00064100269231, "arb_diff": -42.041954065492305 } } } ], "scoreboard": { "current_best_code_sha256": "bf0d179f3e47d28819f6fa7e0d25548c2ced43374c216d36460b00a26a26ffae", "candidate_paired_differences": { "h50": { "mean_diff": -32.8138755972555, "se": 3.1569144793846244 }, "h100": { "mean_diff": -41.9082551241535, "se": 4.1517826127777315 }, "h50safe": { "mean_diff": -32.7877908789785, "se": 3.1579851611629355 } } }, "selection_rule": "Largest overall paired mean difference among variants with no failures on all 200 tuning seeds; no per-spread tuning." }, "paired": { "seeds": { "start": 430000, "end_inclusive": 430499 }, "n": 500, "complete": true, "missing_seeds": [], "failed_seeds": [ 430009, 430108, 430216 ], "statistics_include_partial_edges_on_failed_runs": true, "complete_means": "All 500 requested seed pairs attempted once; failures are not successful 2000-step runs", "overall": { "n": 500, "sum_d": -17318.8369245721, "sum_d2": 1689723.700242409, "mean_diff": -34.6376738491442, "se": 2.089998586073646, "incumbent_mean": 95.54201511346659, "candidate_mean": 60.9043412643224, "retail_diff": -14.0172621364234, "arb_diff": -20.6204117127372 }, "by_spread": { "1": { "n": 131, "sum_d": -573.5923277093, "sum_d2": 85042.83198770859, "mean_diff": -4.378567387093893, "se": 2.2014157528683014, "incumbent_mean": 38.277936657049615, "candidate_mean": 33.89936926995573, "retail_diff": 0.2571812686832072, "arb_diff": -4.635748655753435 }, "2": { "n": 120, "sum_d": -3034.5576480661, "sum_d2": 151609.70196744482, "mean_diff": -25.287980400550833, "se": 2.2897877600142116, "incumbent_mean": 77.8558643979175, "candidate_mean": 52.56788399736667, "retail_diff": -16.186723266488332, "arb_diff": -9.101257134088334 }, "3": { "n": 133, "sum_d": -6339.9282644568, "sum_d2": 615043.7548515502, "mean_diff": -47.668633567344365, "se": 4.2212400411709705, "incumbent_mean": 146.55140363646615, "candidate_mean": 98.8827700691218, "retail_diff": -23.567060090997746, "arb_diff": -24.101573476406767 }, "4": { "n": 116, "sum_d": -7370.7586843399, "sum_d2": 838027.4114357055, "mean_diff": -63.54102314086121, "se": 5.264245777537733, "incumbent_mean": 120.02204692465259, "candidate_mean": 56.48102378379138, "retail_diff": -16.94393129593017, "arb_diff": -46.59709184493276 } }, "successful_pairs_only": { "overall": { "n": 497, "sum_d": -17187.3936096338, "sum_d2": 1679389.141725968, "mean_diff": -34.58228090469578, "se": 2.0979629359920127, "incumbent_mean": 95.78644213205513, "candidate_mean": 61.20416122735936, "retail_diff": -13.911739800740644, "arb_diff": -20.670541103968812 }, "by_spread": { "1": { "n": 130, "sum_d": -573.4095022814, "sum_d2": 85042.7985625715, "mean_diff": -4.410842325241538, "se": 2.218176923658692, "incumbent_mean": 38.538409508369234, "candidate_mean": 34.127567183127695, "retail_diff": 0.2643593750607703, "arb_diff": -4.675201700281539 }, "2": { "n": 120, "sum_d": -3034.5576480661, "sum_d2": 151609.70196744482, "mean_diff": -25.287980400550833, "se": 2.2897877600142116, "incumbent_mean": 77.8558643979175, "candidate_mean": 52.56788399736667, "retail_diff": -16.186723266488332, "arb_diff": -9.101257134088334 }, "3": { "n": 132, "sum_d": -6303.6226427857, "sum_d2": 613725.6566866252, "mean_diff": -47.75471699080076, "se": 4.252456674980703, "incumbent_mean": 147.39383805961742, "candidate_mean": 99.63912106881666, "retail_diff": -23.50028391310379, "arb_diff": -24.254433077749997 }, "4": { "n": 115, "sum_d": -7275.8038165006, "sum_d2": 829010.9845093265, "mean_diff": -63.26785927391826, "se": 5.3030685858101965, "incumbent_mean": 119.97546219064174, "candidate_mean": 56.70760291672348, "retail_diff": -16.557018532327824, "arb_diff": -46.710840741585216 } } } }, "failures": [ { "phase": "tuning", "seed": 320109, "strategy": "h100", "error": "orderbook_pm_challenge.market.OrderBookError: Insufficient free cash for order collateral" }, { "phase": "tuning", "seed": 320151, "strategy": "h50", "error": "orderbook_pm_challenge.market.OrderBookError: Insufficient free cash for order collateral" }, { "phase": "tuning", "seed": 320151, "strategy": "h100", "error": "orderbook_pm_challenge.market.OrderBookError: Insufficient free cash for order collateral" }, { "phase": "tuning", "seed": 320154, "strategy": "h100", "error": "orderbook_pm_challenge.market.OrderBookError: Insufficient free cash for order collateral" }, { "phase": "holdout", "seed": 430009, "strategy": "h50safe", "error": "orderbook_pm_challenge.market.OrderBookError: Insufficient free cash for order collateral" }, { "phase": "holdout", "seed": 430108, "strategy": "h50safe", "error": "orderbook_pm_challenge.market.OrderBookError: Insufficient free cash for order collateral" }, { "phase": "holdout", "seed": 430216, "strategy": "h50safe", "error": "orderbook_pm_challenge.market.OrderBookError: Insufficient free cash for order collateral" } ], "check_values": [ 17.1160017503, 20.0171113118, 2.3903113616, 241.4847522404 ], "python_version": "3.12.3", "seconds_per_simulation": 0.3109347607898219, "validation": { "simulator_commit": "b244530ee2a6247b5f930830eeca9dd950da0e7e", "default_config": true, "default_parameter_variance": true, "configured_steps_per_simulation": 2000, "serial": true, "sandbox": false, "off_equality": { "seeds": { "start": 310000, "end_inclusive": 310009 }, "steps_each": 2000, "comparison": "Two independent instances see the same StepState; N actions drive the unmodified engine. Off overlay actions equal N at every callback." }, "synthetic_checks": [ "buy and sell trigger", "three-step window boundary and expiry", "hold and linear recovery", "retained order shrink", "retrigger reset", "conservative collateral cap: 38 synthetic market states" ], "holdout_selected_before_running": true, "tuning_observations": 800, "holdout_observations": 1000, "holdout_successful_runs": 997, "no_duplicate_seed_strategy_pairs_tuning_and_holdout": true, "equality_check_simulations": 20, "equality_check_details": "Original and collateral-capped overlays each checked switched off on the ten specified check seeds, with N as an independent action comparator.", "no_finalist_seeds_used": true, "timing_includes": "Engine.run wall time, including 20 reference/equality runs; excludes downloading and assembly." } }