{ "key": "f-microprice-shift", "mech": "microprice shift", "approach": "StepState does not expose rival sizes, so imbalance is rebuilt from the simulator's rule that every rival level holds quote_notional/price shares: top sizes are proportional to 1/bid and 1/ask, imb=(a-b)/(a+b) (quote_notional cancels). imb>+t moves N's new asks one tick up, imb<-t moves new bids one tick down; N is shown its shifted orders at their original prices so it does not chase them. This proxy is always >=0, so only asks shift, at low prices or wide spreads. Every t tried lost about 30 of edge (retail falls, arbitrage barely changes).", "tried": [ "t=0.05: tuning -42.50 (se 5.60)", "t=0.1: tuning -37.34 (se 5.49)", "t=0.2: tuning -30.35 (se 5.00), frozen; holdout -30.88 (se 3.09)" ], "next": "Drop this mechanism: raising N's ask gives up retail edge and saves little arbitrage loss. A real size-imbalance signal would need the rival's remaining sizes, which StepState does not provide. A partial-depletion estimate built from rival best-quote moves could be screened as a separate idea.", "spreads_acted": [ 1, 2, 3, 4 ], "off_equals_n": { "seeds": "310000..310009", "equal_every_step": true, "steps_per_seed": 2000 }, "code": "# Overlay f-microprice-shift on incumbent N.\n# The rival's resting sizes are not in StepState; the simulator sizes every\n# rival level at quote_notional/price, so the rival's top-of-book sizes are\n# rebuilt as proportional to 1/bid and 1/ask (quote_notional cancels).\n# imb = (Sb - Sa)/(Sb + Sa). imb > +T: our new asks go one tick up;\n# imb < -T: our new bids go one tick down. N sees its own resting orders at\n# the prices it asked for, so it does not chase the shifted ones.\nimport dataclasses\n\nMP_T = 0.2\nMP_ON = True\nMP_SPREADS = (1, 2, 3, 4)\n\n\nclass Strategy(_N):\n def __init__(self, *a, **k):\n super().__init__(*a, **k)\n self._mp_shift = {}\n self._mp_k = 0\n\n def on_step(self, state):\n if not MP_ON:\n return super().on_step(state)\n if self._mp_shift and state.own_orders:\n live = set()\n views = []\n for o in state.own_orders:\n d = self._mp_shift.get(o.order_id)\n if d:\n live.add(o.order_id)\n o = dataclasses.replace(o, price_ticks=o.price_ticks - d)\n views.append(o)\n self._mp_shift = {i: self._mp_shift[i] for i in live}\n state = dataclasses.replace(state, own_orders=tuple(views))\n acts = super().on_step(state)\n if getattr(self, \"s\", None) not in MP_SPREADS:\n return acts\n b = state.competitor_best_bid_ticks\n a = state.competitor_best_ask_ticks\n if b is None or a is None or b <= 0 or a <= 0:\n return acts\n sb = 1.0 / b\n sa = 1.0 / a\n imb = (sb - sa) / (sb + sa)\n if imb > MP_T:\n side, d = Side.SELL, 1\n elif imb < -MP_T:\n side, d = Side.BUY, -1\n else:\n return acts\n out = []\n for x in acts:\n if isinstance(x, PlaceOrder) and x.side is side and 1 <= x.price_ticks + d <= 99:\n cid = x.client_order_id\n if cid is None:\n self._mp_k += 1\n cid = \"mp-%d\" % self._mp_k\n x = PlaceOrder(side=x.side, price_ticks=x.price_ticks + d,\n quantity=x.quantity, client_order_id=cid)\n self._mp_shift[cid] = d\n out.append(x)\n return out\n", "code_sha256": "032ddde692ad498208dda62a07a9a59819fe2fb6b8bfb6ca870bc108a82c6d39", "assembled_sha256": "81ca9d5bb8ff5eccaf768e19ac4281a6d2a8928f3050f05be5694493768987d6", "tuning": { "seeds": "320000..320199", "selected": { "t": 0.2 }, "variants": [ { "t": 0.05, "overall": { "n": 200, "sum_d": -8499.510067, "sum_d2": 1607862.340855, "mean_diff": -42.49755, "se": 5.596692, "incumbent_mean": 98.470718, "candidate_mean": 55.973168, "retail_diff": -45.419535, "arb_diff": 2.921985 }, "by_spread_mean_diff": { "1": -14.196971, "2": -23.702312, "3": -54.677637, "4": -76.538833 } }, { "t": 0.1, "overall": { "n": 200, "sum_d": -7468.670297, "sum_d2": 1476422.190924, "mean_diff": -37.343351, "se": 5.485286, "incumbent_mean": 98.470718, "candidate_mean": 61.127367, "retail_diff": -39.715137, "arb_diff": 2.371786 }, "by_spread_mean_diff": { "1": -14.8877, "2": -20.426636, "3": -48.081594, "4": -65.289174 } }, { "t": 0.2, "overall": { "n": 200, "sum_d": -6070.187145, "sum_d2": 1177699.587048, "mean_diff": -30.350936, "se": 4.996139, "incumbent_mean": 98.470718, "candidate_mean": 68.119783, "retail_diff": -32.093555, "arb_diff": 1.742619 }, "by_spread_mean_diff": { "1": -10.137253, "2": -16.410898, "3": -40.333672, "4": -53.976217 } } ] }, "paired": { "seeds": "430000..430499", "n": 500, "complete": true, "missing_seeds": [], "overall": { "n": 500, "sum_d": -15437.669677, "sum_d2": 2853642.679767, "mean_diff": -30.875339, "se": 3.086592, "incumbent_mean": 95.542015, "candidate_mean": 64.666676, "retail_diff": -32.743386, "arb_diff": 1.868046 }, "by_spread": { "1": { "n": 131, "sum_d": -1698.026052, "sum_d2": 406750.158341, "mean_diff": -12.962031, "se": 4.753095, "incumbent_mean": 38.277937, "candidate_mean": 25.315906, "retail_diff": -16.660331, "arb_diff": 3.6983 }, "2": { "n": 120, "sum_d": -3451.62474, "sum_d2": 434674.41935, "mean_diff": -28.76354, "se": 4.846332, "incumbent_mean": 77.855864, "candidate_mean": 49.092325, "retail_diff": -29.895527, "arb_diff": 1.131987 }, "3": { "n": 133, "sum_d": -5642.933895, "sum_d2": 1136871.486736, "mean_diff": -42.428074, "se": 7.149785, "incumbent_mean": 146.551404, "candidate_mean": 104.123329, "retail_diff": -45.863275, "arb_diff": 3.4352 }, "4": { "n": 116, "sum_d": -4645.084989, "sum_d2": 875346.61534, "mean_diff": -40.043836, "se": 7.188507, "incumbent_mean": 120.022047, "candidate_mean": 79.978211, "retail_diff": -38.809576, "arb_diff": -1.234261 } } }, "failures": [], "check_values": { "310000": 17.116002, "310001": 20.017111, "310002": 2.390311, "310003": 241.484752 }, "python_version": "3.14.4", "seconds_per_simulation": 1.39, "notes": "Simulator at commit b244530 unmodified; serial engine runs, no --workers/--sandbox; up to 4 independent single-process runs at once on disjoint (seed,strategy) sets; each (seed,strategy) run exactly once. Screen, not a promotion." }