MERCURY / EXECUTION STUDY

Execution policy comparison

Five execution policies on shared synthetic market streams. Cost includes deadline closeout, explicit fees and a penalty for residual inventory.

Market regime

128 evaluation seeds per regime · 2,000 shares · 60 virtual seconds · stress run · maker fee 0 · taker fee 0 ticks/share

Baseline

Policy	Mean	SE	P95	CVaR95	Fill rate	Δ vs TWAP	Paired SE
adaptive	1.089	0.102	3.265	3.845	96.006%	-0.131	0.025
aggressive	2.500	0.000	2.500	2.500	100.000%	1.280	0.103
passive	3.910	0.428	10.000	10.000	58.488%	2.689	0.455
pov	1.017	0.104	2.895	4.020	99.024%	-0.203	0.050
twap	1.220	0.103	2.967	4.143	96.410%	0.000	0.000

95% bootstrap intervals

Percentile intervals conditional on these observed runs. Paired differences resample per-seed policy costs against TWAP; negative differences mean lower cost. Small holdouts may miss rare tails. Intervals do not account for model error or regime selection.

Policy	CVaR95, ticks/share	Mean difference vs TWAP, ticks/share
adaptive	[3.003, 4.800] (512 draws)	[-0.184, -0.088] (512 draws)
aggressive	[2.500, 2.500] (512 draws)	[1.077, 1.480] (512 draws)
passive	[10.000, 10.000] (512 draws)	[1.806, 3.559] (512 draws)
pov	[2.935, 5.187] (512 draws)	[-0.304, -0.106] (512 draws)
twap	[3.059, 5.451] (512 draws)	[0.000, 0.000] (512 draws)

Selected

Policy	Mean	SE	P95	CVaR95	Fill rate	Δ vs TWAP	Paired SE
adaptive	6.241	0.345	13.023	16.929	99.684%	0.517	0.088
aggressive	2.191	0.000	2.191	2.191	100.000%	-3.533	0.301
passive	-0.988	0.010	-1.000	-0.717	100.000%	-6.712	0.301
pov	3.136	0.244	7.623	8.586	100.000%	-2.588	0.315
twap	5.724	0.301	12.053	15.022	100.000%	0.000	0.000

95% bootstrap intervals

Policy	CVaR95, ticks/share	Mean difference vs TWAP, ticks/share
adaptive	[13.200, 19.976] (512 draws)	[0.322, 0.696] (512 draws)
aggressive	[2.191, 2.191] (512 draws)	[-4.138, -3.014] (512 draws)
passive	[-1.000, -0.254] (512 draws)	[-7.378, -6.173] (512 draws)
pov	[7.301, 9.856] (512 draws)	[-3.165, -1.881] (512 draws)
twap	[12.052, 17.221] (512 draws)	[0.000, 0.000] (512 draws)

Neighbor Low Market

Policy	Mean	SE	P95	CVaR95	Fill rate	Δ vs TWAP	Paired SE
adaptive	5.364	0.350	12.622	14.495	100.000%	0.259	0.058
aggressive	2.191	0.000	2.191	2.191	100.000%	-2.914	0.328
passive	-0.942	0.048	-1.000	0.593	100.000%	-6.047	0.329
pov	2.606	0.291	6.571	7.793	100.000%	-2.499	0.255
twap	5.105	0.328	12.402	13.327	100.000%	0.000	0.000

95% bootstrap intervals

Policy	CVaR95, ticks/share	Mean difference vs TWAP, ticks/share
adaptive	[12.291, 15.971] (512 draws)	[0.140, 0.371] (512 draws)
aggressive	[2.191, 2.191] (512 draws)	[-3.568, -2.251] (512 draws)
passive	[-0.782, 2.362] (512 draws)	[-6.671, -5.477] (512 draws)
pov	[6.713, 8.876] (512 draws)	[-2.980, -1.985] (512 draws)
twap	[11.756, 14.646] (512 draws)	[0.000, 0.000] (512 draws)

Neighbor High Market

Policy	Mean	SE	P95	CVaR95	Fill rate	Δ vs TWAP	Paired SE
adaptive	5.757	0.329	13.201	15.592	99.577%	0.260	0.108
aggressive	2.191	0.000	2.191	2.191	100.000%	-3.306	0.288
passive	-1.021	0.012	-1.000	-1.000	100.000%	-6.518	0.288
pov	3.121	0.253	8.012	8.962	100.000%	-2.376	0.321
twap	5.497	0.288	12.662	14.404	99.987%	0.000	0.000

95% bootstrap intervals

Policy	CVaR95, ticks/share	Mean difference vs TWAP, ticks/share
adaptive	[12.931, 17.754] (512 draws)	[0.058, 0.482] (512 draws)
aggressive	[2.191, 2.191] (512 draws)	[-3.831, -2.799] (512 draws)
passive	[-1.000, -1.000] (512 draws)	[-7.112, -5.929] (512 draws)
pov	[7.658, 9.903] (512 draws)	[-3.062, -1.799] (512 draws)
twap	[12.198, 15.973] (512 draws)	[0.000, 0.000] (512 draws)

Method and interpretation

Mean error bars show one sample standard error. Δ vs TWAP uses paired seed differences; a negative number means lower cost. Tail estimates from 128 seeds are noisy: the worst 5% spans only 6.4 observations, with fractional boundary weighting.

Search uses 32 candidates and 32 training seeds per candidate. Evaluation starts at seed 1000007; search and evaluation seeds are disjoint for stress runs. Neighbors change only market-order intensity by ±10% within the search bounds.

Fixed submission delay: 0 ms. At most one submission is in flight; cancels are immediate. External events precede equal-time arrivals. At the deadline, pending submissions expire and the common immediate closeout runs. No feed delay, cancellation delay or network jitter is modeled.

External event times, prices, quantities and cancellation IDs are frozen before execution. Policy orders affect replayed queues and fills, but do not change external arrival rates. Fees are explicit per-filled-share inputs, not calibrated exchange schedules. No empirical calibration, price prediction, live trading or GPU performance is claimed.

Raw runs · Summary · First-seed decisions · Individual fills · First-seed market events · Configuration

Uncertainty exports: CVaR95 intervals · Paired mean intervals

Replay files: baseline tape · neighbor high market tape · neighbor low market tape · selected tape

Full configuration

{
  "format_version": 2,
  "engine_version": "0.3.0",
  "mode": "stress",
  "seed": 7,
  "first_evaluation_seed": 1000007,
  "runs": 128,
  "workers": 4,
  "candidates": 32,
  "training_runs": 32,
  "execution": {
    "side": "buy",
    "quantity": 2000,
    "horizon_ns": 60000000000,
    "interval_ns": 1000000000,
    "participation": 0.15,
    "maker_fee_ticks": 0,
    "taker_fee_ticks": 0,
    "residual_penalty_ticks": 10
  },
  "regimes": {
    "baseline": {
      "limit_rate": 8,
      "cancel_rate": 6,
      "market_rate": 2.5,
      "flow_imbalance": 0,
      "initial_level_size": 500,
      "mean_market_size": 120,
      "cancel_imbalance_sensitivity": 1.2,
      "market_imbalance_sensitivity": -0.8,
      "limit_spread_sensitivity": 0.15,
      "impact_ticks_per_unit": 0.05,
      "hawkes_alpha": 0,
      "hawkes_beta": 5,
      "mean_limit_size": 200,
      "initial_depth_levels": 10,
      "initial_half_spread": 1,
      "tick_size": 0.01
    },
    "selected": {
      "limit_rate": 15.15774177718482,
      "cancel_rate": 15.67714527034423,
      "market_rate": 9.40181005860155,
      "flow_imbalance": -0.23508463896202794,
      "initial_level_size": 603,
      "mean_market_size": 341.97570879044355,
      "cancel_imbalance_sensitivity": 1.2,
      "market_imbalance_sensitivity": -0.8,
      "limit_spread_sensitivity": 0.15,
      "impact_ticks_per_unit": 0.05,
      "hawkes_alpha": 0,
      "hawkes_beta": 5,
      "mean_limit_size": 200,
      "initial_depth_levels": 10,
      "initial_half_spread": 1,
      "tick_size": 0.01
    },
    "neighbor_low_market": {
      "limit_rate": 15.15774177718482,
      "cancel_rate": 15.67714527034423,
      "market_rate": 8.461629052741396,
      "flow_imbalance": -0.23508463896202794,
      "initial_level_size": 603,
      "mean_market_size": 341.97570879044355,
      "cancel_imbalance_sensitivity": 1.2,
      "market_imbalance_sensitivity": -0.8,
      "limit_spread_sensitivity": 0.15,
      "impact_ticks_per_unit": 0.05,
      "hawkes_alpha": 0,
      "hawkes_beta": 5,
      "mean_limit_size": 200,
      "initial_depth_levels": 10,
      "initial_half_spread": 1,
      "tick_size": 0.01
    },
    "neighbor_high_market": {
      "limit_rate": 15.15774177718482,
      "cancel_rate": 15.67714527034423,
      "market_rate": 10.341991064461705,
      "flow_imbalance": -0.23508463896202794,
      "initial_level_size": 603,
      "mean_market_size": 341.97570879044355,
      "cancel_imbalance_sensitivity": 1.2,
      "market_imbalance_sensitivity": -0.8,
      "limit_spread_sensitivity": 0.15,
      "impact_ticks_per_unit": 0.05,
      "hawkes_alpha": 0,
      "hawkes_beta": 5,
      "mean_limit_size": 200,
      "initial_depth_levels": 10,
      "initial_half_spread": 1,
      "tick_size": 0.01
    }
  }
}