Adaptive execution has to earn its complexity.
Reinforcement learning, order allocation, and the practical importance of realistic fills.
| Section | Sources cited | Equations | Code listings |
|---|---|---|---|
| 1. Decompose implementation shortfall into spread, impact, and adverse selection | 1, 2 | 1 | 0 |
| 2. Enforce external risk envelopes and deterministic fallbacks | 1, 2 | 0 | 1 |
01 /Decompose implementation shortfall into spread, impact, and adverse selection
The path from a portfolio target to an executed position contains its own optimization problem. Order size, book depth, bid-ask spread, queue priority, and venue rules can alter or reverse the expected value of a strategy. Execution design must be evaluated alongside signal discovery under a unified cost ledger.
Recent reinforcement learning research studies how agents allocate between market and limit orders in simulated limit order books [1]. While these experiments clarify optimal state-dependent routing under stylized fill intensities, simulation results do not establish live outperformance unless passive queue priority and post-fill adverse selection are modeled explicitly.
A rigorous engagement begins with deterministic baselines such as TWAP and closed-form Almgren–Chriss trajectories [1, 2]. Every candidate policy is evaluated on identical parent-order mandates over horizon , measuring expected implementation shortfall and variance alongside terminal unexecuted inventory penalties.
Child fills q_t incur temporary impact and spread costs, while any leftover inventory q_T at the deadline T is forced through a penalty sweep ξ.
02 /Enforce external risk envelopes and deterministic fallbacks
We compare implementation costs, passive fill probabilities, order rejection behavior, and participation caps across calm and stressed liquidity regimes [1, 2]. Any adaptive policy requires evaluation on unseen volatility episodes where historical spread relationships widen.
In production, the execution service must also reconcile partial fills asynchronously, handle interrupted socket connections, and enforce hard exposure limits outside the learned policy. Intelligence is useful only when the complete order state machine remains observable and bounded.
- Benchmark adaptive routing policies against TWAP and Almgren–Chriss schedules on identical parent orders.
- Penalize unexecuted deadline inventory explicitly so policies cannot game cost metrics by withholding trades.
- Maintain external participation caps and kill-switch state machines independent of the policy network.
# Illustrative only; compares front-loaded risk-averse trajectory against uniform TWAP.
import numpy as np
Q0, T = 10_000.0, 10
eta, sigma, lam = 2.5e-6, 0.012, 1e-4
kappa = np.arccosh(1.0 + 0.5 * (lam * sigma**2) / eta)
t_grid = np.arange(T + 1, dtype=float)
ac_inventory = Q0 * np.sinh(kappa * (T - t_grid)) / np.sinh(kappa * T)
ac_trades = -np.diff(ac_inventory)
twap_trades = np.full(T, Q0 / T)
assert np.isclose(np.sum(ac_trades), Q0)
assert np.all(ac_trades >= 0.0) and np.isclose(ac_inventory[-1], 0.0)
assert ac_trades[0] > twap_trades[0]
print(f"first_slice_ac={ac_trades[0]:.1f}, first_slice_twap={twap_trades[0]:.1f}")Research is complete only when the assumptions, experiment, and limitations can be reconstructed.
This perspective is research commentary. It does not constitute investment advice or evidence of live strategy performance. Experimental methods require project-specific testing. Equations and code illustrate research formulations, not measured live returns.