Robustness or Crowding: Experimental Design for Trading Strategy Capacity

Battle Count: 82

This paper is directly and fundamentally relevant to quantitative trading. It addresses the central practical question of how much capital a trading strategy can absorb before its edge disappears (capacity), which is the primary constraint on fund sizing and allocation for quant strategies. The paper provides explicit design rules for measuring capacity through controlled experiments, prices the cost of such experiments in forgone edge, establishes what within-date contrasts can and cannot identify, and benchmarks against the impact models that trading desks already run. The calibration on 13 long-short strategies (momentum, size, value, profitability, investment) directly mirrors the strategies run by quant funds. The finding that contemporaneous assignment is essential (staggered assignment causes variance saturation) and that arms near the operating scale cost almost nothing while still separating the design are immediately actionable for any institution considering capacity measurement.

The theoretical framework is highly complex, involving switchback experimental design, continuous treatment causal inference, interference/exposure mapping, partial identification on finite grids, accumulation kernels, and constrained optimization. Implementing the full design requires: (1) running multiple parallel trading sleeves with randomized deployment scales, (2) maintaining block randomization with fixed hold periods, (3) computing within-date arm contrasts with factor adjustment, (4) deattenuating using transported or estimated kernels, (5) constructing simultaneous confidence bands over a prespecified candidate set, (6) pricing the experiment in forgone edge, and (7) potentially running a separate ramp design for path dependence testing. The calibration requires reconstructing implied borrow rates from option data and screening 125 candidate drivers. However, the core estimation (block average contrast, deattenuation by GL) is straightforward once the design is set up. The main practical barrier is institutional: running many parallel sleeves simultaneously with randomized deployment.

The paper provides extremely detailed simulation parameters (Section G), calibration inputs (Table 2), and design specifications. However, no actual deployment experiment is run. The Fama-French portfolio data is publicly available, but the reconstructed implied borrow rate series (1996-2024) and the specific panel construction are not publicly distributed. The pre-registration protocol (Section F) is outlined but not executed. Monte Carlo exercises are fully specified with replication counts, grids, and algorithms, enabling reproduction of the simulation results given the calibration inputs. The two transported bridges (persistence and scale response) are stated but not testable from the panel alone.