Hypotheses: two literature-backed overlays should cut drawdowns:
- Volatility targeting (Moreira & Muir, 2017): scale exposure down when recent portfolio volatility exceeds a fixed annual budget.
- Turbulence de-risking (Kritzman & Li, 2010): cut exposure when the cross-asset Mahalanobis distance spikes above its trailing 80th percentile.
I originally ran both against a private strategy of mine (its rules aren't published; any long-only allocation strategy would do — the lesson is about the method).
The trap this case study exists to demonstrate: any rule that lowers average market exposure will show a smaller drawdown — even a coin flip would. The honest control is not the base strategy; it is a matched-exposure twin: the same average exposure held constantly, with zero timing intelligence. If the overlay's timing adds anything, it must beat its twin.
What the pipeline said (two independent decade-long halves; Calmar = CAGR / |maxDD|):
| candidate | first half, Calmar vs twin | second half, Calmar vs twin | verdict |
|---|---|---|---|
| vol-targeting | 0.78 vs 0.84 | 1.06 vs 1.12 | loses to its own twin |
| turbulence de-risk | 0.64 vs 0.84 | 0.82 vs 1.13 | loses badly; made two separate crisis windows worse |
Both overlays' timing subtracted value. Their entire apparent "protection" was the de-grossing — which a constant, brainless scale-down delivered better and cheaper.
Verdict: REJECTED, both. Lesson: never benchmark a safety overlay against the base
strategy. Benchmark it against doing the same amount of nothing — that comparison is built
into this framework as controls.matched_exposure_control.