FX Carry (Hardened v2)

The story in one line

Run exactly as pre-registered over a full 15-year sample, hardened cross-sectional FX carry loses money: final $3,392 (−32%), CAGR −2.5%, profit factor 0.844, Sharpe −0.29, max drawdown −44.5%. It passes 3 of 11 gates — and all three are trivial (≥100 weeks; low concentration, which here only means it loses broadly). All eight substantive gates fail.

The decisive result is the carry-label placebo: the real strategy (PF 0.844) is beaten by 33% of random sign-permutations and sits below the placebo 95th percentile (0.955). Over 2011–2026, ranking currencies by interest rate was worse than random. The high-yielders depreciated more than they paid in carry — the forward-premium puzzle did not pay on G10 majors in this era.

Per the frozen multi-testing rule, v2 failed → no v3. Carry is retired — the 6th archetype to fail honestly.

Gate scorecard — 3 / 11

#	Gate	Result	Pass
1	≥ 100 weekly rebalances	796	✅
2	PF ≥ 1.20 net	0.844	❌
3	Annualized Sharpe ≥ 0.6	−0.29	❌
4	Max DD ≤ 12%	−44.5%	❌
5	Positive in ≥ 60% of years	38%	❌
6	Block-bootstrap 95% LB Sharpe > 0	−0.83	❌
7	Carry-label placebo: real PF > p95	0.844 < 0.955	❌
8	2× cost stress: PF > 1.0	0.839	❌
9	Deflated Sharpe positive	DSR 0.012	❌
10	No currency > 40% of P/L	max 33% (CHF)	✅
11	No calendar year > 40% of P/L	max 14%	✅

Gates evaluated on the base pre-registered config (overlay + 75 bp financing drag). No configuration was selected post-hoc to pass.

Why it fails — root cause

(a) The carry edge in G10 is real but tiny post-GFC, and FX drift dwarfs it. Decomposition: cumulative carry +$848, FX price −$2,261, costs −$195. The interest differential earned a slow trickle because 2011–2021 was a near-universal ZIRP era — there was almost no carry to harvest — while the FX leg lost 4× what carry earned. This matches the well-documented “lost decade” for G10 carry.

(b) The signal underperforms random. 33% of random currency-sign assignments beat the real rate-ranked strategy. Whatever P/L exists is not attributable to the carry signal.

(c) The crashes are real and concentrated in the FX leg. CHF −$961 (short-CHF blew up at the Jan-2015 SNB de-peg — USDCHF gapped ~−15% intraday); CAD −$626 and AUD −$591 (long high-yielders through the 2014–15 USD super-cycle). These are the canonical carry-crash mechanism.

(d) The risk overlay made it worse. Raw (no overlay) finishes $5,996 (PF 1.068); pre-registered (overlay + drag) finishes $3,392 (PF 0.844). The VIX de-gross and 12-month drawdown circuit are reactive — they cut exposure after drawdowns begin and stay de-grossed through recoveries, locking in losses. A reactive overlay cannot rescue a signal with no in-sample edge — it can only bleed it. This transferable lesson held in every test, including the 2008 GFC stress appendix (−17.0% with overlay vs −17.5% without).

Reconciling v1 (PF 1.17) with v2 (PF 0.844)

Not a contradiction — sample selection. 2023–2026 was a rare confluence: the largest G10 differentials in 15 years and no grinding carry crash. Measure a strategy on its best regime and it looks deployable; measure it across regimes and the truth surfaces. The two runs also used independent rate sources and v2 used real USDCHF — a genuine independent validation that overturns v1.

Honest caveats

Hand-coded policy rates (small errors can’t flip a placebo-below-random result), 2025–26 rates are a partial reconstruction affecting only the last ~10% of the sample, daily resolution (captures the 2015 CHF gap; intraday was worse). A broad EM carry basket — where the literature’s premia actually live — is a different archetype and instrument set, out of scope for a $5k retail MT5 account.

Verdict: FAIL (PF 0.844). Do not deploy. Not close. Carry is retired as a standalone deployable archetype. What it taught us — that reactive risk overlays destroyed value in every test — is kept as a project-wide lesson.