The Carry Trade (FX and Commodities) — a real premium with a documented crash tail and a post-2008 fifty-percent haircut

Borrow the low-rate currency, buy the high-rate one (FX carry), or hold futures in structural backwardation (commodity carry): both collect a documented premium, and both carry a well-quantified left tail. Pre-crisis FX carry Sharpe ratios (0.91-1.48) rival any factor in this wiki, but the same episode that built the reputation — 2008 — is also the textbook crash. The premium has since roughly halved.

FX Carry: Interest Rate Differential Strategy, Historical Sharpe and Returns

An unhedged FX carry portfolio (1990-2007) delivered 4.78% annualized excess return at 5.07% volatility, Sharpe = 0.91 (Jurek 2007, https://www.nber.org/system/files/working_papers/w14054/w14054.pdf). Restricting to 1999-2007, equal-weighted and spread-weighted carry portfolios reached Sharpe 1.26 and 1.48 respectively — the most favorable pre-crisis window (same source). Over the much longer 1900-2012 sample, Quantpedia reports only 2.4% annualized excess return at Sharpe 0.26 — averaging in crisis episodes drags the number far below the pre-2008 figures. A practitioner benchmark, the Deutsche Bank Currency Carry USD Index, shows 7.27% indicative annual return over 1989-2009 (Quantpedia; not peer-reviewed).

Skewness is the tell

Brunnermeier et al. report -0.41 weekly and -0.88 quarterly skewness for carry portfolios, with annualized Sharpe of only 0.44 (weekly) and 0.41 (quarterly) on their own formation methodology (https://www.nber.org/system/files/working_papers/w14473/w14473.pdf). The headline Sharpe ratios above and the negative-skew Sharpe ratios here come from different studies/methodologies and should not be averaged — they document the same phenomenon (return concentrated with a fat left tail) from different angles.

Crash Risk: 2008 and Other Carry Unwind Magnitudes

The 2008 unwind is the textbook case: as the crisis hit, USD/JPY fell sharply and carry positions took large, fast losses. Brunnermeier et al. 2008 document the mechanism: carry trades take substantial losses when VIX rises, because margins increase and funding constraints bind (https://www.nber.org/system/files/working_papers/w14473/w14473.pdf).

Despite this, Jurek (2007) estimates the crash-risk premium in G10 currencies at only 0.20%-0.50% per annum, depending on portfolio weighting and hedging — under 10% of gross unhedged carry returns. If that estimate is right, most of the carry premium is not compensation for crash risk, which makes the risk-reward asymmetry worse, not better, than a pure risk-premium story would suggest. (unverified) A widely repeated narrative — that a trillion-dollar yen position turned the subprime shock into a global crash in 2008 — lacks precise documentation and the causal direction is disputed; treat it as narrative, not evidence.

Commodity Carry (Roll Yield) Evidence and Numbers

An equally-weighted commodity futures portfolio of 35 contracts, 1877-2015, generated 4.8% average annual return at 17.6% volatility (arithmetic mean); the geometric mean over the same full sample is 3.3% (Levine, Ooi, Richardson, Sasseville, "Commodities for the Long Run," NBER WP 22793, via https://www.cxoadvisory.com/commodity-futures/some-commodities-over-the-long-run/). Over the 139-year sample, the carry component (roll yield plus short rates) contributes more to average returns than the excess-of-cash spot-return component: excess spot return averages 2.1% annualized versus 3.9% for net convenience yield (carry). This is the central finding for this page — carry, not spot appreciation, is the return engine in commodities.

Curve shape drives return

backwardation periods (47% of the sample, average duration 4.4 months) delivered 7.9% annualized returns; contango periods delivered only 1.5%. Returns are also strongly state-dependent: +14.1% in high-inflation regimes vs -4.5% in low-inflation ones, +8.9% in economic expansions vs -7.3% in contractions (same source, 1877-2015).

Nearly 50% of cumulative trend-following performance in commodity futures is attributed to roll yield rather than spot trend (academic research cited in commodity futures literature); momentum strategies that explicitly exploit expected roll yield reportedly earn higher risk-adjusted returns than momentum using only nearest-contract prices (multiple sources, not individually named in notes — treat as directionally supported, not precisely quantified).

Cost and Capacity, and Whether the Premium Has Decayed

Carry performance fell 5.4% per year from the pre-GFC to the post-GFC era for a volatility-matched carry portfolio (academic research on post-2008 performance). The mechanical driver: post-crisis interest-rate spreads between high- and low-rate currencies narrowed substantially as rates compressed toward zero everywhere, shrinking the gross differential available to harvest.

This sits inside a broader pattern documented elsewhere in this wiki: McLean-Pontiff find returns to 97 financial characteristics decline about 58% after academic publication (see Out-of-Sample vs Post-Publication Decay: The Two Numbers That Tell You If a Premium Is Real). Carry-specific research reports a comparably severe 50% post-publication decline in carry-trade returns after papers disseminated the strategy to practitioners (2025 research, https://www.sciencedirect.com/science/article/abs/pii/S0927539825000623). In the post-2008 period, unhedged carry lost its crash-risk compensation, and hedged crash-neutral overlays showed no significant premium above unhedged carry — consistent with either the crash-risk premium being arbitraged away or hedging costs consuming it.

Attempts to restore the edge with risk management exist: a managed (risk-adjusted) carry strategy reached Sharpe 1.07 versus 0.76 for a benchmark strategy across a 34-currency sample (1999-2018) — practitioner-attested, not peer-reviewed, and itself a post-publication result, so it should be read as "someone tried to fix the decayed edge," not as proof the fix works out-of-sample. On costs: carry has low turnover relative to short-term reversal strategies, which protects it from the transaction-cost erosion that kills faster strategies (see Transaction Cost Accounting — the arithmetic that separates a real edge from a paper one), but spread compression and increased competition have still eroded margins. (unverified) Claims that carry trade "still works" in 2024-2025 after the August 2024 yen unwind lack long-term statistical validation and may just be mean-reversion rather than a restored premium — too recent to judge.

What does NOT work

Treating pre-crisis Sharpe ratios (0.91-1.48) as the strategy's ongoing character is the main trap: those numbers come from a specific 1990-2007 (or 1999-2007) window that ended in the crash that defines the strategy's tail risk. The 139-year FX-adjacent long sample and the post-GFC 5.4%/year decline both show the true unconditional picture is far weaker. Averaging skewness/Sharpe figures across the Jurek and Brunnermeier et al. methodologies as if they were one estimate is also wrong — they use different formation rules and periods.

Related

- Out-of-Sample vs Post-Publication Decay: The Two Numbers That Tell You If a Premium Is Real — carry's ~50% post-publication decline is a direct data point for this wiki's decay pattern (McLean-Pontiff 58% baseline), and the mechanism (crowding once a paper reveals the trade) is the same one this page documents for carry specifically. - Transaction Cost Accounting — the arithmetic that separates a real edge from a paper one — carry's low turnover is why it survives costs better than reversal-type strategies even as competition compresses its gross spread; the cost-survival question this wiki requires for every technique. - Time-Series Momentum in Futures — a backtest with Sharpe near 1.0 that live CTAs never matched — commodity carry (roll yield) and time-series trend-following overlap in futures markets; nearly half of trend-following's cumulative performance traces to roll yield rather than price trend, so the two techniques share return sources and should not be counted as fully independent edges.