← All articles

Carry on 28 currency pairs: a real premium, an unbearable risk

13 min read · Published 28 Oct 2026

Carry is the oldest premium on the currency market: you buy a currency with a high interest rate against one with a low rate and collect the difference every night. On 28 pairs in 2005–2026, our rule picked the sides of pairs clearly better than chance (placebo p = 0.004), so the premium exists even after costs at a CFD broker. It earned, however, an average of 2.9% a year with a drawdown of 37.6%. After cutting the risk to a level an account can bear, 0.2% a year was left. This chapter is about the fact that a real signal and a profitable strategy are two different things.

Abbreviations: swap is the fee or credit charged every night for holding a currency position; it comes from the interest-rate difference between the two currencies. The markup is the part of that difference the broker keeps. Sharpe is the average annual result divided by its volatility; 0 means no edge. Calmar is the average annual profit divided by the largest drawdown. R is a position’s result in units of risk (−1 R = a loss equal to the distance to the stop). Placebo is the same procedure with randomly assigned signals.

Hypothesis

Theory says carry should not work. If one currency pays 5% a year and another 0%, the first should lose roughly 5% a year in value, and the gain from interest rates will vanish in the exchange rate. In the data, however, something else happens. Eugene Fama (1984, Journal of Monetary Economics) showed that higher-yielding currencies on average do not lose as much as this principle predicts, and often even gain. Since then, the phenomenon has been called the forward premium puzzle.

Later literature explains what this premium pays for. Lustig, Roussanov and Verdelhan (2011, Review of Financial Studies) showed that high-yielding currencies react more strongly to a common, global risk factor. Menkhoff, Sarno, Schmeling and Schrimpf (2012, Journal of Finance) linked carry returns to the volatility of the whole currency market: carry loses precisely when uncertainty rises around the world. Brunnermeier, Nagel and Pedersen (2008, NBER Macroeconomics Annual) described crash risk: the distribution of returns is skewed to the left, because carry positions are closed abruptly and all at once when risk appetite and access to funding fall.

So carry is payment for the risk of a sudden collapse. Our question was: does this payment survive a CFD broker’s costs, and is the risk worth it?

The hypothesis, frozen before the run: a portfolio of the three pair sides with the highest positive net carry (the rate difference minus the broker’s markup), rebuilt every week, achieves a Sharpe of at least 0.40 after all costs and beats a placebo in which carry was randomly assigned to other pairs.

Why not the usual criterion. For our other strategies we require a PF of at least 1.25 and at least 100 trades. Carry, however, earns little and regularly, and loses rarely and a lot. With weekly positions, PF depends practically only on the position’s Sharpe: a Sharpe of 0.40 gives a PF of about 1.15, and a PF of 1.25 would require a Sharpe of about 0.65 on each position separately, more than carry had in its best decades. For a similar reason, we gave up on splitting the data into training and holdout samples. To distinguish a Sharpe from zero (at p < 0.05), with 21 years of data about 0.36 is enough, while with a 7.7-year holdout sample you would need 0.59. That is why the test was a single shot on the whole history, with parameters taken in advance, and with a criterion that checks the tail.

How we tested it

Data. 28 pairs from eight major currencies (EUR, GBP, AUD, NZD, USD, CAD, CHF, JPY), daily candles from 2004. We built the cross rates from the dollar pairs, because the broker’s history of crosses had gaps of 42–136 days in 2006–2009. From 2011, the synthetic crosses matched the real ones with a median difference of no more than 0.025%.

Interest rates. Until mid-2010 we used three-month interbank rates from the OECD database, and from July 2010 the rate difference derived from currency futures prices. The reason for the switch came out in validation before the test: interbank rates systematically understated the difference for the yen, the franc and the euro (by a median of 0.52 percentage points a year on the yen). Futures prices also include the so-called cross-currency basis, the extra price of borrowing dollars, and it was with them that the broker’s current swaps agreed. The measurement gate: the model had to reproduce the broker’s current swaps. The sign matched on 28 pairs out of 28, and the error was at most 0.32 percentage points on 27 pairs and 0.52 on one.

The broker’s markup was derived from the account’s current swaps: from 0.59 to 2.14% a year per side, median 1.01%. We assumed it constant over the whole history, so one of the stress tests was raising it by half.

The rule:

• decision at the close of the last day of the week, execution at the open of the first day of the next;
• candidates: pair sides with positive net carry; variant A takes the three best, variant B only those on the right side of the 100-day moving average of closes;
• a position stays as long as it is in the top three;
• a stop on the daily close, at a distance of 3 × the average daily change in the exchange rate over 20 days; after a stop, the side can return at the next rebuild at the earliest;
• equal risk per position; in the report, 0.5% of capital per position.

Costs: the spread from ticks of the real pair once per trade, a commission of $7 per lot round trip, and swap charged every night according to net carry, triple on Wednesday.

Criteria, frozen before the run, all required to pass: a Sharpe of at least 0.40; placebo p below 0.05 (1,000 draws, with a correction for two variants); a result better than the best passive exposure to any of the 28 pairs; a Calmar of at least 0.33, meaning the worst drawdown is recovered in at most three years of average profit; at least 60% of years in profit, and the best year at most 40% of the profit; a positive Sharpe without the five best weeks and without the best pair; a Sharpe of at least 0.25 with the markup raised by half and with carry lowered by 0.5% a year. We also wrote down a rejection condition in advance: a Sharpe below 0.25, placebo p of at least 0.20, or a Calmar below 0.15.

Result

Both variants met the rejection condition on Calmar, so the verdict was KILL. At the same time, the signal passed the key placebo check: choosing sides by carry beat random choice with p = 0.004, and the reverse ranking, holding the sides most expensive to fund, clearly lost. The premium exists even after costs at a CFD broker. But it cannot be collected safely. One thing in the table may seem strange: a higher markup slightly raised the Sharpe. Cost does not help, but with a higher markup fewer sides have positive net carry, so in the zero-rate years the portfolio stood aside more often and its composition changed.

While writing this chapter, we also calculated, on the saved test data, where variant A’s result came from. In closed positions, swap brought in about +61% of capital, exchange-rate moves took away about 9.5%, and spread and commissions about 2%. So the high-yielding currencies did lose value, but only by one sixth of what the interest brought in. This is Fama’s puzzle, measured on our data. The problem lies in how this premium was distributed over time:

Why it doesn’t work

The premium is paid for with the tail. The 37.6% drawdown lasted from July 2007 to January 2009. In August 2007 a single week cost 9.3% of capital, and in October 2008 another one 8.1%. Later came smaller but painful episodes: the week of the Swiss central bank’s decision in January 2015 (−5.1%) and summer 2024, when the market abruptly closed yen-funded positions: two consecutive weeks brought −3.2% and −5.2%. A stop on the daily close does not protect against a violent move within a single session: the worst single position lost 6.8 R. Correction of 28 Oct 2026: this used to say “against a gap”. That position did not jump past the stop at the open, as it started the day still in profit; the whole loss came from a move within one session.

A Calmar of 0.08 means 13 years of recovery. With an average profit of 2.9% a year, a 37.6% drawdown is almost thirteen years of profit. Our rejection threshold (0.15) allowed seven years, and the strategy did not meet even that. A simulation with a random ordering of periods showed a drawdown of 74% at the 95th percentile. To fit it within 6%, risk would have to be cut to 0.04% per position, and then 0.24% a year is left.

It is one bet, not a portfolio. In 71.5% of position-weeks the funding currency was the yen or the franc. Most often the strategy held long New Zealand and Australian dollars against the yen, short the euro against the New Zealand dollar, and long the New Zealand dollar against the franc. Three positions meant in practice the same bet three times: risky currencies against safe ones. In a panic, all three lose at once.

The broker takes a large part of the premium. A markup of 0.6 to 2.1% a year per side was small compared with the rate differences of 2005–2008, when the three best sides yielded about 5–6% a year net. In the zero-rate years almost nothing was left: in 2021, on average only 4 sides out of 56 had any positive net carry at all, and the three best yielded 0.3% a year.

The trend filter moved the losses, but did not reduce them. Variant B limited the 2008 loss to −2.4% instead of −27.2%, but gave it back in other years (2011: −12.6%, 2017: −9.5%, 2020: −13.2%). The Calmar stayed the same.

The third criterion was weak, but it was not the deciding one. The strategy did not beat the best passive exposure (Sharpe 0.34). We later calculated, however, that with zero drift in exchange rates, the best of 28 pairs picked after the fact typically has a Sharpe of about 0.38, so such a benchmark mostly measures chance. The Calmar decided the verdict.

Second attempt: a vehicle without a markup. The only permissible way back to the topic was to change the trading vehicle, not the parameters. So we recalculated the same rule on currency futures, where the rate difference is built into the price and there is no broker markup. In 2010–2026, variant A gave a Sharpe of 0.28, against 0.30 at the broker over the same period. The markup was gone, but a cross on futures means two trades instead of one, so the cost doubled. The 0.30 threshold was not reached, and we closed the topic of carry after two attempts. The same pattern, a real premium that after CFD costs yields almost zero, repeated with economic momentum in currencies (chapter 39); we describe it as a whole in chapter 43.

The lesson for you

A placebo test answers the question of whether a signal is real. It does not answer the question of whether you can make money on it with a risk your account can bear. Carry passed the first test and failed the second.

How to evaluate a carry-type strategy yourself:

1. Judge by the worst period, not by the average. Calculate how many years of average profit the largest drawdown takes away. If it is more than a few, the strategy is a bet that a crisis will not come.
2. Calculate your broker’s markup. Take the annual long and short swaps of the same pair, in percent. The markup per side is minus their average. Example: a long position earns +3%, a short costs −5%, so the broker keeps 1% a year per side. Compare this with the rate difference.
3. Count the funding currency. If most positions are funded with the same currency, you have one bet, even across several pairs.
4. Don’t rely on the stop. A stop checked on the daily close, and in practice any stop during a weekend gap or a central bank decision, can cost several times more than the planned 1 R.
5. Check compliance with the challenge rules. Carry holds positions over weekends and for weeks. One bad week from 2007 (−9.3% at 0.5% risk per position) would eat almost the entire typical drawdown limit of a challenge.

In your trading journal, record swap separately from the result of the price move. After a few months you will see whether you are making money on direction or only on interest, and how much of it the broker takes. The “Luck or edge?” calculator will help you judge whether the result stands out from chance, but only the worst period in the history will tell you about tail risk.

Educational content, not investment advice. Past results do not guarantee future results. We describe the result of our reconstruction of a single, strictly defined carry strategy, on our data and with our broker’s costs — not an assessment of every way of trading interest-rate differences.