The Carry Trade Hides in Plain Sight

The standard account of the August 2024 unwind begins by saying that volatility had been unusually low, which made leverage cheap and positions large. On the funding currency most closely associated with the episode, that account is not supported. In the ten weeks before 5 August 2024, sixty-day realised volatility in USD/JPY averaged 8.7 per cent, the 48th percentile of its distribution since 2005. On a twenty-day window it was the 53rd percentile; on a hundred-and-twenty-day window, the 43rd. It was, by every measure tried here, ordinary.

Figure 1 reports both halves: the unwind itself, which was violent, and the volatility measure that preceded it, which was unremarkable. The second half is a negative result and is kept because it bears directly on whether the mechanism this article proposes can be used as a warning system.

Section 01Fixing what the word means

Carry is used loosely enough to mean almost nothing, so this article attaches it to one observable object: the yen-funded position whose unwind the Bank for International Settlements documented in Bulletin 90. Borrow in a low-rate currency, hold higher-yielding assets, and the position earns a rate differential while remaining exposed to the funding currency appreciating. USD/JPY is therefore a price proxy for the position: a sharp fall, meaning yen strength, is the loss signature.

It is a proxy and nothing more. No public daily series measures aggregate carry positioning, and the exchange rate moves for many reasons besides the closing of levered positions. What the proxy supports is a description of the episode's shape, not an estimate of who was positioned or by how much.

The shape is clear enough in the upper panel. From its 10 July 2024 level, USD/JPY had already fallen 7.3 per cent by 31 July. On 5 August it was 9.8 per cent below that level, the VIX closed at 38.57, and the Nasdaq-100 was 13.4 per cent lower. The yen went on strengthening after the equity market had stabilised, reaching 12.9 per cent below the July level by mid-September.

Section 02The sizing rule, and what it did not do

The proposed amplification channel is mechanical and easy to state. Volatility-targeting strategies scale exposure inversely to a risk forecast, and value-at-risk limits bind less tightly when recent volatility is low. A quiet stretch therefore permits a larger position at unchanged risk appetite, and the same rule demands a reduction once volatility rises.

The second half of that story survives the data. Sixty-day realised volatility in USD/JPY averaged 11.9 per cent between 5 August and 30 September, the 80th percentile of the sample. A constant volatility target would have cut the position by roughly 27 per cent on that move alone, and would have done so while the market was moving against it. The procyclicality is real.

The first half does not survive. There was no unusual calm to be read as permission. Whatever made the position large by early August, this volatility measure would not have shown it, and would not have shown it on any of the three windows tested. That matters practically: the same statistic used to size the position is the one that failed to flag it, and a risk system built on it would have registered nothing until the loss had already happened.

vol-target sizing:     position  =  target_vol / realised_vol
                       => position rises as realised_vol falls

measured, USD/JPY realised volatility, annualised:

    window        1 May - 10 Jul 2024      percentile of 2005-2026 sample
     20-day             8.5%                        53rd
     60-day             8.7%                        48th
    120-day             8.4%                        43rd

    5 Aug - 30 Sep 2024, 60-day            11.9%     80th percentile

    implied position change on that move:  -27%

Section 03What is left of the argument

Removing the calm-before-the-storm claim leaves the article's substance intact and changes what it is for. Negative-skew exposures do accrue smoothly and lose abruptly. Backward-looking risk estimates do understate the risk of a position whose losses are concentrated in rare states, because the rare state has not happened yet and so contributes nothing to the estimate. Sizing rules built on those estimates are procyclical.

What fails is the inference from a quiet tape to a crowded position. The July 2024 volatility reading was ordinary and the unwind was severe, so the reading carried no information about the exposure. A measure that is uninformative about a build-up cannot be used to anticipate it, however good its account of the aftermath.

The volatility that mattered was not low beforehand. It was simply not measuring the thing that was accumulating.

Whether some other observable would have shown the build-up is a fair question this research cannot answer. Futures positioning data, cross-currency basis, and prime-brokerage leverage have all been proposed, and none is available at daily frequency in a form this journal can verify. The result stands as a narrow negative: on the most obvious candidate, realised volatility in the funding currency, there was nothing to see.

This research is analysis and commentary for general information. It is not investment advice, an offer, or a solicitation, and it contains no price forecasts. Volatility statistics are the author's own calculations from the sources cited.