Are there ways to enhance the currency carry trade (long currencies offering high interest rates and short those offering low rates)? In the May 2012 version of their paper entitled “Average Variance, Average Correlation and Currency Returns”, Gino Cenedese, Lucio Sarno and Ilias Tsiakas investigate the ability of components of the currency exchange market risk (variance of the average return for all exchange rates) to predict carry trade returns. Their baseline carry trade portfolio involves U.S. dollar nominal exchange rates, rebalanced monthly. They decompose the market variance into two components: average variance of individual exchange rate returns, and average correlation of exchange rate returns. They examine the effects of changes in these risk components on the entire future distribution of currency trade returns (via quantile breakdowns), focusing on the large losses in the left tail and large gains in the right tail. Using daily spot and forward exchange rates for 33 currencies relative to the U.S. dollar as available during 1976 through February 2009 (15 active exchange rates at the beginning and 22 at the end), they find that:
- Over the entire sample period, the baseline carry trade generates a gross mean annual return of 8.6%, with standard deviation 7.8% and gross Sharpe ratio 1.09. Performance derives mostly from differences in interest rates across countries (accounting for 13.7% per year) and exchange rate variation (accounting for -5.1% per year).
- The overall correlation between the average variance and average correlation components of currency market risk is 0.19, indicating substantial independence but fluctuating considerably for subsamples. The product of these components captures more than 90% of the behavior of currency market variance.
- High average variance significantly predicts large future losses (amplified left tail) for the carry trade, while low average correlation significantly predicts large future gains (amplified right tail) via elevated diversi cation. Overall currency market variance is a weaker predictor than either component separately. Similarly, implied volatility indexes (VIX for the equities market and VXY for the currency exchange market) do not significantly predict carry trade returns.
- An enhanced carry trade strategy, implemented out-of-sample using inception-to-date monthly signals from average variance and average correlation starting with three years of data, considerably outperforms the baseline carry trade. Based on median trading frictions for each currency over the entire sample period (averaging 0.06% for spot rates and 0.094% for forward rates across all currencies):
- A strategy that exploits the relationship between current average variance and next-month carry trade return by exiting positions in the left tail (lowest past returns) after increases in average variance generates a net annual Sharpe ratio as high as 0.98 (depending on the threshold for the left tail), compared to 0.74 for the baseline carry trade.
- A strategy that exploits the relationship between current average correlation and next-month carry trade return by doubling positions in the right tail (highest past returns) after decreases in average correlation generates a net annual Sharpe ratio as high as 0.97 (again depending on the threshold for the right tail).
- A strategy that combines average variance and average correlation signals generates a net annual Sharpe ratio as high as 0.96.
In summary, evidence indicates that investors can enhance the risk-adjusted performance of the currency carry trade by exploiting information in the average variance of individual exchange rate returns and the average correlation of exchange rate returns.
Cautions regarding findings include:
- Data collection and calculations appear to be potentially burdensome (or costly if delegated).
- The study calculates enhanced carry trade results for multiple tail definition thresholds, thereby incorporating data snooping bias. The best outcomes therefore likely overstate reasonable expectations for future performance. Moreover, even out-of-sample tests can incorporate data snooping bias when strategy construction relies on prior studies involving the same set of data.
- If trading frictions vary over time, then applying median frictions for the entire sample period may mislead.
