An Examination of G10 Carry Trade and Covered Interest Arbitrage Before, During, and After Financial Crises

Abstract

This paper examines and compares the trading strategies of carry and covered interest arbitrage. This study constructs portfolios for G10 countries based on interest rates’ spot and forward exchange rates. We extend the prior literature by focusing on the profitability of the strategies during and around the two crisis periods, comparing both carry trade (CT), i.e., unhedged, and covered interest arbitrage (CIAT), i.e., hedged. We find that both CT and CIAT have variable profits during the period examined, with both strategies’ profits generally concentrated in the pre-crisis period and most losses in the post-crisis period.

1. Introduction

Studies on the forward premium puzzle have shown that investors can profit from trading strategies involving arbitrage techniques such as carry trade (CT) or covered interest arbitrage trade (CIAT), which is a contradiction to uncovered interest parity (UIP) theory (Burnside et al., 2007, 2011; Xanthopoulos, 2011). Reasons cited in the literature for this puzzle include a time-varying risk premium (C. M. Engel, 1984; Fama, 1984), methodological issues (Bilson, 1981), investor behavior (Burnside et al., 2011), momentum (Baillie & Chang, 2011), peso events (Burnside et al., 2011), and size of economy (Hassan, 2013).

Uncovered interest rate parity (UIP) posits that interest rate differentials should equal expected change in foreign exchange rates. However, studies find contrary evidence where higher interest rate differentials lead to FX appreciation instead of depreciation, a violation of UIP, known as the forward premium puzzle. Thus, UIP and the forward premium puzzle is important to explore for several reasons, such as how trading strategies around this violation, like CT, are linked to currency crashes (Brunnermeier et al., 2008), or policy implications linked to FX movements associated with trade and capital flows that can affect international trade competitiveness (C. Engel, 1996).

These ongoing discussions highlight the importance of examining these unresolved questions and providing evidence to the literature. Although there are numerous explanations for the premium puzzle, only more recently has there been research examining how economic conditions like the crisis affect the profitability of these strategies. Coudert and Mignon (2013) finds that sovereign default risk, assumed by investing in high yield currencies, increases profits during boom periods and losses during crises. See also Faugere (2013). Cho et al. (2019) and Chen and Lin (2022) both use Markov switching models, finding evidence of two regimes: a low-volatility regime in which carry trades are profitable and UIP is violated, and a high-volatility regime related to negative carry-trade returns and financial crisis. Fan et al. (2022) examines the decline in performance of the carry-trade strategy since the 2008 Global Financial Crisis (GFC), attributing it to the disappearance of the downside risk in the post-crisis carry-trade portfolio. Elias et al. (2024) finds that the currency carry-trade hypothesis explains the forward premium anomaly only when interest rate differentials are positive, and that negative interest rate differentials exist during crisis periods, postulating that investors are seeking safe-haven currencies. However, none of these studies conducted a simultaneous comparison of CT and CIAT before, during, and after crises that would allow for a comprehensive examination of the UIP in such an environment.

In the study, we add to the literature by comparing the performance of unhedged vs. hedged currency trading strategies during the periods surrounding two crises. In regard to the tech bubble and the GFC, the latter, in particular, remains one of the most significant events for the past century, and it likewise offers lessons for future shocks, such as the recent COVID-19 pandemic. We do this first by replicating the CT strategy of an international commercial bank and then comparing its performance with CIAT around the same period. This allows us to examine profitability using a strategy from institutional investors in the industry, and then, second, by contrasting the performance of these strategies between crisis and non-crisis sub-periods. Finally, we use regime modeling to investigate evidence of state dependence in trading strategies and determine if a switch in trading regimes has occurred in our data, and then we test for the presence of ARCH and GARCH effects. Our work has implications for the reliability of using forward rates as effective and consistent measures of investor expectations during periods of crisis and non-crisis.

Motivation for this paper stems from the continued popularity and widespread use of both CT by practitioners worldwide (Galati & Melvin, 2004) and CIAT. For example, in Austria, 12% of households have their housing loans denominated in foreign currencies, a form of carry trade, notably in Swiss francs and Japanese yen (Beer et al., 2008). Likewise, huge sums were invested in carry-trade strategies involving Japanese yen of up to JPY one trillion by early 2007 (The Economist, 2007). Then, in response to the GFC, many carry-trade positions were reversed (Pilling, 2008; McKinnon et al., 2010). These examples highlight how CT and CIAT can affect the volume of currency trading, consequently affecting the flows of currency that can affect FX rates and the international competitiveness of economies. Thus, it is important to examine the impact of financial crises on CT and CIAT by utilizing the approach of industry participants.

We find in this paper that both CT and CIAT have variable profits during the period examined, with both strategies’ profits generally concentrated in the pre-crisis period and most losses in the post-crisis period. The crisis period generally yields profits for both strategies, counter to some of the literature. Using regression analyses, we find that the coefficient on the crisis dummy variable has a positive significant relation with profits, and we find a positive significant relation between risk premium and profits. Treating the crisis period as an exogenous shock, we find that the crisis period and risk premium statistically explain some variation in profits in the sample. In sub-period analyses of crisis vs. non-crisis, we find that the statistically positive coefficient on risk premium is due to the crisis period. This finding supports the concept of time-varying risk posited by Fama (1984). Given that the crisis period has a statistical association with profits, we examine profitability of the strategies, for which results suggest systematic change over time. Our regime modeling of the data reveals states differentiated by volatility levels, and because of the time-varying volatility we have uncovered, we use ARCH and GARCH models to analyze the data for a process of continuous time variation, as they are popular models to treat time series that exhibit this feature. The results indicate evidence of ARCH effects in the regression models with CIAT and CT, and the difference in profits from the CT and CIAT as dependent variables. However, in general, we do not find evidence of GARCH effects in our models of CT and CIAT strategies.

To our knowledge, no prior study compares the profitability of carry- and covered-trade strategies in a direct match-up, nor do previous studies compare the profitability and volatility of the two strategies simultaneously during periods of market turmoil. The contributions of the paper are as follows. First, we contribute to the literature on profitability for both CT and CIAT by comparing crisis and non-crisis periods from a practitioner standpoint. Second, we provide evidence on the performance of both strategies and draw comparisons prior to, during, and after the 2008 financial crisis, reporting the trends of profits during these periods. Third, we document factors driving the differences in the profitability between crisis and non-crisis periods. Furthermore, using regime modeling, we find evidence of crisis and non-crisis states, and we find evidence of ARCH effects in the data.

Our study has implications for practitioners and traders in designing trading strategies that account for structural breaks. Likewise, our study has potential policy implications where with regime-dependent outcomes would suggest that central bank interventions may inadvertently create arbitrage opportunities.

The remainder of the paper is organized as follows. Section 2 provides more background on carry trade, covered interest arbitrage, and a literature review. Section 3 describes the data and methodology. Section 4 presents and discusses the results. Section 5 concludes the paper.

2. Literature Review and Hypothesis

2.1. Literature Review

Forward premium puzzle is well documented and has been the subject of several studies (Baillie & Chang, 2011; Bilson, 1981; Burnside et al., 2011; C. M. Engel, 1984; Fama, 1984; Hassan, 2013). Likewise, carry trade is a widely used strategy (Galati & Melvin, 2004), though its existence conflicts with theory on interest rate parity. While conventional parity conditions predict no profitability, several studies have documented the profitability of both carry trade and covered interest arbitrage strategies (Bilson, 1981; James et al., 2009). Bilson (1981) show that buying the currency whose interest rate was relatively high will yield expected profits without bearing much risk. James et al. (2009) shows that carry trade has been successful in returning profits for three decades and is a robust trading strategy.

In general, some of the reasons cited for the existence of profitable trading strategies that are in conflict with theory are risk premia, methodology, and behavioral biases/market anomalies such as momentum. Bilson (1981) posits that transaction costs, information costs, and risk aversion, as well as forecast errors, could be causes of the forward premium puzzle, while Fama (1984) argues the forward market to be efficient and finds that most of the variation in the forward rates is due to premia. In addition, Fama finds a negative relationship between expected future spot rate and the premium component of forward rates. Thus, the lock between the premium in the forward rate and interest rates on the nominal bonds of two countries is the direct consequence of interest rate parity. Ma and Meredith (2002) and Snaith et al. (2013) both find that the forward premium bias decreases as the exchange-rate horizon is extended.

However, several studies reject these explanations. Domowitz and Hakkio (1985) test the alternative hypothesis, which explains UIP using risk premia without sacrificing the notion of market efficiency. They posit that, contrary to many rational expectation models, exchange rates do not incorporate expectations about the future. They argue that this is evident from the relation between the time when decisions are made and the resolution of uncertainty, since all decisions are made after the resolution of uncertainty, and thus there is no speculative component in the demand function they use. They find evidence against the unbiasedness hypothesis for a majority of the currencies and find little support for the conditional variance of the exchange-rate forecast error being an important determinant of the risk premium. Froot and Frankel (1989) find that, for the period 1981–1985, the variation in the forward dollar discount of the four most actively traded currencies (Deutschmark, Swiss franc, Yen, British pound) reflects changes in expected depreciation rather than risk premia, and that forward discount bias is primarily attributable to irrationality. Zhuang (2015) finds that the forward premium puzzle is caused by large but infrequent shocks to spot exchange rates, which dissipate over longer horizons as market participants learn to make better forecasts.

Melvin and Taylor (2009) reviews events linked to the global financial crisis and impact on exchange rates. They document a short period of carry-trade unwinding, followed by a period of good returns. They explain that the unwinding in carry trade was fueled by volatility in the currency markets and subsequently increased volatility in other types of financial assets. Baillie and Chang (2011) find that carry and momentum trading somewhat explains the level and length in forward premium anomaly (i.e., UIP deviation). They find that these strategies have some short-term profitability, but the likelihood of reversals to UIP increases. They attribute this to two different regimes, where exchange rate volatility differs. Menkhoff et al. (2012) examine volatility risk associated with foreign exchange and link it to cross-sectional variation in excess returns associated with carry trades. They find that risk explains returns due to speculation surrounding currencies and returns, serving as compensation for bearing these risks, with carry-trades performing poorly during periods of high market turmoil. Our study examines a similar argument, but we focus on contrasting the carry trade with CIAT.

Mancini et al. (2013) examine liquidity in the foreign exchange (FX) market. They document that there is illiquidity in the market despite the perception that the FX markets are liquid, and that this illiquidity can heighten losses in a market downturn by more than 25%, using a carry-trade strategy they develop. They find that more liquid currencies are less sensitive to systematic market FX liquidity, while the opposite is true for less liquid FX rates. They also find that low interest-rate currencies insure against liquidity risk, and high interest rates create exposure to liquidity risk. Hassan (2013) shows that country size introduces cross-country variation, explaining a large percentage of the cross-sectional variation in currency returns. Hassan posits that debts issued by larger economies hedge against consumption risk, and this leads to a permanent difference in both real and nominal interest rates compared with smaller economies, yielding consistent deviations from UIP. Hassan and Mano (2019) find that currencies with persistently higher forward premia due to cross-sectional asymmetries have greater expected carry-trade returns than others.

Du et al. (2018) report violations of covered interest parity (CIP) after the GFC and examine what leads to the deviations. They find consistent systematic deviation from CIP after the crisis among G10 currencies, and that credit risk has no explanatory power for the persistence of this CIP violation. They attribute this persistent violation to high costs tied to regulatory restrictions on financial institutions after the crisis and funding disparities in savings in investments across currencies. Yamani (2019) examines the carry-trade crashes and how momentum strategy can be incorporated in a diversification role. The paper combines momentum with carry trade to develop a strategy that yields meaningful profits during crisis periods and finds the strategy to be a good hedge, providing diversification benefits in crisis periods.

2.2. Hypotheses

Although theory suggests that when parity conditions are met, CT and CIAT will yield zero profits, the underlying assumptions and mechanisms for both outcomes are different. Therefore, it is essential to establish whether investors can earn profits using either strategy. This will allow for a test of the efficiency while observing the differences in the profits accruing to each strategy, so that we can establish the differences between CT and CIAT performance. Note that evidence from previous research is mixed. Brunnermeier et al. (2008) and Bansal et al. (2012) report the profitability of CT, while Brunnermeier et al. (2008), and Melvin and Shand (2017) report losses from CT due to a variety of conditions. Froot and Thaler (1990) found that profits to CIAT do not exist due to market frictions and transaction costs, among other factors. Therefore, it is important to establish the profitability of CT and CIAT strategies. Subsequently, we test the following hypothesis:

• H1a: CT and CIAT strategies will yield profits.
• H1b: CT and CIAT strategies will yield losses.

Since the mechanisms driving profits differ for CT and CIAT, we further hypothesize the following:

• H1c: There is a statistically significant difference in profits from both strategies.

In examining any hypotheses on CT and CIAT profitability using regression models, an implicit assumption is that the relationship is linear in the independent variables. Given evidence that the macroeconomic environment can influence the profitability of these strategies because of changes in risk appetite or risk perceptions, due to a crisis or event, we also explore the potential for non-linear or state-dependent relationships using regime modeling.1 The regime model posits two states where all parameters are endogenously determined by the data, including the period(s) at which a change in state occurs and the computed Markov Probabilities being in a state. We fit two models in this paper, one in which the states are solely determined by volatility levels, and the other which additionally allows states to be defined by different coefficients on the independent variables. Likewise, because of the non-linear nature of volatility, we examine the data for the presence of ARCH and GARCH processes that might lead to the appearance of state-like patterns.

Consequently, we predict the following:

• H2: Crisis will have significant impact on profitability of both strategies.
• H3: We will observe different regimes around crisis.
• H4: We will observe ARCH or GARCH effects in the data.

Note that CT relies more on market movements and risk to generate profits, and benefits from volatility, as compared to CIAT, which instead, seeks to exploit market imperfections and arbitrage opportunities at low risk. Therefore, given the potential non-linear nature of volatility, it is important to contrast the differential impact of volatility and state dependence on the profitability of these two strategies. These tests also contribute to studies on the risk associated with CT, such as that by Daniel et al. (2014), who find that investors expect compensation for different risks.

3. Data and Methodology

3.1. Data

We obtained monthly data for the period from 2000 to 2018 from Datastream for the spot and forward exchange rates, and Libor rates, for the G10 currencies: Euro (EUR), Japanese yen (JPY), Great Britain pound (GBP), Australian dollar (AUD), Canadian dollar (CAD), Swiss franc (CHF), New Zealand dollar (NZD), Norwegian krone (NOK), and Swedish krona (SEK). The currencies’ quotes are relative to the USD.2 The exchange and forward rates are in indirect quotes (quoted in foreign currency per dollar). We use Libor and T-Bill rates for interest rates. We also obtain the Consumer Price Index (CPI), Treasury Yields, and Trade Weighted Dollar Index from the St. Louis Federal Reserve Bank.

3.2. Methodology

To construct the portfolio, we first ranked the ten currencies based on their Libor interest rates. Then, we invested USD 200 in each of the five currencies with the highest interest rates and financed our long positions by shorting USD 200 in each of the five currencies with the lowest interest rates; thus, the portfolio required no upfront investment. At the end of each month, the portfolios are rebalanced due to changes in the ranking of the currencies based on their prevailing interest rates, for both the short and long portfolios. This aligns with real-world financial portfolio-formation processes. This approach is similar to that used to form the Deutsche Bank Currency Returns Carry Trade Index. For ease of comparison, we also calculate the returns to the long and short positions separately and take the average. Portfolio construction is the same for the CT and CIAT strategies, except that CT involves no hedging using forward rates, whereas CIAT uses forward rates for hedging. We assume the position of a US investor and assume no transaction costs or taxation. Figure 1 illustrates the portfolio construction process.

We first perform a univariate comparison of profits between the two strategies. For multivariate analyses, we employ the following model:

PROFIT_t = α + β₁CRISISt + β₂SP500t + β₃INFLATIONt + β₄RISKPREMt + e_t

(1)

where PROFIT is the profit (or loss) of CT or CIAT strategy; and CRISIS is a dummy variable defined following National Bureau of Economic Research (NBER)3 criteria to be from March 2001 to November 2001 and from December 2007 to June 2009. Other variables are monthly data for the SP500 is the return on the S&P500 index; INFLATION is the US inflation rate; and RISKPREM is the yield spread between 5-year treasuries and 1-year treasuries (Fama & Bliss, 1987).

4. Results and Discussion

Table 1. Summary statistics.

	Sample (N = 216)		Non-Crisis (N = 188)		Crisis (N = 28)
Panel A: Spot Rates
Country	Mean	STD	Mean	STD	Mean	STD
UK	0.6261	0.0744	0.6288	0.0739	0.6083	0.0769
Canada	1.2210	0.1892	1.2160	0.1820	1.2546	0.2330
Norway	6.9206	1.1972	6.9748	1.2026	6.5566	1.1125
Australia	1.3264	0.2858	1.3098	0.2795	1.4376	0.3080
Japan	106.2776	13.9935	105.9819	14.2763	108.2631	11.9578
Sweden	7.7653	1.1977	7.7472	1.1599	7.8872	1.4447
Euro	0.8404	0.1338	0.8443	0.1275	0.8142	0.1706
New Zealand	1.5506	0.3570	1.5248	0.3514	1.7233	0.3523
Switzerland	1.1661	0.2568	1.1557	0.2612	1.2356	0.2160
Panel B: Forward Rates
Country	Mean	STD	Mean	STD	Mean	STD
UK	0.6268	0.0737	0.6293	0.0733	0.6099	0.0758
Canada	1.2220	0.1901	1.2172	0.1827	1.2546	0.2350
Norway	6.9433	1.2069	6.9921	1.2128	6.6155	1.1332
Australia	1.3336	0.2858	1.3161	0.2786	1.4513	0.3098
Japan	104.9494	13.8464	104.6265	14.0663	107.1175	12.2771
Sweden	7.7701	1.2023	7.7465	1.1613	7.9283	1.4629
Euro	0.8397	0.1337	0.8432	0.1273	0.8164	0.1717
New Zealand	1.5611	0.3597	1.5346	0.3540	1.7390	0.3533
Switzerland	1.1625	0.2551	1.1518	0.2593	1.2344	0.2160
Panel C: Interest Rates
Country	Mean	STD	Mean	STD	Mean	STD
UK	0.0282	0.0225	0.0262	0.0226	0.0419	0.0166
US	0.0410	0.0140	0.0415	0.0138	0.0375	0.0150
Norway	0.0342	0.0224	0.0304	0.0206	0.0592	0.0182
Australia	0.0455	0.0172	0.0438	0.0164	0.0569	0.0187
Canada	0.0237	0.0156	0.0234	0.0161	0.0262	0.0122
Japan	0.0028	0.0024	0.0024	0.0019	0.0056	0.0034
Sweden	0.0205	0.0168	0.0179	0.0155	0.0379	0.0145
Euro	0.0238	0.0173	0.0217	0.0173	0.0381	0.0094
New Zealand	0.0481	0.0206	0.0458	0.0194	0.0641	0.0215
Switzerland	0.0077	0.0126	0.0065	0.0124	0.0158	0.0103

reports summary statistics of LIBOR, spot, and forward exchange rates, in Panel A–C, respectively. During the sample period, Australia and New Zealand have the highest interest rates, while Japan and Switzerland have the lowest.

Table 2. Distribution of currencies in short and long positions.

Currencies	Number of Months Shorted	% of the Currency Being Shorted	Number of Months Being Longed	% of the Months Being Longed	Total
Great Britain Pound (GBP)	114	52.78%	102	47.22%	216
Canadian dollar (CAD)	106	49.07%	110	50.93%	216
Norwegian krone (NOK)	42	19.44%	174	80.56%	216
Australian dollar (AUD)	0	0.00%	216	100.00%	216
Japanese yen (JPY)	216	100.00%	0	0.00%	216
Swedish krona (SEK)	171	79.17%	45	20.83%	216
Euro (EUR)	151	69.91%	65	30.09%	216
New Zealand dollar (NZD)	0	0.00%	216	100.00%	216
Swiss franc (CHF)	216	100.00%	0	0.00%	216
US dollar (USD)	64	29.63%	152	70.37%	216

presents the frequency and percentage of times that we short or long each currency in the sample. Note that for the entire sample period, we short the Swiss Franc and Japanese yen and long the New Zealand and Australian dollar. These results correspond to the notably low and high average interests of these currencies, as observed in Table 1.

In

Table 3. Summary statistics of CT and CIAT profits.

Variable	Mean	Median	STD	t-Stat	Min	Max
Ending CT long position	859.36	800.08	91.57	137.93	800.03	1000.20
Ending CT short position	940.77	1000.01	91.53	151.06	800.02	1000.11
Net CT USD profit	−81.42	−199.95	183.10	−6.54	−199.97	200.14
Net CT profit %	−8.14%	−19.99%	18.31%	−6.54	−20.00%	20.01%
Ending CIAT long position	855.58	799.09	89.89	139.89	790.66	1002.49
Ending CIAT short position	944.22	1002.28	92.66	149.76	796.87	1012.30
Net CIAT USD profit	−88.64	−205.23	182.48	−7.14	−215.12	197.02
Net CIAT profit %	−8.86%	−20.52%	18.25%	−7.14	−21.51%	19.70%
USD Profit difference between CT and CIAT	7.23	7.19	2.95	36.01	−2.30	20.61
% Profit difference between CT and CIAT	0.72%	0.72%	0.29%	36.01	−0.23%	2.06%

, we present a summary of the statistics of the profits for both strategies. For CT, the average ending dollar values of the long and short positions are USD 859.36 and USD 940.77, respectively, leading to a loss for the overall strategy of USD 81.42. In comparison, the average ending dollar values of the long and short positions for CIAT are USD 855.58 and USD 944.22, resulting in a loss of USD 88.64. The dollar losses for both strategies are statistically significant. Overall, the results suggest that the two trading strategies are not profitable over the sample period. However, high standard deviations indicate that the mean values may be misleading. Supporting our concern is the disparity of the mean and median values of the net profits from both trading strategies. Note that the dollar profit difference of USD 7.23 between the strategies is statistically significant.

In

Table 4. Net CT and CIAT profits by year.

Year	N	USD Net CT Profit	% Net CT Profit	USD Net CIAT Profit	% CIAT Profit	USD Profit Difference	% Profit Difference
2000	12	−133.2781	−13.33%	−140.9334	−14.09%	7.6553	0.7655%
2001	12	200.0921	20.01%	192.9214	19.29%	7.1707	0.7171%
2002	12	200.1020	20.01%	191.7862	19.18%	8.3158	0.8316%
2003	12	200.0848	20.01%	192.0000	19.20%	8.0848	0.8085%
2004	12	200.0947	20.01%	192.9736	19.30%	7.1211	0.7121%
2005	12	−166.5925	−16.66%	−177.0444	−17.70%	10.4519	1.0452%
2006	12	−199.9358	−19.99%	−209.7607	−20.98%	9.8249	0.9825%
2007	12	33.4217	3.34%	25.3055	2.53%	8.1162	0.8116%
2008	12	0.0953	0.01%	−10.9028	−1.09%	10.9981	1.0998%
2009	12	−199.9570	−20.00%	−206.1892	−20.62%	6.2322	0.6232%
2010	12	−199.9486	−19.99%	−206.0389	−20.60%	6.0902	0.6090%
2011	12	−199.9519	−20.00%	−205.5651	−20.56%	5.6133	0.5613%
2012	12	−199.9570	−20.00%	−207.0323	−20.70%	7.0753	0.7075%
2013	12	−199.9661	−20.00%	−206.1804	−20.62%	6.2143	0.6214%
2014	12	−199.9585	−20.00%	−206.6506	−20.67%	6.6921	0.6692%
2015	12	−199.9549	−20.00%	−205.9647	−20.60%	6.0098	0.6010%
2016	12	−199.9584	−20.00%	−203.7319	−20.37%	3.7735	0.3774%
2017	12	−199.9617	−20.00%	−204.5883	−20.46%	4.6266	0.4627%

, we report the mean profits of the two strategies by year. Interestingly, the results show that both strategies experience profitability changes around the period of the financial crisis. Profits that were different in the years leading up to 2008 were 2000, 2005, and 2006. For CT, the highest profit occurs in 2004, with the greatest loss occurring in 2017. For CIAT, the highest profit is in 2009, while the greatest loss is in 2006. The change in sign and stickiness near the global crisis is an interesting observation that warrants further exploration.

Table 5. Net CT and CIAT profits: non-crisis vs. crisis periods.

Variable	Non-Crisis Period (N = 188)		Crisis Period (N = 28)		Tests of Differences in Means
	Mean	Median	Mean	Median	Diff. Mean	t-Stat
USD Net CT profit	−97.81	−199.95	28.65	200.09	−126.47	(−3.50) ***
%Net CT profit	−9.78%	−20.00%	2.87%	20.01%	−12.65%	(−3.50) ***
USD Net CIAT profit	−104.86	−205.41	20.25	187.81	−125.12	(−3.47) ***
%Net CIAT profit	−10.49%	−20.54%	2.03%	18.78%	−12.51%	(−3.45) ***
USD profit difference	7.05	7.05	8.40	8.34	−1.35	(−3.40) ***
%Profit difference	0.71%	0.70%	0.84%	0.83%	−0.13%	(−2.28) **

Note: ** p < 0.05; *** p < 0.01.

compares the average and median profitability of CT and CIAT in the non-crisis and crisis periods, where we follow the NBER crisis period definition. The average (median) dollar net profit for CT in the non-crisis period is USD −97.81 (199.95), compared to the average (median) dollar net profit in the crisis period of USD −28.65 (200.09). Meanwhile, the average (median) dollar net profit for CIAT in the non-crisis period is USD −104.86 (−205.41), in contrast to the average (median) dollar net profit in the crisis period of USD −20.25 (187.81). The differences in mean tests for CT and CIAT profits are statistically different from the non-crisis to the crisis period at the 5% level, suggesting a material difference in profits between periods.

In

Table 6. Cross-sectional analyses CT and CIAT profits.

	FULL		Before Crisis		During Crisis		After Crisis
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
	CT	CIAT	CT	CIAT	CT	CIAT	CT	CIAT
crisis	128.54 ***	127.55 ***
	(3.93)	(3.92)
sp500ret	−176.73	−158.38	−296.15	−286.72	276.35	308.55	−0.02	6.57
	(−0.67)	(−0.61)	(−0.75)	(−0.72)	(0.49)	(0.56)	(−1.47)	(0.99)
cpichg	11,723.82 ***	11,679.12 ***	−508.53	−500.88	13,088.60 *	13,127.07 *	−0.09	72.75
	(3.27)	(3.27)	(−0.10)	(−0.09)	(1.97)	(2.00)	(−0.34)	(0.63)
riskprem	6818.98 ***	6872.71 ***	18,265.17 ***	18,269.35 ***	−12,789.96	−12,772.33	0.39 ***	−47.75
	(4.41)	(4.47)	(9.08)	(9.04)	(−1.44)	(−1.45)	(3.37)	(−0.97)
twexb	6.62 ***	6.65 ***	−3.18	−3.10	7.24 **	7.34 **	−0.00	0.07 ***
	(6.16)	(6.22)	(−1.32)	(−1.28)	(2.43)	(2.48)	(−1.26)	(2.75)
intercept	−920.96 ***	−932.50 ***	252.72	234.17	−637.28 *	−655.75 *	−199.95 ***	−212.67 ***
	(−7.59)	(−7.73)	(0.93)	(0.86)	(−1.95)	(−2.02)	(−28,883.34)	(−72.66)
adj. R-sq	0.27	0.27	0.55	0.55	0.17	0.18	0.14	0.08
N	216	216	86	86	28	28	102	102

Note: * p < 0.10; ** p < 0.05; *** p < 0.01.

, we report the results of multivariate analyses of strategy profitability and the relationship to crisis and risk. Regression models (1) and (2) report the results for the entire sample period. Each model for the CT and CIAT uses the dollar profit for each strategy as dependent variables. Consistent with the univariate analyses in Table 4 and Table 5, there is a statistically significant positive relationship between the crisis dummy variable and CT and CIAT profits. These results imply that the profits (and/or losses) accrued for the two trading strategies are significantly affected in the crisis period. The other control variables, cpichg, the proxy for inflation; riskprem, the proxy for risk premium; and twexb, the proxy for weighted dollar index, are positive and significant. We interpret this as compensation for risk factors and inflation tied to the strategies. We also note a statistically significant negative relationship between market returns and profits that could indicate hedging potential for both CT and CIAT strategies.

In subsequent regressions, we examine the profitability of the two strategies before the crisis (columns 3 and 4), during the crisis (columns 5 and 6), and post-crisis (columns 7 and 8). Dependent variables are, again, dollar profits from CT and CIAT. From sub-period analyses, we observe that the positive relation between profits and riskprem observed in the full sample is mostly from the pre-crisis period. We observe similar results for CT after the crisis. There is a statistically positive relation between the riskprem for both CT and CIAT strategies before the crisis, whereas post-crisis, this relationship exists only for the CT strategy.

Because the univariate analyses conducted in Table 4 provide evidence of state-dependent changes in profitability, we examine the data using regime modeling. Figure 2 presents the smoothed probabilities from regime modeling using non-switching and switching regressors,4 which differentiate State 1 from State 2 solely based on fitted volatility levels. Note again that regime modeling determines all parameters endogenously from the data (including the volatility levels and Markov Probabilities of being in a state)—meaning that we do not instruct the model as to the characteristics or parameters defining a state, the period(s) during which a change in state occurs, or the corresponding probabilities of remaining in a state. All parameters are determined by the data. Figure 2a,b and Figure 3a,b plot the Markov Probabilities for CT and CIAT for switching and non-switching regressors, respectively, and show a distinct change in state starting in 2001 and 2008, corresponding to the financial crises. We only show the smoothed probabilities of being in State 1, since the smoothed probability of being in State 2 is a mirror image of the State 1 graphs. These results are consistent with Baillie and Chang’s (2011) assertion that carry trade breaks down when markets become more turbulent, and a reversion to UIP is more likely to be observed during such periods of volatility.

In

Table 7. Regime modeling of profits.

Panel A: Regime Modeling with Switching Regressors
		Carry Trade (CT)		Covered Trade (CIAT)
		State 1	State 2	State 1	State 2
Markov Prob	1	0.9753	0.0247	0.9790	0.0210
	2	0.0066	0.9335	0.0551	0.9449
		Regime 1	Regime 2	Regime 1	Regime 2
		Common (Switching Regressors)
		Regime 1	Regime 2	Regime 1	Regime 2
SP500ret		−25.37	46.10	−9.84	60.95
z-statistic		(−0.82)	(1.01)	(−0.29)	(1.46)
CPIchg		−632.23	2404.87	−650.71	2172.76
z-statistic		(−1.61)	(3.04)	(−1.50)	(3.00) ***
RiskPremium		−990.20	−855.47	−935.11	−818.57
z-statistic		(−5.61) ***	(−2.61) ***	(−4.80) ***	(−2.73) ***
TWEXB		−1.74	1.75	−1.81	1.68
z-statistic		(−92.43) ***	(39.13) ***	(−86.92) ***	(41.14) ***
Log (Sigma)		2.68	2.74	2.78	2.65
z-statistic		(45.46) ***	(30.16) ***	(47.14) ***	(29.17) ***
Log-Likelihood		−923.6155		−932.94
Panel B: Regime Modeling with Non-Switching Regressors
		Carry Trade (CT)		Covered Trade (CIAT)
		State 1	State 2	State 1	State 2
		0.9790	0.0210	0.9793	0.0207
		0.0551	0.9449	0.0542	0.9458
		Regime 1	Regime 2	Regime 1	Regime 2
		Common (Non-Switching Regressors)
		Carry Trade (CT)		Covered Trade (CIAT)
SP500ret		−36.38		−222.50
z-statistic		(−1.15)		(−0.65)
CPIchg		−535.14		−537.50
z-statistic		(−1.37)		(−1.25)
RiskPremium		−973.96		−912.83
z-statistic		(−5.61) ***		(−4.76) ***
TWEXB		−1.74		−1.80
z-statistic		(−93.17) ***		(−87.59) ***
		Regime 1	Regime 2	Regime 1	Regime 2
Log (Sigma)		2.67	6.02	2.77	6.01
z-statistic		(44.60) ***	(67.12) ***	(46.48) ***	(67.20) ***
Log-Likelihood		−1132.94		−1147.69

Note: *** p < 0.01.

, we report the coefficient results of our regime modeling. In Panel A, we report results for the model with switching regressors, which permit all of the regressor coefficients to have state-dependent values. Again, the residual volatility levels (sigma) of the post-crisis state appear to be lower in comparison to the pre-crisis state. Moreover, the post-crisis state appears to be largely less sensitive to our regressors than the pre-crisis state (with the exception of S&P500ret), as indicated by a decline in regression coefficients. Panel B presents the results for the model with non-switching regressors. Overall, regime modeling provides a picture of state dependence in the data, with a significant (and possibly permanent) change in state occurring around the time of the crisis periods.

While regime-switching captures potential state dependence in volatility regimes, ARCH/GARCH complements it by examining the data for persistence of volatility clustering in the framework of continuous time. We present the results of our ARCH/GARCH models in

Table 8. Testing for ARCH—M and GARCH effects (z-statistics in parentheses).

	(1)		(2)
	CT %	z-Statistic	CIAT %	z-Statistic
GARCH^2	0.518	(6.068) ***	0.518	6.05
crisis			0.033	(0.32)
sp500ret	−23.473	−11.150	−23.425	−1.15
cpichg	−1228.656	(−3.006) ***	−1228.313	(−2.99) ***
riskprem	−3019.786	(−19.220) ***	−3020.538	−18.59
TWEXB	−1.564	(−68.87) ***	−1.564	(−68.47) ***
Variance Equation
constant	4.784	1.081	4.801	1.075
Residual2(−1)	0.747	3.943	0.750	(3.89) ***
GARCH(−1)	0.370	5.449	0.369	(5.217) ***
N	216.000		216.000
Adj. R2	0.249		0.246
Log-Likelihood	−118.135		−1181.142

Note: *** p < 0.01.

. CT is the dependent variable for model (1), while CIAT is the dependent variable for model (2). Heteroscedasticity tests indicate ARCH in mean and GARCH effects at lag 1 for all models. The presence of ARCH-M effects in our data is consistent with our regime modeling results, in that both high- and low-volatility periods for CT and CIAT exhibit some persistence. This result is consistent with protracted periods of profitability and losses for each of the strategies that we observe in the data.

5. Conclusions

We analyzed returns associated with carry trade (CT) and covered interest arbitrage (CIAT) in the periods before, during, and after two crises. Our results show a positive relationship between CT and CIAT profits, and find different profit levels before, during, and after the crisis for each of the two trading strategies. We find that most profits are concentrated in the pre-crisis period and losses in the post-crisis period. Notably, we find that the strategies were profitable during crisis when compared with non-crises periods. We further use regime modeling, finding evidence of a structural break, resulting in change in the profitability of CT and CIAT trading strategies, and finding the presence of ARCH effects. Our results highlight the fragility of arbitrage strategies, offering lessons applicable to future disruptions to the currency markets, and our work has implications for traders and market participants who rely heavily on following the same strategy (or currency) regardless of environment or trading regime. For instance, the value to a trader could be the reversal of their trading strategy by recognizing a regime change, which in turn could lead to increased profitability or reduced losses. We believe this to be a topic ripe for refined future research in being able to provide early signals or flags indicating a change in the trading regime. Another extension could examine the relationship between returns to CT and CIAT in combination with portfolio hedging strategies for financial instruments.

We acknowledge certain limitations of the paper, such as the frequency of the data,5 that we use different length sample periods before and after the crisis, and that we assume frictionless trading for simplicity. Regarding the latter, traders should take into consideration market frictions, such as transaction costs and taxes when executing CT and CIAT. Again, we believe that addressing these issues provides fertile ground for future applied research.