Before we discuss our formal tests of predictive ability to compare the performance of our alternative models, it is useful to look at a summary of out-of-sample performance measures based on a fixed-window rolling regression that starts December 2000 to December 2004 and rolls from March 2005 (owing to our three month holding period) until March 2009. Table 6 calculates out-of-sample average annual carry trade returns, annualized Sharpe ratios, skewness, and the proportion of correctly called long-short positions. Notice that the out-of-sample period includes the turbulent period of the second half of 2008 and the first quarter of 2009 during which carry trade unwinds saw crashing returns to currency speculation.

RW

VAR

VECM

VECM+X

Threshold

0.059

0.055

0.053

0.054

0.055

-4.51%

4.72%

3.87%

5.00%

5.43%

-0.43

0.40

0.36

0.43

0.50

-0.99

0.87

0.91

1.06

1.05

42.3%

55.5%

51.9%

53.8%

57.9%

# RMSE

# Annual Return

Annual Sharp Ratio Skewness

Direction Correctly Called

Notes: RW refers to the null random walk model; VAR, VECM and VECM+X are the three linear models described in the text. “Threshold” refers to the nonlinear threshold model using VX as the threshold variable. Estimation sample for first window: December 2000 to December 2004. Prediction window: March 2005 to March 2009.

Table 6 – Summary of Sample Statistics for Fixed-Window, Rolling-Sample, Out-of- Sample Panel Forecasts and Associated Giacomini-White (2006) Statistics

(a) Summary Statistics

VAR

VECM

VECM+X

Threshold

Annual Return

1.80

2.91

3.08

2.89

Annual Sharp Ratio

1.85

2.88

3.10

2.94

Skewness

1.28

0.42

0.67

0.37

Direction Correctly Called

5.98

6.20

7.31

8.72

(b) Giacomini-White (2006) Statistics

Notes: The numbers reported are t-ratios whose asymptotic distribution is critical value for the 95% confidence level applies. Estimation sample for to December 2004. Prediction window: March 2005 to March 2009.

N(0, 1). Hence, the usual 1.96 first window: December 2000

21