Journal of Applied Economic Sciences
Volume IV/ Issue 1(7)/ Spring 2009
We do not have the same impressive results in currency markets. The regression on the factors
currently used in literature produced a pseudo squared-R of 0.14 with GDP growth and inflation
failing to be statistically significant as leading indicators of currency crises. Ceteris paribus, increases
in stock returns or real interest rates contribute to lower probability of currency crisis while increases
in terms of trade, current account or unemployment rate signal a higher probability of currency crises.
Changes in CDS premiums are statistically significant at 5 percent confidence level in the third model
and they have the expected positive sign, where an increase in the premiums corresponds to a higher
probability of currency crises. However, a test on the slope coefficient of changes in CDS premiums
in the second model cannot reject the null hypothesis that the coefficient is significantly different
from zero. So, changes in CDS premiums are insignificant in presence of currently used factors in
literature. Hence, the inclusion of the factor in the base model does not improve forecasts of currency
crises in emerging markets.
However, one cannot draw final results from conclusions by looking at the coefficients and
marginal effects of the factors only. Statistics on predictive ability of our models should also be
considered. To obtain meaningful statistics about predictions from our data-sample, we set a
probability threshold above which, it is decided that the model predicts a crisis. Other studies choose
a threshold that maximizes the share of correctly classified observations. We follow Peltonen (2006)
and use four different probability thresholds. Table 6 and Table 7 show statistics on predictive ability
of our models for different probability thresholds.
Table 5. Logit regression models predicting currency market crises2
|
Dependent variable: Crisis Indicator Independent variables (t-1): |
Model (1) |
Model (2) |
Model (3) |
|
Coefficients Marginal Effects |
Coefficients Marginal Effects |
Coefficients Marginal 0fects | |
|
Constant |
-6.738983*** |
-6.487072*** |
-4.23099*** |
|
[1.077809] |
[1.086092] |
[0.2203864] | |
|
Sovereign Credit Defalut Swaps |
0.8368797 0.0046655 [1.062977] [0.00602] |
1.63882** .0236229*** [0.6597416] [0.00923] | |
|
Terms of Trade |
0.8058111** 0.0046468 ** [0.3655891] [0.00215] |
0.5876072** 0.0024199** [0.3637605] [0.00172] | |
|
Current Account |
0.0000107*** 6.20e-08*** [4.79e-06] [0.00000] |
8.91e-06*** 3.67e-08*** [5.11e-06] [0.00000] | |
|
Unemployment Rate |
0.0830174** 0.0004787** [0.0354746] [0.0002] |
0.0954109** 0.0003929** [0.0356845] [0.00018] | |
|
Change in Consumer Prices |
0.0759931 0.0004382 [0.0569267] [0.00036] |
0.1838124 0.000757 [0.0814281] [0.00045] | |
|
Real Interest Rate |
-0.01649704 ** -0.0009513** [0.0788175] [0.00045] |
-0.1489212** -0.0006133** [0.0946067] [0.00039] | |
|
GDP Growth |
3.375028 0.0194626 [4.032087] [0.0232] |
3.520752 0.0144992 [4.609704] [0.01952] | |
|
One Month Stock Returns |
-8.855517* -0.0510666* [3.613523] [0.02639] |
-9.64336* -0.0397134* [3.751803] [0.02234] | |
|
Observations |
1209 |
1080 |
1522 |
|
Log Likelihood |
-90.690727 |
-69.647606 |
-120.9699 |
|
Pseudo squared-R |
0.1439 |
0.1636 |
0.0197 |
|
Chi-square |
30.48 |
27.25 |
4.87 |
|
P-Value |
0.0000 |
0.0006 |
0.0274 |
The statistics on the predictive ability of stock market crises show that the second model
outperforms the base model by the share of correctly predicted crises at 0.1 and 0.15 probability
2
Standard errors are in brackets, * significant at 10%, ** significant at 5%, *** significant at 1%, marginal effects give the
estimated slope coefficient, coefficients are estimated for the original logistic model
133