threshold. The second model correctly predicts 27.27% and 33.33% while the base model predicts
21.05% and 16.67% of the crises at 0.1 and 0.15 thresholds respectively. At 0.25 probability
threshold, the base model predicts one crisis and makes no false alarm while the second and third
models correctly predict two crises out of five. Adding CDS premium changes to the base model
improves the predictive ability and at best, it predicts 40% of the crises.
Changes in CDS premiums are statistically insignificant in the second model, and hence the
predictive ability of the base model and the second model is the same. However, the CDS premiums
are significant in the third model but underperform the base model in the predictive ability. The model
does not predict any crises at a probability threshold of more than 0.1. At a threshold of 0.05, the third
model correctly predicts only 4.76% of the crises while the base model predicts 16.13%. At best, the
base model correctly predicts 42.46% of the crises at 0.1 and 0.15 probability thresholds.
Table 6. Forecasts of crises probabilities in currency market models3
|
Number of crises |
Crises Predicted |
Share of correctly | |
|
Currency Market |
predicted |
Correctly False Alarms Pr( D +) Pr(~D∣ +) |
Sensitivity Specificity classified obs. pr( +∣ D) pr( -∣~D) |
Model (1)
Threshhold
|
0.05 |
62 |
16.13% |
83.87% |
47.62% |
95.62% |
94.79% |
|
0.1 |
7 |
42.86% |
57.14% |
14.29% |
99.66% |
98.18% |
|
0.15 |
7 |
42.86% |
57.14% |
14.29% |
99.66% |
98.18% |
|
0.25 |
5 |
40.00% |
60.00% |
9.52% |
99.75% |
98.18% |
|
Model (2) | |||||||
|
Threshhold |
0.05 |
62 |
16.13% |
83.87% |
47.62% |
95.62% |
94.79% |
|
0.1 |
7 |
42.86% |
57.14% |
14.29% |
99.66% |
98.18% | |
|
0.15 |
7 |
42.86% |
57.14% |
14.29% |
99.66% |
98.18% | |
|
0.25 |
5 |
40.00% |
60.00% |
9.52% |
99.75% |
98.18% | |
|
Model (3) | |||||||
|
Threshhold |
0.05 |
21 |
4.76% |
95.24% |
4.17% |
98.66% |
97.17% |
|
0.1 |
1 |
0.00% |
100.00% |
0.00% |
99.93% |
98.36% | |
|
0.15 |
0 |
0.00% |
0.00% |
0.00% |
100.00% |
98.42% | |
|
0.25 |
0 |
0.00% |
0.00% |
0.00% |
100.00% |
98.42% | |
Table 7. Forecasts of crises probabilities in stock market models*
|
Stock Market |
Number of crises |
Crises Predicted _______________Pr( D +) |
False Alarms _____________Pr(~D∣ +) |
Sensitivity Pr( +∣ D) |
Specificity Pr( -∣ ~D) |
Share of correctly |
|
Model (1) Threshhold 0.05 |
112 |
12.50% |
87.50% |
37.84% |
93.40% |
92.04% |
|
0.1 |
19 |
21.05% |
78.95% |
10.81% |
98.99% |
96.84% |
|
0.15 |
6 |
16.67% |
83.33% |
2.70% |
99.66% |
97.30% |
|
0.25 |
1 |
100.00% |
0.00% |
2.70% |
100.00% |
97.63% |
|
Model (2) Threshhold 0.05 |
93 |
11.83% |
88.17% |
29.73% |
94.47% |
92.90% |
|
0.1 |
22 |
27.27% |
72.73% |
16.22% |
98.92% |
96.91% |
|
0.15 |
9 |
33.33% |
66.67% |
8.11% |
99.60% |
97.37% |
|
0.25 |
5 |
40.00% |
60.00% |
5.41% |
99.80% |
97.50% |
|
Model (3) Threshhold 0.05 |
23 |
8.70% |
91.30% |
5.41% |
98.61% |
96.37% |
|
0.1 |
8 |
25.00% |
75.00% |
5.41% |
99.60% |
97.34% |
|
0.15 |
7 |
28.57% |
71.43% |
5.41% |
99.67% |
97.41% |
|
0.25 |
5 |
40.00% |
60.00% |
5.41% |
99.80% |
97.54% |
3* Crises predicted correctly were classified when there was a prediction that a crisis will happen and the crisis in
fact happened. False alarms were considered when a crisis was predicted but it did not take place in reality. Sensitivity
measures the percent probability of a crisis to have been predicted when a crisis happens. Specificity measures the
probability that no crisis is predicted when no crisis is taking place. Share of correctly classified observations measures the
percentage of correctly predicted situations in the market out of all the data points used. D stands for “crises happening”
while ~D is the opposite. + stands for “predicted crisis” while - stands for “no crisis predicted”.
134