4.1. Result of MS-Regression
We conducted MS-Regression modeling on the observation values and change rates of four macroeconomic indicators: GDP, CPI, interest rates, and unemployment rates. The model incorporated a trend term; moreover, the regimes can be fitted under the assumption of either equal variance or differing variances. We conducted model fitting under both scenarios and selected the model with the higher likelihood. In our case, we set the order to zero, indicating that the current regime’s likelihood does not depend on past regimes. The results are discussed as follows:
As illustrated in
Figure A1, U.S. GDP has shown a predominantly upward trend over the past 50 years; modeling the observation values of GDP led to a poor fit: the model classified the first 25 years as a recession period and the subsequent 25 years as an expansion period due to the absolute values of GDP being higher in the latter period. This approach was deemed unsuitable. Therefore, we modeled the change rates of GDP using MS-Regression, and the results, the result of the smoothed marginal probabilities for Regime 1 and Regime 2, is shown in
Figure A3. It indicates that the economic crises during the 2008 subprime mortgage crisis and the 2020 COVID-19 pandemic were correctly identified as Regime 2. However, the model classified all periods before 1985 as Regime 2, while most of the periods after 1985 (except for the crises) were classified as Regime 1. This pattern suggests that the MS-Regression model did not provide a satisfactory fit for GDP.
For CPI, modeling the change rates yielded poor results: the model identified only six data points as Regime 2, with all other data points classified as Regime 1. Consequently, we modeled the observation values of CPI using MS-Regression. The result of the smoothed marginal probabilities for Regime 1 and Regime 2 is presented in
Figure A4. The model successfully identified the economic crisis of 2020 as Regime 2 but classified all periods before 1985 as Regime 2 and most periods after 1985 as Regime 1, and the model cannot identify the crisis of 2008. This suggests that the MS-Regression model also performed poorly for CPI.
When modeling the change rates of U.S. interest rates, the results were similar to those of CPI’s change rates. The model identified only 13 data points as Regime 2, with all other data points classified as Regime 1, leading to an unsatisfactory fit. Modeling the observation values of interest rates using MS-Regression produced results similar to those of GDP and CPI. The result of the smoothed marginal probabilities for Regime 1 and Regime 2 is shown in
Figure A5; the model correctly identified the economic crises during 2008 and 2020 as Regime 2 (albeit with some delay for the COVID-19 crisis, extending the period to 2024). However, the model classified all periods before 1999 as Regime 2 and most periods after 1999 as Regime 1. This outcome indicates that the MS-Regression model also performed poorly for interest rates.
For unemployment rates, modeling the change rates yielded ambiguous results: the probabilities of Regime 1 and Regime 2 hovered around 50% for most time points, making regime classification unclear and the model unfit for further analysis. In contrast, modeling the observation values of unemployment rates using MS-Regression produced significantly better results. The result of the smoothed marginal probabilities for Regime 1 and Regime 2 is shown in
Figure 1. This approach not only effectively identified the 2008 subprime mortgage crisis and the 2020 COVID-19 pandemic as Regime 2 but also identified three periods between 1973 and 1998 as Regime 2. The alternation between expansion periods and recession periods aligns better with realistic economic cycles. Additionally, unlike other macroeconomic indicators, the unemployment rate results do not suffer from overly short Regime 2 periods, making it more suitable for further research. Therefore, we decided to define economic regimes based on unemployment rates in subsequent analyses.
In the MS-Regression model for unemployment rate observation value, The log-likelihood of the model is −320.104. The model’s information criteria, including the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and Hannan-Quinn Information Criterion (HQIC), are 650.209, 666.872, and 656.947, respectively. During the modeling process, we found that the results assuming equal variance yielded a higher likelihood compared to those with differing variances. Therefore, the two regimes exhibit the same level of volatility. The parameters are as follows:
Regime Transition Probabilities:
The empirical findings from our model indicate that, during Regime 1, the mean of the unemployment rate model is relatively low, reflecting an economic expansion phase in the U.S. economy. In contrast, during Regime 2, the mean of the unemployment rate model is relatively high, suggesting signs of an economic recession in this phase.
4.2. Result of Quantile-VAR
In our study, we consider the expansion periods (recession periods) across different time frames as continuous time series. Therefore, we integrate the expansion periods (recession periods) from each period into a single time series, resulting in two distinct time series: one for expansion periods and another for recession periods.
Based on the Quantile-VAR model, we set the quantiles at 0.05, 0.5, and 0.95 to model these periods separately. These three models serve to distinguish between the extreme return spillovers associated with negative shocks, average normal conditions, and positive shocks. Consequently, this approach allows us to estimate the correlations under the extreme lower tail, median conditions, and the extreme upper tail. During the estimation process, we select the optimal parameters according to the Schwarz Criterion (SC) for our models. We set the forecast horizon .
Figure A6,
Figure A7 and
Figure A8 represent the impulse response functions for the entire dataset, expansion periods, and recession periods, respectively.
Table 1,
Table 2 and
Table 3 present the return spillovers in the quantile VAR for the entire dataset, while
Table 4,
Table 5,
Table 6,
Table 7,
Table 8 and
Table 9 provide the return spillovers in the quantile VAR for the expansion periods and recession periods, respectively.
The Impulse Response Function allows for the examination of the individual spillover effects between various currencies at different points in time. However, to observe the impact of each individual currency on all other currencies or to assess the overall effect of all currencies on a single currency, our analysis will focus on the Connectedness Index from
Table 1,
Table 2,
Table 3,
Table 4,
Table 5,
Table 6,
Table 7,
Table 8 and
Table 9. The results indicate that, regardless of whether the full dataset, expansion periods, or recession periods are considered, the TCI is around 70% when the quantile is set at 0.05 or 0.95, and it is approximately 50% when the quantile is set at 0.5. These findings underscore the greater influence of extreme shocks on the system of return spillovers. Notably, the contributions to and from other variables in both the extreme lower and extreme upper tails are significantly stronger compared to those at the median. This result aligns with the findings reported in [
1].
Now, we examine the results for the entire dataset. Regardless of the quantile, the impact value of the US dollar on other currencies remains stable around 60. For other currencies, the influence from other currencies is relatively smaller when the quantile is 0.5; however, this influence increases significantly under extreme lower and upper tail conditions. On the other hand, under extreme lower and upper tail conditions, the US dollar’s impact on other currencies is relatively smaller, with the “contribution to others” values being either relatively small or the smallest. In contrast, under median conditions, the US dollar’s impact on other currencies is relatively larger, reaching 80. This suggests that under extreme conditions, there is a stronger mutual influence among various currencies, whereas in normal conditions, the US dollar is the primary factor influencing other currencies. An additional interesting point is that the Australian dollar has a minimal impact on other currencies under normal conditions, with a value of only 21.24.
Next, we observe the results during the expansion period. We find that under extreme lower and upper tail conditions, the US dollar is less influenced by other currencies, whereas under median conditions, the US dollar is most affected by other currencies. Under median and extreme upper tail conditions, the US dollar has the greatest impact on other currencies, but under extreme lower tail conditions, its impact is the smallest. Similarly, it can be observed that under extreme upper and lower tail conditions, the influence of each country’s currency on others expands, while under median conditions, this influence is smaller, with the Australian dollar having the lowest impact at 27.85.
Finally, we examine the results during the recession period. Under median conditions, the US dollar’s influence on other currencies and its susceptibility to influence from other currencies remain the highest. Under extreme lower and upper tail conditions, the mutual influence among various currencies significantly increases, and the degree of influence becomes more similar. The Australian dollar, under median conditions, has the lowest impact on other currencies, with a value of only 2.95.
In the next step, we employed a rolling window analysis to investigate the time-varying characteristics of return spillovers among five currencies. This analysis covers three datasets: the entire dataset, the expansion period, and the recession period, examining the conditions at the median as well as the extreme lower and extreme upper tails. A window size of 20 periods was selected for this analysis.
Figure A9 illustrates the changes in TCI for the entire dataset,
Figure A10 shows the changes in TCI during the expansion period, and
Figure A11 presents the changes in TCI during the recession period.
We found that, compared to the TCI at the median, which exhibits significant fluctuations, the TCI at the extreme lower and extreme upper tails is much higher but with a smaller range of variation. Although the variations at the extreme lower and extreme upper tails are somewhat similar, they do not occur simultaneously when showing larger or smaller values, regardless of whether the entire dataset, the expansion period, or the recession period is considered. Thus, initially, we observed that the TCI at the extreme lower and extreme upper tails shows higher values in any period, but this is not symmetrical. The differences in their behavior can be seen from the TCI graphs, indicating asymmetry between the extreme lower and extreme upper tails.