Results for the intervention reaction function

4. Methodology and Empirical Results

4.2 Empirical Results

4.2.2 Results for the intervention reaction function

Table 3 presents the estimates of the reaction functions for our sample countries. It can be seen from the negative and statistically significant coefficient of MOVAG that central banks in India, Malaysia, Philippines, Singapore, Taiwan, and Thailand tend to intervene in the markets when their own currencies appreciate against the US dollar. In addition, exchange rates in India, Philippines, and Taiwan are more volatile as shown in the ten-day rolling standard deviation, which also increases the probability of interventions. It is interesting to note that only the coefficient of ∆SAM_t for Malaysia, Singapore, and Taiwan is statistically significant at the conventional significance level, which suggests that there are more intervention reports for these countries when their currencies have higher exchange rate changes (∆SAM_t).

‧

國

立政治大學

‧

N a tio na

l C h engchi U ni ve rs it y

Table 3 Results of Asian Countries’ Intervention Reaction Functions Using Firm News Category

We estimate the intervention function using Heckman’ s (1979) two-step approach:

𝑍𝑡∗= â0+ MOVAGt× â1+ SIGt× â2+ 𝑡 (9)

𝑦𝑡= 𝑔̂0+ ∆𝑆𝐴𝑀𝑡× 𝑔̂1+ 𝜀𝑡 (10)

Equation (9) is a probit model for the decision to intervene. MOVAG is the ten-day moving average of the 9:00 a.m.exchange rates. SIG is the ten-day rolling standard deviation of the 9:00 a.m.exchange rates in equation (10) that the number of intervention news. ∆𝑆𝐴𝑀_𝑡 is the opening exchange rate of day t minus the opening exchange rate of day t-1. The sample period is from January 2005 to April 2011. *,**, and ***

denotes significance at the 90%, 95%, and 99% confidence levels, respectively.

Panel A: India

SIG -20.31 -1.37

Panel G: Taiwan

Coefficient t-statistic

Equation (9)

MOVAG -0.11 -3.08 ***

SIG 2.36 4.22 ***

CONS 2.17 1.87 *

Equation (10)

ΔSAM -0.59 -2.15 **

CONS 1.25 7.61 ***

Log likelihood -683.55

Panel H: Thailand

Coefficient t-statistic

Equation (9)

MOVAG -0.16 -6.83 ***

SIG -1.13 -2.45 **

CONS 3.88 5.08 ***

Equation (10)

ΔSAM 0.46 0.72

CONS 1.66 4.23 ***

Log likelihood

‧

-413.92

國

立政治大學

‧

N a tio na

l C h engchi U ni ve rs it y

4.2.3 Factors influencing the odds of successful intervention

Having estimated the intervention reaction function, we obtain NEWSHAT to proxy for the number of intervention reports and which adjusts for sample selection bias. Table 4 presents the odds of successful interventions, according to the broad criterion 1. It can be seen that the odds of successful interventions according to this criterion increase significantly with the number of firm intervention reports in India, Indonesia, Malaysia, Singapore, South Korea, Taiwan, and Thailand. However, the coefficient estimates of NEWSHAT for both Indonesia and South Korea are virtually zero to make any economic significance. Based on the estimated coefficient, the increase in the odds of successful intervention from having no firm intervention report to having one firm intervention report is about 45% for India, 24% for Malaysia, 46% for Singapore, 49% for Taiwan, and 47% for Thailand, all things remaining unchanged. The probability of having no firm intervention report is 0.5, ceteris paribus. Consequently, having a firm intervention report significantly improves the likelihood of intervention success in five of the eight Asian countries. In the Philippines case, the negative coefficient of -38.039 suggests that the odds of successful interventions decrease by about 50% when the number of firm intervention reports increases from zero to one. Philippines is the only country that exhibits the opposite and perverse effect from having a firm intervention report on the probability of intervention success.

The odds of successful interventions increase significantly with the number of coordinated interventions (i.e. COUN) and with the first intervention day in the past five business days (i.e. FIRS) for all the countries. When no country is jointly intervening on the same day, the probability of successful intervention is 0.5. When another country jointly intervenes on the same day, the probability of successful intervention goes up by 31% for South Korea (the highest increase in the odds) and 23% for Taiwan (the lowest

‧

國

立政治大學

‧

N a tio na

l C h engchi U ni ve rs it y

increase in the odds). When two countries jointly intervene on the same day, the increase in probability from that of having just one country jointly intervene is about 14% for all countries. The dummy for the first intervention day in the past five business days displays significant explanatory power and has a dominant influence on the probability of successful intervention. The probability increases by as much as about 48% in all countries except South Korea and Taiwan. The latter two countries observe an increase in their probability of successful intervention on the first intervention day by about 45%.

This result suggests that most of the information is critical for the shaping market perception about exchange rates in Asian countries is prevalent in the first of five business days of interventions.

When an official statement supporting interventions has appeared in Reuters’ news reports or a neutral report is made, six of the eight countries show statistically significant results. Although the results suggest that both types of news increase the odds of successful intervention, the increase in the odds of success is higher for a supporting intervention statement than for a neutral statement.

‧

國

立政治大學

‧

N a tio na

l C h engchi U ni ve rs it y

Table 4 The Effects of Individual Variables on the Odds Ratio Using Firm News Category

NEWSHAT is the number of reports after adjusting for sample selection bias. COUN is the number of countries jointly intervening on the same day. FIRS is a dummy variable that equals one when a firm report on intervention on day t is the first intervention day in the past five business days. SUPP is a dummy variable that equals one when there is an official statement supporting interventions in Reuters’ news reports. NEUT is a dummy variable that equals one when there is an official announcement of neutral reports. Announcement days refer to announcements on the repurchase rate (RR), interest rate (IR), interbank overnight rate (IOR), gross domestic product (GDP), consumer price index (CPI), balance of trade (BOT), and current account balance (CAB). The LR test is the likelihood ratio test for the null that all the coefficients of the covariates other than the intercept are equal to zero. The sample period is from

Panel C: Malaysia

Macro Announcements

Table 5 provides the results for the joint significance of the explanatory variables.

The results are presented according to a number of benchmark covariates. Taking the case of India as an example, the first benchmark covariate is FIRS, as this variable is not only statistically significant, but it also has the highest magnitude for its coefficient than the other variables in Table 4. In each benchmark regression, additional variables are added one at a time to the benchmark covariate(s) and model adequacy is evaluated using the likelihood ratio test. LR test1 reports the test statistic for the null that the coefficients of the covariates other than the intercept are all equal to zero. LR test2 reports the test statistic for the null that coefficients of the additional covariates other than the benchmark covariates are equal to zero. The purpose of LR test2 is to determine the joint significance of the additional variables other than the benchmark covariate(s). For this reason, LR test2 provides the basis of the subsequent benchmark covariates.

Following on from the example of India, the second set of benchmark covariates is determined by the results of the first set of benchmark regressions with FIRS as the sole benchmark covariate. Here, COUN displays the highest likelihood ratio test value from LR test2 (i.e. 78.66) in Table 5 compared to the other variables, and hence the second benchmark covariates are FIRS and COUN. The third benchmark covariates are determined using the LR test2 results of the second set of benchmark regressions.

NEWSHAT clearly has the highest test statistic value (of 18.81) for LR test2, and FIRS, COUN, and NEWSHAT form the third set of benchmark covariates. Following this process of inference, the last benchmark covariates for India are FIRS, COUN, NEWSHAT, and SUPP. We apply the same process of inference to set up the benchmark covariates to the other countries in our sample.

Referring to the benchmark covariates (2) to (4) for India, whenever the variable NEUT is added to the regressors, the test statistic of LR test2 fails to become statistically

‧

國

立政治大學

‧

N a tio na

l C h engchi U ni ve rs it y

significant, implying that neutral reports of intervention by Reuters do not have any explanatory power on the odds ratio. It can be seen that adding NEWSHAT, COUN, and SUPP to FIRS significantly improves the models’ fit for India and South Korea. On the other hand, adding COUN and NEWSHAT to FIRS significantly improves the models’ fit for Indonesia, Malaysia, and Singapore. This result suggests that both neutral and supporting reports of intervention by Reuters do not have any impact on the odds ratio.

Only FIRS and COUN are jointly significant at the 99% confidence level for Philippines.

As for Taiwan, adding COUN, FIRS, and SUPP to NEWSHAT significantly improves the models’ fit. Finally, Thailand is the only country for which both neutral and supporting reports of intervention by Reuters, in addition to FIRS and COUN, are jointly significant at the 99% confidence level for improving the model’s fit. In sum, two determinants are commonly shared by all countries in terms of their predictive power in explaining the odds of intervention success: the coordinated intervention and the first day of five days of interventions.³

3 Results based on firm and suspected news categories in the sample are not reported here for brevity.

These results remain qualitatively unchanged and they are available from the authors upon request.

‧

國

立政治大學

‧

N a tio na

l C h engchi U ni ve rs it y

Table 5 Joint Significance of Variables on the Odds Ratio Using Firm News Category

We use the logit regression to predict the success of central bank intervention:

𝐿̂𝑖= ln (_1−𝑃̂^𝑃̂^𝑖

𝑖) = 𝑋𝑖𝑏̂ + 𝑢𝑖, where _X_i is a (1× K) vector of variables, which include NEWSHAT, COUN, FIRS, SUPP, and NEUT. NEWSHAT is the number of reports after adjusting for sample selection bias.

COUN is the number of countries jointly intervening on the same day. FIRS is a dummy variable that equals one when a firm report on intervention on day t is the first intervention day in the past five business days. SUPP is a dummy variable that equals one when there is an official statement supporting interventions in Reuters’ news reports. NEUT is a dummy variable that equals one when there is an official announcement of neutral reports. The LR test is the likelihood ratio test for the null that all the coefficients of the covariates other than the intercept are all equal to zero. LR test1 is the likelihood ratio test statistic that compares the log likelihood function with only an intercept in the covariate to that of a regression that includes a constant and covariates. The term ‘plus’ implies that the variable is added to the existing list of regressors in the logit regression. LR test2 is the likelihood ratio test statistic for the null that the benchmark regression contains the covariates in the same row as ‘–’. The sample period is from January 2005 to April 2011. *,**, and *** denote significance at the 90%, 95%, and 99% confidence levels,

Panel B: Indonesia

plus COUN -18.06 95.82 *** 10.72 ***

Panel H: Thailand

Log Likelihood LR test1 LR test2

(1) FIRS -197.51 125.73 *** -

plus NEWSHAT -195.21 130.32 *** 4.60 **

plus COUN -152.53 215.68 *** 89.96 ***

plus SUPP -190.57 139.59 *** 13.87 ***

plus NEUT -190.71 139.33 *** 13.60 ***

(2) FIRS, COUN -152.53 215.68 *** -

plus NEWSHAT -152.42 215.91 *** 0.23

plus SUPP -149.36 222.03 *** 6.34 **

plus NEUT -146.35 228.04 *** 12.36 ***

(3) FIRS, COUN, NEUT -146.35 228.04 *** -

plus NEWSHAT -146.11 228.52 *** 0.48

plus SUPP -142.98 234.79 *** 6.75 ***

(4) FIRS, COUN, NEUT, SUPP -142.98 234.79 *** -

plus NEWSHAT -142.67 235.41 *** 0.62

‧

國

立政治大學

‧

N a tio na

l C h engchi U ni ve rs it y

5. Robustness check

A natural question that arises is whether the results are robust to different categories of intervention news reports. That is, do other news categories (suspected, supports and neutral) have different effects than firm intervention reports? To address this issue, we use firm and suspected news reports and provide results in Table 6. It is seen that results are qualitatively unchanged based on these intervention reports.

We can reject the null hypothesis at the 95% confidence level in favor of a negative forecast value (pl + p2 < 1) for purchase of US dollar intervention in Indonesia and Taiwan, and sales of US dollar intervention in Taiwan using the broad success criterion 1.

In terms of criterion 2, we show that the null hypothesis can be rejected at the 95%

confidence level in favor of a negative forecast value for purchase of US dollar intervention in Indonesia, Malaysia, Philippines, South Korea, Taiwan, and Thailand, and for sales of US dollar intervention in India, Indonesia, Philippines, and Taiwan.

Results based on criterion 3 indicate that interventions are more effective in reducing local currency depreciation (appreciation) against the U.S. dollar. We can reject the null hypothesis at the 95% confidence level in favor of a positive forecast value (pl + p2 > 1) in five of the eight countries for purchases of US dollar intervention and six of the

eight countries for sales of US dollar intervention. These results imply that purchases (sales) of foreign exchange intervention lead to a smaller local currency depreciation (appreciation) against the US dollar, which is consistent with the notion that Asian countries adopted a leaning-against-the-wind intervention policy.

‧

國

立政治大學

‧

N a tio na

l C h engchi U ni ve rs it y

Table 6 Asian Central Bank Intervention and Alternative Success Criteria (Firm and Suspected News Category)

Figures in columns 2, 3, and 4 are the total number of actual intervention, the number of days of successful exchange rate intervention based on a success criterion, and the percentage of successful intervention, respectively. Figures in columns 5, 6, and 7 are total business days, the number of days of successful exchange rate interventions based on a success criterion, and the percentage of virtual successful intervention, respectively. p1 is the probability that a central bank purchases (or sells) US dollars on day t conditional on the exchange rate conforming to the criterion. p2 is the probability that a central bank does not purchase (or sell) US dollars on day t conditional on the foreign exchange rate not conforming to the criterion. p1+ p2 denotes the prediction value for intervention. The final column is the value of the test statistic given by one minus the cumulative density function (CDF). ***, ** and * denote significance at the 99%, 95% and 90% confidence levels, respectively.

Panel A: India

Intervention Virtual

p1 p2 p1+p2 1-CDF Total Successes Total Successes