Global financial cycle, media coverage and currencies of emerging markets | Macroeconomic Dynamics
3.1 Data
A balanced panel of 17 EMs, according to the International Monetary Fund (IMF) groupingFootnote
4
is examined with monthly data from January 1998 to April 2019.
We obtain country-specific sentiments and buzz data without inherent biases from Thomson Reuters Refinitiv MarketPsych Indices (TRMI). Michaelides et al. (Reference Michaelides, Milidonis and Nishiotis2019) use the same dataset as public information in the currency market. The TRMI employs lexical analysis to extract sentiment indices in real-time by reckoning several sources comprising conventional newspaper articles and social media platforms, such as Twitter, every minute, totaling over two million news articles and posts daily. MarketPsych has been given emerging attention in recent research since it generates a wider news and social media sentiment index by converting these volumes and professional news into manageable information flows that are capable of influencing decisions. The TRMI scores the news and social media content as sentiment indices, which are described as overall positive references, net of negative references. This is a bipolar ranging from −1 to 1, such that a positive score value in the bipolar is termed a positive opinion and a negative value shows a negative opinion and 0 refers to neutrality. The buzz is the number of words used to compute the sentiment index, representing the intensity of media attention to the economy.
We generate a global sentiment factor from 95 countries’ sentimentsFootnote
5
by applying a DFM as shown in the equation below (see Appendix A for details of this model):
\begin{equation} x_{t}^{i} = \alpha _{{global}}^{x,i}\, f_{t}^{x, {global}} + e_{t}^{x,i} \end{equation}
where
$x_{t}^{i}$
includes the sentiment for the respective 95 countries at month
$t$
.
$f_{t}^{x, {global}}$
is the global common factors for the variables. This helps to integrate the impact of the co-movement factor among the global sentiments.
Although there have been several measures of the global financial cycle, we rely on the global factor estimated by Miranda-Agrippino and Ricco (Reference Miranda-Agrippino and Ricco2015), which has received more attention in the literature. This is also utilized in Miranda-Agrippino and Rey (Reference Miranda-Agrippino and Rey2020). The global factor explains a significant share of the common variation in a large cross-section of risky asset prices globally. Miranda-Agrippino and Ricco estimate the global factor using aDFM at monthly frequency data for a larger and an unbalanced (heterogeneous) panel of risky asset prices traded globally.Footnote
6
They use the log-returns of the asset prices.Footnote
7
These asset prices comprise equity prices covering Europe, North and Latin America, Asia Pacific, and Australia; commodity prices excluding precious metals; and corporate bond indices. This is then extended by Miranda-Agrippino et al. (Reference Miranda-Agrippino, Nenova and Rey2020) with two dimensions: time by estimating the data to cover up to April 2019 and cross-section, through including a larger and richer set of price series that is updated to reflect the relevant adjustments in global markets, especially via the inclusion of Chinese stocks. One of the advantages of using the global factor is that it accounts for major changes in global risky asset prices.Footnote
8
Furthermore, the impact of the global factor of risky assets could be influenced by global economic fundamentals, including interest rate, inflation, economic conditions indicator, consumer and business survey sentiments, investor sentiment, and unemployment rate. As with generating the global sentiment factor shown in Equation (1), we apply the DFM presented in Equation (2) to obtain common factors for the respective global economic fundamentals (see Table A2 in Appendix A for details on the data set of the global economic fundamental indicators).
\begin{equation} y_{t}^{i} = \alpha _{{global}}^{y,i}\, f_{t}^{y, {global}} + v_{t}^{y,i} \end{equation}
where
$y_{t}^{i}$
denotes the global economic fundamentals including the interest rate series, consumer and business survey sentiment and unemployment rates at month
$t$
.
$f_{t}^{y, {global}}$
is the global common factors for the economic fundamentals.
We filter out these global economic fundamental factors from the global factor proposed by Miranda-Agrippino and Ricco to generate a “pure” global factor shock. Sometimes we refer to Miranda-Agrippino and Ricco’s global factor as a “general global factor” to ensure a distinction from the “pure global factor shock.”
Rey (Reference Rey2015) identify VIX to co-move with the global financial cycle. To have a comparative analysis, we employ VIX as a proxy for the global financial cycle. VIX is the Chicago Board Options Exchange (CBOE) Volatility Index that measures market expectations with respect to volatility represented by the Standard & Poor 500 index option prices. This is usually used as a global financial market risk in the literature. In this case, we employ the direct measure of the global financial cycle by Miranda-Agrippino and Ricco and the VIX as an uncertainty measure. We retrieved VIX data from FRED, Federal Reserve Bank of St. Louis.Footnote
9
The target variable in our empirical analysis is the EM currencies, which, for robustness checks, we rely on both the nominal bilateral exchange rates against the USD and the effective exchange rates (broad indices).Footnote
10
Next, we measure the value of the currencies as the average log-returns of the exchange rates from the first day of the current month to the first day of the subsequent month, such that a decrease in the bilateral exchange rate and an increase in the effective exchange rate mean appreciation of the domestic currency and vice versa. We control for the economy’s macroeconomic indicators, such as the inflation rate and the industrial production index.Footnote
11
This helps factor in business cycle variation in the analysis. To ensure comparable units, we standardized the data variables employed in this study.
3.2 Methodology
We set the background by looking at the EMs and global sentiments and the global financial cycle shock transmission on the EMs’ currencies. To achieve this, we employ the local projection introduced by Jordà (Reference Jordà2005), which is less sensitive to misspecification and free from dynamic restrictions, giving us the flexibility to apply it to both linear and nonlinear models. This has been used in several studies, including Jordá and Taylor (Reference Jordà and Taylor2016), Ramey and Zubairy (Reference Ramey and Zubairy2018), etc. We begin our empirical analysis with the baseline specification by generating the impulse response for the linear local projection.
\begin{align} FX_{j,t+h}^{i,Comp} = \alpha _{h} + \lambda _{j,h} + \sum _{k=1}^r \beta _{k,h}^{Comp}Sent_{j,t-k} + \sum _{k=1}^r \phi _{k,h}^{Comp}GSent_{t-k} + \sum _{k=1}^r \gamma _{k,h}^{Comp}GFact_{t-k} \nonumber \\ +\sum _{k=1}^r \chi _{k,h}^{Comp}FX_{j,t-k}^{i,Comp} + \sum _{k=1}^r \mu _{k,h}^{Comp}INF_{j,t-k} + \sum _{k=1}^r\delta _{k,h}^{Comp}IPI_{j,t-k} + \epsilon _{j,t+h}^{Comp}, \end{align}
where
$FX_{j,t+h}^{i,Comp}$
is the cumulative log-returns of the bilateral or effective exchange rate for the country
$j$
. Our local projection coefficients of interest include
$\beta _{k,h}^{Comp}, \phi _{k,h}^{Comp}$
and
$\gamma _{k,h}^{Comp}$
, giving the cumulative effect of the country-specific, that is, domestic sentiment (
$Sent_{j,t}$
), global sentiments (
$GSent_{t}$
) and the global financial cycle variables (
$GFact_{t}$
), respectively, from time
$t$
to
$t+h$
for each element of the log-returns of the exchange rates. As stated earlier, shock transmission may be affected by economic fundamentals, especially the variation of the business cycle, to some extent. EMs with a suitable inflation and production growth track record influence their capacity to issue more securities in local currency (Burger and Warnock, Reference Burger and Warnock2007; Koepke, Reference Koepke2019). In this vein, we control for key domestic macroeconomic variables such as inflation (
$INF_{j,t}$
) and industrial production (
$IPI_{j,t}$
).Footnote
12
Again, we include the lag of the exchange rates as control variables to develop auto-correlation, which accounts for the persistence in the realized monthly exchange rate. Additionally, we account for country-fixed effects,
$\lambda _{j,h}$
.
$r$
is the lag of the variables employed in the system. We use 6-month lags; however, changing the number of lags does not significantly change the results. We perform a
$h$
steps ahead forecast horizon, thus in this study we estimate the cumulative response of the exchange rates with robust standard errors up to 20 months ahead. The cumulative response provides a total effect of the exchange rate movement, thus capturing both temporary and long-run effects over time, making the study results relevant for policy implications. To ensure robustness of the standard error, we apply the Variance-Covariance matrix estimator, Spatial and Serial Correlation Consistent (vcovSCC). This helps to address the serially correlated, cross-sectional dependence and heteroskedasticity associated with the errors.
This baseline linear model discussed so far assumes that both positive and negative sentiments play the same role in the exchange rate response to shocks in the global sentiments and global financial cycle. Equation (3) tests three hypotheses: (i) both domestic and global sentiments cause both bilateral and effective exchange rates to appreciate such that higher sentiments appreciate the domestic currency and vice versa. (ii) An increase in the global risky asset market returns (global factor or pure global factor shock) causes the emerging economies’ currencies to appreciate in the short run due to the demand for the domestic currency to invest in EMs, which corresponds to the study by Ranaldo and Söderlind (Reference Ranaldo and Söderlind2010). (iii) The VIX shock leads to the depreciation of the currencies of emerging economies.
Investors may choose not only to invest in risky EM economies based solely on the global financial cycle, such as the returns on risky assets. They also perceive the value of risky assets by considering media sentiment or perceptions regarding specific EMs or the global economy to determine whether the investment is good or bad (Shiller, Reference Shiller1990). This demonstrates how EMs and global sentiments interact with the global financial cycle.
Positive vs. Negative News: The size or sign of sentiments and global financial cycle effects on the currency is expected to be influenced differently during positive and negative sentiment states of EM economies. Therefore, by analogy, we create a threshold country sentiment for positive and negative as below:
\begin{equation} I_{t} = \begin{cases} 1 &{if}\, S_{t-1} \gt 0 \quad ({Positive}),\\ 0 &{if}\, S_{t-1} \leq 0 \quad ({Negative}). \end{cases} \end{equation}
where
$S_{t-1}$
is the lag of the country sentiment.
$I_{t}$
is a dummy of 1 when it is a positive sentiment and otherwise it becomes a negative sentiment.
High vs. Low Media Intensities: Another important analysis is to examine the interaction between the degree of media intensity, sentiments, the global financial cycle, and the EMs’ exchange rates. This helps to investigate how the level of media attention in the EMs influences sentiments and the impact of the global financial cycle on exchange rates. To do this, we employ the country-specific media buzz. We also contribute to Miranda-Agrippino and Rey (Reference Miranda-Agrippino and Rey2020) by looking into the monetary policies’ buzz, which is termed as rates buzz. The country-specific buzz, as media coverage explains the number of words used to compute the EMs’ sentiment indicators, which directly measures the country-specific media intensity. On the other hand, the monetary policies’ buzz is a sum of all references underlying the central bank, interest rates and their forecast values, debt defaults, and monetary policy loose and tight. We generate a threshold of higher and lower buzz:
\begin{equation} I_{t} = \begin{cases} 1 &{if}\, Buzz_{t-1}^{i} \gt \overline {Buzz}^{i},\\[6pt] 0 &{if}\, Buzz_{t-1}^{i} \leq \overline {Buzz}^{i}, \end{cases} \end{equation}
where
$Buzz_{t-1}^{i}$
measures the media attention given to the EMs at time
$t-1$
and
$I_{t}$
in Equation (5) refer to higher buzz (
$HigherBuzz_{t-1}^{i}$
) where
$i$
denotes the country-specific buzz or monetary policies buzz.
$\overline {Buzz}^{i}$
is the mean of the country-specific buzz or monetary policies’ buzz. A dummy 1 is given when the buzz at time
$t-1$
is greater than the average buzz and we term it a higher buzz regime; otherwise, we denote it as a lower buzz regime. The higher buzz regime explains the higher media intensity, while the lower buzz regime represents the lower media intensity.
In the second stage, we build on the baseline model by extending the local projection estimation with state-of-the-art regime switching as in Auerbach and Gorodnichenko (Reference Auerbach and Gorodnichenko2012), Ahmed and Cassou (Reference Ahmed and Cassou2016), Ramey and Zubairy (Reference Ramey and Zubairy2018), Cloyne et al. (Reference Cloyne, Jordà and Taylor2023), Gonçalves et al. (Reference Gonçalves, Herrera, Kilian and Pesavento2024), Inoue et al. (Reference Inoue, Rossi and Wang2024), etc. This state dependence explains the distinguished nonlinear response of exchange rates to sentiments and global financial cycle shocks when the emerging economy is in a state of positive or negative by incorporating Equation (4) into the local projection. This is represented by the regime-switching model in Equation (6). This specification describes the asymmetric impact of the sentiments, summarizing the different influences of positive and negative sentiment.Footnote
13
\begin{align} FX_{j,t+h}^{i,Comp} &= I_{t}\Biggl [\alpha _{h}^{P} + \lambda _{j,h}^{P} + \sum _{k=1}^r \beta _{k,h}^{P,Comp}Sent_{j,t-k} + \sum _{k=1}^r \phi _{k,h}^{P,Comp}GSent_{t-k} + \sum _{k=1}^r \gamma _{k,h}^{P,Comp}GFact_{t-k} \nonumber \\ & \qquad+\sum _{k=1}^r \chi _{k,h}^{P,Comp}FX_{j,t-k}^{i,Comp} + \sum _{k=1}^r \mu _{k,h}^{P,Comp}INF_{j,t-k} + \sum _{k=1}^r\delta _{k,h}^{P,Comp}IPI_{j,t-k}\Biggr ] \nonumber \\ & \qquad+(1 – I_{t})\Biggl [\alpha _{h}^{N} + \lambda _{j,h}^{N} + \sum _{k=1}^r \beta _{k,h}^{N,Comp}Sent_{j,t-k} + \sum _{k=1}^r \phi _{k,h}^{N,Comp}GSent_{t-k} \nonumber \\ & \qquad + \sum _{k=1}^r \gamma _{k,h}^{N,Comp}GFact_{t-k} +\sum _{k=1}^r \chi _{k,h}^{N,Comp}FX_{j,t-k}^{i,Comp} \nonumber \\ & \qquad+ \sum _{k=1}^r \mu _{k,h}^{N,Comp}INF_{j,t-k} + \sum _{k=1}^r\delta _{k,h}^{N,Comp}IPI_{j,t-k}\Biggr ] + \epsilon _{j,t+h}^{T,Comp}, \end{align}
where
$I_{t}$
denotes the threshold dummy variable.
$P$
and
$N$
added to the coefficients suggest positive or higher intensity and negative or lower intensity, respectively.
$\epsilon _{j,t+h}^{T,Comp}$
represents the error of the asymmetric model. The other notations remain the same as explained above. Equation (6) tests our fourth hypothesis that there is an asymmetrical impact of sentiments and the global financial cycle, which is in line with Laakkonen & Lanne, Reference Laakkonen and Lanne2009) and Akhtar et al. (Reference Akhtar, Faff and Oliver2011).
