Uncertainty in an emerging market economy: evidence from Thailand

In large part, this is due to the lack of long samples and reliable data for these countries. Uncertainty indicators proposed in the literature are also only mainly available for the US and few other developed economies. See www.​policyuncertaint​y.​com.

This is a distinction that earlier measures of uncertainty such as survey-based ones often fail to address. First moment shocks can be thought of as a deterioration in the expected outcome which is not uncertainty, just bad news. Second moment shocks on the other hand are uncertainty and are defined as a greater range of expected outcomes. Disentangling the two can be difficult, especially since market participants tend to become more pessimistic in the face of greater uncertainty.

The LDA algorithm involves two main steps. First, for each newspaper article, the process starts by randomly choosing a distribution over topics. In the second step, each word in each document is then drawn from one of the topics, where the selected topic is chosen from the per-document distribution over all topics. In this way, all documents will always share the same number of topics but with different proportions. Readers are referred to Blei (2012) for more details on the algorithm.

These are (i) Bangkok Biz News, (ii) Daily News, (iii) Matichon and (iv) Thairath.

They are in Thai language and are available upon request.

Note that although there are more than five topics in the full corpus, since we choose to discard the remaining topics that cover other uncertainty-related issues such as those on education, the environment, healthcare and housing, we scale the total count of uncertainty-related articles in our five selected topics to 100. In this way, we can focus on analyzing how important each of these five topics are in relation to each other. This is also for visual purposes of viewing Fig. 2, since the number of other uncertainty-related topics that we discard are quite large.

In the literature, the PCA has often been employed to gauge the overall level of uncertainty across a swathe of uncertainty proxies that are often available, and has been shown to capture the common movements among the various indicators of uncertainty well (see Haddow et al. 2013; Forbes 2016 and Redl 2017, among others).

That is, our empirical results are robust to other global uncertainty measures that includes applying the PCA to combinations of the BBD US news-based uncertainty, the Euro area news-based uncertainty and the VIX index as well.

Another popular ordering of the VAR in the literature is one of the reverse orders with uncertainty ordered last. We performed the reverse-order analysis as a robustness check, where the interpretation of our findings did not change significantly. Nevertheless, it is still unclear whether uncertainty should be placed before or after the real activity variable. Also, there is a recent strand of the literature that debates whether the recursive VAR identification approach is a relevant approach at all, given that there may be reverse caUSlity between uncertainty and real activity. Few studies that address this potential endogeneity of uncertainty propose novel identification procedures (Carriero et al. 2018b; Cesa-Bianchi et al. 2018; Mumtaz 2018; Angelini et al. 2019; Ludvigson et al. forthcoming), but so far have delivered empirical results that are quite mixed.

The majority of VARs in the literature that investigate the effects of uncertainty are in levels except for some that consider growth or HP-filtered variables. According to Sims et al. (1990), VARs in log levels provide consistent estimates of the IRFs even in the presence of co-integrating vectors.

Information criterion tests suggest either VARs with 1 or 2 lags, so we select a VAR with 1 lag due to the large number of endogenous variables in the VAR. The results are also robust to VARs with 2 lags. To ensure no model misspecification, we also perform multivariate Portmanteau tests to ensure no serial correlation in the error terms of the VAR models.

We also obtain the impulse responses of inflation and the policy rate to global and local shocks but we do not wish to discuss them here since the Thai CPI and policy rates have been relatively stable throughout the sample under investigation.

There appears to be some minor evidence of overshooting for the impact of macroeconomic uncertainty shocks on consumption and investment (Fig. 5), but is only marginally significant. For studies that find evidence of overshooting, we notice that they tend to use volatile implied or realized financial market volatility measures as proxies for uncertainty, whereas studies that use alternative proxies for similar countries find no such effect (Jurardo et al. 2015; Cuaresma et al. 2019). Thus, the rebound effect may depend on the type of uncertainty measure used. Alternatively, it may depend upon cross-country differences or the sample period under investigation. Carrière-Swallow and Cèpedes (2013) offer evidence that real activity tends to occur in the medium run for developed economies, while emerging economies do not display a similar pattern. Caggiano et al. (2014) show that if the sample period includes the GFC where most developed central banks switched to unconventional monetary policy measures in the presence of the effective zero lower bound, the overshoot vanishes.

Note, however, that not all types of capital flows may react alike to uncertainty. For example, Hlaing and Kakinaka (2019) shows that for 50 developed and emerging markets, global uncertainty clearly increases the likelihood of contractions in FPI for all countries, while it increases FDI in only advanced economies

We also examine the FEVD results at other horizons, but the findings that they deliver do not qualitatively change our discussion of results. Results are available upon request.

To provide a guide for factor estimation, we use the Bai and Ng (2002) information criterion (IC) to select the number of factors. The IC suggests 3 factors which explains only 21% of the variation in the dataset, where the first three factors load heavily on real activity measures such as retail sales and the manufacturing production index, the SET index and return on its components, and government bond rates, respectively. Since the variation explained by the three factors are rather low we also consider extracting 18 factors which can explain at least half of the variation of series in the dataset. However, we find whether using 3 or 18 factors provides aggregate uncertainty measures that are not statistically significantly different; thus, we use 3 factors in our empirical investigation.

Other weighting schemes are also possible such as by employing the principal component analysis (PCA) approach. We follow JLN and construct these measures as part of our robustness checks and find that final indices do not differ significantly.

Similar to JLN, we find that the response of economic variables to macroeconomic and financial uncertainty of various horizons do not differ in a significant way.

We follow Wallach et al. (2009) and calculate the perplexity value as \(perplexity(W) = exp\{ \frac{\sum log P(w_d | \Phi , \alpha )}{\sum N_d}\} \), where W is the test set which contains a random selection of articles \(w_d\) which amount to 10% of all articles in the full dataset. For each article \(w_d\), the probability \(P(w_d | \Phi , \alpha )\) is computed as \(P(w_d | \Phi , \alpha ) = \int d\theta P (w_d | \theta ,\Phi ) P(\theta |\alpha )\), in which the integral is approximated via the iterated pseudo-count method. \(\sum N_d\) is the total number of words in the whole set and parameters \(\Phi \) and \(\alpha \) are learned from the training set.