Uncertainty in an emerging market economy: evidence from Thailand

Abstract

This paper constructs various measures of domestic and global uncertainty and provides a comprehensive study of their impacts on the Thai economy. Based on a small open economy VAR, global uncertainty delivers deeper and more long-lasting effects when compared to within-country ones. In addition, we find that uncertainty shocks first generate sudden and large declines for stock prices and foreign portfolio investment, before gradually affecting the real economy through investment and trade channels. There is also meaningful heterogeneity among different types of domestic uncertainty. While financial uncertainty matters most for the Thai economy overall, consumption demand largely responds to macroeconomic uncertainty, while economic policy and political uncertainty generates the most persistent effects on investment. Furthermore, fiscal policy uncertainty is a key driver of trade flows while monetary policy uncertainty plays an important role for capital markets.

Appendices

Appendix A: Model-based uncertainty measure

The JLN approach estimates uncertainty from a large number of macroeconomic and financial time series based on a diffusion index and stochastic volatility models. We apply the JLN approach to construct macroeconomic and financial uncertainty indices for Thailand. In doing so, we first let $y^C_{jt}$ be a variable in either the macro or financial category. Its forecast, $E[y^C_{jt+h}|I_t]$ can be estimated from the following factor augmented forecasting model:

$$\begin{aligned} y^C_{jt+1} = \phi _j^y(L) y_{jt} + \gamma _j^F(L){\hat{\mathbf{F}}_t}+ \gamma _j^W(L)\mathbf{W_t }+ v^y_{jt+1} \end{aligned}$$

(1)

where $\phi _j^y(L), \gamma _j^F(L), \gamma _j^W(L)$ are finite-order polynomials. The factors ${\hat{\mathbf{F}}_t}$ are drawn from the information set $I_t$ which is approximated by the full dataset which contains both macroeconomic and financial time series variables.¹⁶$\mathbf{W_t }$ contains additional predictors that are meant to capture possible nonlinearities such as the squares of the first component of ${\hat{\mathbf{F}}_t}$. In the model, the prediction error for $y^C_{jt+1}, {\hat{\mathbf{F}}_t}, \mathbf{W_t }$ are permitted to have time-varying volatility $\sigma ^y_{jt+1}, \sigma ^F_{kt+1}, \sigma ^W_{lt+1}$, respectively, which generates time-varying uncertainty in the overall series $y^C_{jt}$.

From Eq (1), we compute the forecastable component $E[y^C_{jt+h}|I_t]$ which form the basis of our uncertainty measures. More specifically, we calculate the forecast error as $V^{y^C}_{jt+h} = y^C_{t+h} - E[y^C_{jt+h}|I_t]$, where the conditional volatility of this forecast error $E[(V^{y^C}_{jt+h})^2|I_t]$ is then generated based on a parametric stochastic volatility model for the one-step-ahead prediction errors in $y^C_{jt}$ and the factors. Then, using a recursive method, we can estimate $E[(V^{y^C}_{jt+h})^2|I_t]$ for future horizons $h>1$. As discussed in JLN, the stochastic volatility modeling approach allows for shocks to the second moment of a variable to be independent from the first moment, consistent with theoretical models of uncertainty which presumes the existence of an uncertainty shock that independently affects $y_j$.

Finally, uncertainty for the variable $y^C_{jt}$ at horizon h can be computed as:

$$\begin{aligned} U^{y^C}_{jt}(h) \equiv \sqrt{E[(V^{y^C}_{jt+h})^2|I_t]} \end{aligned}$$

(2)

which measures uncertainty as the conditional volatility of the purely unforecastable component of the h-step-ahead realization of each underlying macroeconomic and financial time series based on available information at time t. We follow JLN and assume equal weights $w_j=\frac{1}{N_C}$ to arrive at the aggregate uncertainty measure:¹⁷

$$\begin{aligned} U^{y^C}_{t}(h) \equiv plim_{N_{\rightarrow \infty }} w_j U^{y^C}_{jt}(h). \end{aligned}$$

(3)

Based on Eq. (3), we compute the macroeconomic and financial uncertainty measures by aggregating the conditional variances of the unforecastable components over variables that belong to the either macroeconomic or financial categories. For both measures, we compute uncertainty for the forecasting horizons $h=4$, but also consider various other horizons for robustness checks.¹⁸

The underlying dataset to construct Thai economic uncertainty indices comprises of monthly macroeconomic and financial data obtained from the Bank of Thailand and the Stock Exchange of Thailand databases over the 2002M1–2019M12 sample. We choose to construct the series based on monthly data for a wider information set and then construct quarterly uncertainty indices by taking within-quarter averages. In the full dataset, we have 199 macroeconomic series that represent broad categories that describe the macroeconomy (Groups 1–10) and 22 financial series (Group 11) as listed in Table 5. In the table, each series has one of the following transformation codes which are applied to the data series to ensure stationarity:

Macroeconomic time series transformations

• 1: $X_{it} = X^A_{it}$

• 2: $X_{it} = X^A_{it} - X^A_{it-1}$

• 3: $X_{it} = \Delta ^2 X^A_{it}$

• 4: $X_{it} = ln(X^A_{it})$

• 5: $X_{it} = ln(X^A_{it}) - ln(X^A_{it-1} )$

• 6: $X_{it} = \Delta ^2 lnX^A_{it}$

• 7: $X_{it} = (X^A_{it}-X^A_{it-1})/X^A_{it-1}$

where $X_{it}$ denotes the transformed variable i and $X^A_{it}$ is the actual or raw data series. Note that we use the notation $\Delta = 1-L$ and $LX_{it} = X_{it-1}$.

Financial time series transformations

For the first five financial time series with transformation code 8, we follow the method as described below.

• $D\_log(DIV): \Delta logD^*_t$

• $D\_log(P): \Delta logP_t$

• $D\_DIVreinvest: \Delta logD^{re,*}_t$

• $D\_Preinvest: \Delta logP^{re,*}_t$

• d-p: $log(D^*_t)-log(P_t)$

Note that to obtain the dividend and price series, $(D^*_t$ and $P_t)$, we first construct the return series with dividends ($RETD_t$) and excluding dividends ($RETX_t$) as: $RETD_t = \frac{P_{t=1}+D_{t+1}}{P_t}$ and $RETX_t=\frac{P_{t+1}}{P_t}$, and produce a normalized price series based on the recursive rule: $P_0=1, P_t=P_{t-1}RETX_t$. A dividend series can then be constructed as: $D_t =P_{t-1}(RETD_t - RETX_t)$ where $D^*_t = (D_t+D_{t-1} + D_{t-2} + D_{t-3})$.

For dividends and prices under reinvestment, $(D^{re*}_t$ and $P^{re*}_t)$, we use the recursion $P^{re}_0 =1, P^{re}_t = P_{t-1} RETD_t$. Then, dividends under reinvestment can be defined as $D_t^{re} = P^{re}_{t-1}(RETD_t - RETX_t)$ where as before, $D^{re*}_t = (D_t^{re} + D^{re}_{t-1} + D^{re}_{t-2} + D^{re}_{t-3})$.

Finally, for the remaining 17 financial time series which are industry portfolios, the portfolio returns are constructed from the price and dividend yield series as follows:

$$\begin{aligned} R_{it} = \frac{P_{t+1}+D_{t+1}}{P_t}. \end{aligned}$$

Table 5 List of macroeconomic and financial variables

Appendix B: News-based uncertainty measure

To construct topic-based uncertainty measures, we adopt Azqueta-Gavaldon (2017) by employing the latent Dirichlet allocation (LDA) method, the most popular topic-modeling approach that has been developed by Blei et al. (2003) to help uncover the underlying topics in the Thai-language newspapers. The steps that we follow are as outlined below:

Fig. 13

Number of topics and perplexity. Note Validation perplexity is based on the log probabilities of randomly selected articles in a 10% test set. Details on calculations can be found in Wallach et al. (2009)

• (i) Select articles in the economics and business section that contain any of the following keywords: {“uncertain(ty),” “delayed,” “conflict,” “crisis,” “postpone,” “procrastinate,” “confused,” “unsure,” “bankrupt,” “unclear,” “risk(y),” “halted,” “ambiguous”}.

• (ii) Screen out articles that are too short (less than 50 words) or too long (more than 1000 words), which leaves us with a total of remaining 91,400 news articles. Based on our reading in the Thai newspapers, we find that very short articles often contain uninformative news including advertisement and gossip sections while very long articles contain too many different topics that could potentially introduce noise into the estimation process. Note that, on this issue, other refinement process can be applied. For example, Husted et al. (2020) construct monetary policy uncertainty using proximity refinement by restricting uncertainty words to be within 5-20 words of the phrase ‘federal reserve’ or ‘monetary policy’.

• (ii) Apply standard preprocessing by removing stop words and punctuation, lemmatization and stemming.

• (iii) Employ the LDA approach to uncover underlying topics. The LDA assumes a generative process with the following joint distribution:

  $$\begin{aligned} p(\theta , z, w| \alpha , \beta ) = p(\theta |\alpha ) \prod _{n=1}^N p(z_n|\theta ) p(w_n|z_n,\beta ), \end{aligned}$$

  where each document has a multinomial distribution $\theta$ of topics. A document’s topic distribution is randomly sampled from a Dirichlet distribution with $\alpha$ as a parameter governing the concentration, $\beta = \{\beta _1,...,\beta _k\}$ as the topic-word probabilities of K topics, and $z=\{z_1,...,z_N\}$ and $w=\{w_1,...,w_N\}$ are sets of N topics and words, respectively. In this study, we set $\alpha = 50/K$ and $\beta = 0.1$ as suggested by Griffiths and Steyvers (2004) and compare the goodness of fit of models with different values of K. The goodness of fit for each model is based on calculating the perplexity value as proposed by Wallach et al. (2009).¹⁹

  Figure 13 shows the model’s evaluation based on the perplexity value for K between 5 and 200. We find that the goodness of fit improves monotonically with a larger K. Therefore, we also need to consider the interpretability of the topics derived from the model based on our own subjective judgment. Too few topics can produce results that are too broad, while selecting too many topics can lead to too detailed and redundant topics. In our case, we find that the topics become very specific and semantically less meaningful when K is above 50. Thus, we choose $K = 50$ which yields the most interpretable result. Applying a subjective judgment for K is a valid approach that has also been carried out by Larsen (2021), in which he picked $K=80$ to construct measures of economic uncertainty using the LDA method for Norway. In addition, as cited in Larsen (2021), Chang et al. (2009) conclude that “practitioners developing topic models should thus focus on evaluations that depend on real-world task performance rather than optimizing likelihood-based measures” (p. 296).

• (iv) Construct the uncertainty index for each topic based on the number of articles describing uncertainty for each topic. More specifically, we first label each article d with its most likely topic (the topic with the highest probability $\theta _d$). The index is then the estimated topic proportion for each topic within a quarter.

Appendix C: Impulse response analysis of domestic uncertainty subcomponents

See Figs. 14, 15, 16, 17, 1819, 20, 21, 22, 23, 24 and 25.

Fig. 14

Impulse responses of macroeconomic variables to global uncertainty and domestic macroeconomic uncertainty shocks (production and growth outlook). Note Plotted are the impulse responses to global and domestic macroeconomic uncertainty shocks. Shaded regions correspond to 68% standard error bands

Fig. 15

Impulse responses of macroeconomic variables to global uncertainty and domestic macroeconomic uncertainty shocks (commodity prices and agricultural products). Note Plotted are the impulse responses to global and domestic macroeconomic uncertainty shocks. Shaded regions correspond to 68% standard error bands

Fig. 16

Impulse responses of macroeconomic variables to global uncertainty and domestic financial uncertainty shocks (capital markets). Note Plotted are the impulse responses to global and domestic macroeconomic uncertainty shocks. Shaded regions correspond to 68% standard error bands

Fig. 17

Impulse responses of macroeconomic variables to global uncertainty and domestic financial uncertainty shocks (credit markets). Note Plotted are the impulse responses to global and domestic macroeconomic uncertainty shocks. Shaded regions correspond to 68% standard error bands

Fig. 18

Impulse responses of macroeconomic variables to global uncertainty and domestic monetary policy uncertainty shocks. Note Plotted are the impulse responses to global and domestic macroeconomic uncertainty shocks. Shaded regions correspond to 68% standard error bands

Fig. 19

Impulse responses of macroeconomic variables to global uncertainty and domestic fiscal uncertainty shocks. Note Plotted are the impulse responses to global and domestic macroeconomic uncertainty shocks. Shaded regions correspond to 68% standard error bands

Fig. 20

Impulse responses of financial variables to global uncertainty and domestic macroeconomic uncertainty shocks (production and growth outlook). Note Plotted are the impulse responses to global uncertainty and domestic financial uncertainty shocks. Shaded regions correspond to 68% standard error bands

Fig. 21

Impulse responses of financial variables to global uncertainty and domestic macroeconomic uncertainty shocks (commodity prices and agricultural products). Note Plotted are the impulse responses to global uncertainty and domestic financial uncertainty shocks. Shaded regions correspond to 68% standard error bands

Fig. 22

Impulse responses of financial variables to global uncertainty and domestic financial uncertainty shocks (capital markets). Note Plotted are the impulse responses to global uncertainty and domestic financial uncertainty shocks. Shaded regions correspond to 68% standard error bands

Fig. 23

Impulse responses of financial variables to global uncertainty and domestic financial uncertainty shocks (credit markets). Note Plotted are the impulse responses to global uncertainty and domestic financial uncertainty shocks. Shaded regions correspond to 68% standard error bands

Fig. 24

Impulse responses of financial variables to global uncertainty and domestic monetary policy uncertainty shocks. Note Plotted are the impulse responses to global uncertainty and domestic financial uncertainty shocks. Shaded regions correspond to 68% standard error bands

Fig. 25

Impulse responses of financial variables to global uncertainty and domestic fiscal uncertainty shocks. Note Plotted are the impulse responses to global uncertainty and domestic financial uncertainty shocks. Shaded regions correspond to 68% standard error bands