Using Newspapers for Textual Indicators: Guidance Based on Spanish- and Portuguese-Speaking Countries

Abstract

This paper investigates the role that two key methodological choices play in the construction of dictionary-based indicators: the selection of local versus foreign newspapers, and the breadth of the press coverage (i.e. the amount of newspapers considered). The large literature in this field is almost silent about the robustness of research results to these two choices. These questions are relevant since the production of newspaper-based economic indicators is growing fast. We use as a case study the well-known economic policy uncertainty (EPU) index, taking as examples the six largest Latin American economies (Argentina, Brazil, Chile, Colombia, Mexico and Peru) and Spain. First, we develop EPU measures based on press with different levels of proximity, i.e. local versus foreign, and corroborate that they deliver broadly similar narratives. Second, we examine the macroeconomic effects of EPU shocks computed using these different sources by means of a structural Bayesian vector autoregression framework and find similar responses from the statistical point of view. These two applications should reassure researchers that they can rely on foreign sources to construct EPU indexes. This option may foster the comparability of results across countries and lay the groundwork for cross-country studies of uncertainty. Finally, we show that constructing EPU indexes based on only one newspaper, an option followed by many studies, may yield biased responses. Increasing the number of sources reduces the chances of obtaining biased responses. This suggests that it is important to maximize the breadth of the press coverage when building text-based indicators, since this would improve the robustness and credibility of results.

Appendix

1.1 A Data description

The data used to estimate the impact of policy uncertainty on macro and financial variables are sourced via Refinitiv. A complete description of the data is provided in Table 8.

In the case of quarterly data, real GDP series are taken from the NationaI Institute for Geography and Statistics of Brazil (IBGE in Portuguese), the National Institute of Statistics and Geography of Mexico (INEGI in Spanish) and the Oxford Economic database for Argentina, Chile, Colombia and Peru. Inflation is calculated from the respective domestic consumer price indexes, constructed and published by the respective national statistics offices. In both cases, we take the quarterly rate of change. Macro series are seasonally adjusted; we use Refinitiv to adjust for seasonality any series published without seasonal adjustment. The quarterly rates of each country are averaged to obtain the Latin American inflation and growth rates, thus avoiding the overweighting of Brazil and Mexico.

Net portfolio capital inflows from non residents are extracted from balance of payments publications of the respective national statistics offices and complemented with data from the IMF’s databases to build some series back to 2003. Capital flow series are scaled by the nominal GDP levels, and we avoid taking moving averages to reflect the direct impact of increases in EPU on the evolution of capital inflows and outflows. Latin American series are obtained adding portfolio capital flows.

The bilateral exchange rates vis–à–vis the USD are taken from Reuters. As there is no Latin American exchange rate against the US dollar, we take the quarterly changes in each exchange rate and average them for the six countries analysed. We use the benchmark MSCI equity indices in USD to estimate the change in stock prices, and although MSCI has a Latin American index, we prefer to extract the six individual indexes and calculate the average of the quarterly changes as the aggregate index is biased towards the evolution of the major firms of Brazil and Mexico, which have the highest market capitalization.

Finally, we include the VIX index, which represents the market’s expectations regarding price changes in the S &P 500 index. Because it is derived from the prices of Standard and Poor’s index options with near-term expiration dates, it generates a 30-day forward projection of volatility. The VIX is included in levels and attempts to capture global uncertainty, so it is useful to disentangle the effects of EPU shocks from the effects of international events that, in turn, could affect EPU.

The robustness exercise with monthly data uses the same variables as above but for the activity index and portfolio capital flows of non-residents. For the former, we use the monthly GDP proxies published by the national statistics offices and central banks of the region (Estimador Mensual de Actividad Económica (EMAE) for Argentina; Índice de Atividade Econômica do Banco Central (IBC-Br) for Brazil; Índice Mensual de Actividad Económica (Imacec) for Chile; Indicador de Seguimiento a la Economía (ISE) for Colombia; Indicador Global de la Actividad Económica (IGAE) for Mexico; and the INEI monthly GDP for Peru). In the case of portfolio flows by non-residents, the central banks of all countries except Peru publish monthly data.

Table 4 Descriptive statistics for Latin America

Table 5 Descriptive statistics for Brazil

Table 6 Descriptive statistics for Mexico

Table 7 Descriptive statistics for Spain

Table 8 Data description

Table 9 A sample of source coverage in the related literature

1.2 B Constructing the EPU Indexes: Local Press Coverage by Country

Table 10 Local press coverage by country

Table 11 Keywords across press types for the Latin American region

Table 12 Keywords across press types for Spain

1.3 C Narrative Across Alternative Press Coverage by Country

Table 13 Events of economic policy uncertainty in Argentina: comparison across types of press coverage

Table 14 Events of economic policy uncertainty in Brazil: comparison across types of press coverage

Table 15 Events of economic policy uncertainty in Chile: comparison across types of press coverage

Table 16 Events of economic policy uncertainty in Colombia: comparison across types of press coverage

Table 17 Events of economic policy uncertainty in Mexico: comparison across types of press coverage

Table 18 Events of economic policy uncertainty in Peru: comparison across types of press coverage

1.4 D Our EPU Indexes by Country

This section shows each country-level EPU index with its associated timeline of relevant events. In most cases, the peaks in the indexes are explained by events that might increase economic uncertainty in the country. This exercise is commonly used in the literature to provide evidence of the validity of the index as a proxy of economic uncertainty. Nevertheless, there are a few cases in which the spikes do not correspond to a relevant event in the country’s recent history. This is noise. Following the literature, we manually clean the series by replacing each of these noisy peaks with the average of each series. Below, we provide a list of all noisy peaks for each country.

• Argentina: Jul. 2003, Jul. 2005 (foreign press)

• Chile: Feb. 2021 (foreign press)

• Peru: Sep. 2010, Jul. 2017 (foreign press)

Fig. 4

EPU indexes for Argentina. Note: This figure shows the EPU indexes for Argentina based on the local (red line) and foreign press (black line) against the narrative of events associated with increases in policy uncertainty in Argentina

Fig. 5

EPU indexes for Brazil. Note: This figure shows the EPU indexes for Brazil based on the local (red line) and foreign press (black line) against the narrative of events associated with increases in policy uncertainty in Brazil. PAC stands for Growth Acceleration Program

Fig. 6

EPU indexes for Chile. Note: This figure shows the EPU indexes for Chile based on the local (red line) and foreign press (black line) against the narrative of events associated with increases in policy uncertainty in Chile

Fig. 7

EPU indexes for Colombia. Note: This figure shows the EPU indexes for Colombia based on the local (red line) and foreign press (black line) against the narrative of events associated with increases in policy uncertainty in Colombia

Fig. 8

EPU indexes for Mexico. Note: This figure shows the EPU indexes for Mexico based on the local (red line) and foreign press (black line) against the narrative of events associated with increases in policy uncertainty in Mexico

Fig. 9

EPU indexes for Peru. Note: This figure shows the EPU indexes for Peru based on the local (red line) and foreign press (black line) against the narrative of events associated with increases in policy uncertainty in Peru

Fig. 10

EPU indexes for Spain. Note: This figure shows the EPU indexes for Spain based on the local (Spanish—red line) and foreign (Anglophone—blue line) press against the narrative of events associated with increases in policy uncertainty in this Spain

1.5 E Our EPU Indexes for the Latin American Region

Fig. 11

EPU indexes for the Latin American region. Note: This figure shows the EPU indexes for the Latin American region based on the local (red line) and foreign press (black line) against the narrative of events associated with increases in policy uncertainty in this region. The Latin American region is defined as the following countries: Argentina, Brazil, Chile, Colombia, Mexico and Peru

Table 19 EPU indexes for the Latin American region: descriptive statistics across press coverage types

Fig. 12

Frequency distribution of EPU indexes for Spain and the Latin American region: a comparison. Note: This figure shows the distribution of the EPU index for Spain and of the EPU index for the Latin American region. Both indexes are based on all the available press (foreign and local)

Table 20 Descriptive statistics of the EPU indexes

1.6 F Structural VAR Model with Recursive Identification à la Cholesky

In this section we explain the details of our recursive identification strategy to identify structural shocks in our VAR models. First we present the model mathematically and then we illustrate the identification assumptions through a graphical representation.

The multivariate time-series model for periods $t=1, \dots ,T$ consists of a $n \times 1$ vector of variables $\{y_t\}_{t=1}^T$, $n \times n$ matrices of coefficients $A_{i}, i = 1, \dots ,p$ capturing the dynamics of the pth lagged system, a $n \times 1$ vector of constants C and an $n \times 1$ vector of error terms $\varepsilon _t$ with zero-mean and positive-definite $n \times n$ covariance matrix $\Sigma$. This data-generating process is assumed to evolve according to the following VAR(p):

$$\begin{aligned} y_t = C + A_{1} y_{t-1} + A_{2} y_{t-2} + \dots + A_{p} y_{t-p} + \varepsilon _t, \quad \text {with} \ \varepsilon _t \sim \mathcal {N}(0,\Sigma ). \end{aligned}$$

(F.1)

The open-economy model contains variables for the global block (G) and for the local block (L) consisting of variables for Latin American countries like Brazil (BR) or Mexico (MX), and Spain (ES). Hence, in our benchmark specification we have $y_t = \{y_t^{G}, y_t^{L}\}$, with $L=(BR, MX, ES)$. Each superscript indicates a group of variables (five variables in total, one in the G group and four in the L group). In particular, we have $y_t^{G} = VIX_t$ and $y_t^{L} = (EPU_t, PCF_t, GDP_t, CPI_t)^\top$. The block exogeneity assumption entails imposing zero-restrictions on the coefficient matrices so to cancel out past effects³⁴ of $y_t^{L}$ on $y_t^{G}$:

$$\begin{aligned} \begin{pmatrix} y_t^{G}\\ y_t^{L} \end{pmatrix} =\begin{pmatrix} c_{1} \\ c_{2} \end{pmatrix} +\sum _{i=1}^p \begin{pmatrix} a_{11,i} &{} 0 \\ a_{21,i} &{} a_{22,i} \end{pmatrix} \begin{pmatrix} y_{t-i}^{G}\\ y_{t-i}^{L} \end{pmatrix} +\begin{pmatrix} \varepsilon _t^{G}\\ \varepsilon _t^{L} \end{pmatrix}. \end{aligned}$$

(F.2)

One might be interested in having an economic interpretation of the shocks, and for this, one needs to define the contemporaneous relationships between variables. Let $D_{j}, j = 0, \dots ,p$ capture such structural relationships, K be the constants and $\eta _t$ be the structural innovations with zero-mean and structural covariances $\Gamma$. Therefore, the structural VAR(p) is given by:

$$\begin{aligned} D_{0}y_t = K + D_{1} y_{t-1} + D_{2} y_{t-2} + \dots + D_{p} y_{t-p} + \eta _t, \quad \text {with} \ \eta _t \sim \mathcal {N}(0,\Gamma ). \end{aligned}$$

(F.3)

As a result, the link between the reduced-form VAR and the structural counterpart³⁵ is established by $A_i = D D_{i}$, $C = D K$, $\varepsilon _t = D \eta _t$, and $\Sigma = D \Gamma D^\top$, with $D = D_{0}^{-1}$. To recover the structural model from its reduced-form or disentangle $\eta _t$ from $\varepsilon _t$, one needs to adopt an identification strategy to restrict the contemporaneous matrix D.

This comes from the fact that for the system of equations in $\Sigma = D \Gamma D^\top$ there are $n^2$ elements to be identified in D and $\frac{n(n+1)}{2}$ in $\Gamma$ due to its symmetry about the diagonal. Once the structural covariance is normalised³⁶$\Gamma = I_n$, $I_n$ being a $n \times n$ identity matrix, there are just $n^2$ elements left to identify. Since one can estimate from the reduced model $\frac{n(n+1)}{2}$ distinct elements from $\Sigma$, only $\frac{n(n-1)}{2}$ additional restrictions on D are needed to just identify the system. For this purpose, we impose exclusion restrictions recursively, which implies finding a lower triangular matrix $D \quad \text {s.t.} \quad \Sigma = D D^\top$. One way to achieve this is to apply the Cholesky decomposition of $\Sigma = P P^\top$, with P being a $n \times n$ lower triangular matrix with positive diagonal. It follows that $P = D$ is one possible solution that satisfies the needed $\frac{n(n-1)}{2}$ restrictions for exact identification.³⁷ Accordingly, we impose contemporaneous restrictions such that upper off-diagonal elements in D are zero.

$$\begin{aligned} \begin{pmatrix} \varepsilon _{VIX_t}^{G}\\ \varepsilon _{EPU_t}^{L}\\ \varepsilon _{PCF_t}^{L}\\ \varepsilon _{GDP_t}^{L}\\ \varepsilon _{CPI_t}^{L} \end{pmatrix} =\begin{pmatrix} d_{11} &{} 0 &{} 0 &{} 0 &{} 0 \\ d_{21} &{} d_{22} &{} 0 &{} 0 &{} 0 \\ d_{31} &{} d_{32} &{} d_{33} &{} 0 &{} 0 \\ d_{41} &{} d_{42} &{} d_{43} &{} d_{44} &{} 0 \\ d_{51} &{} d_{52} &{} d_{53} &{} d_{54} &{} d_{55} \\ \end{pmatrix} \begin{pmatrix} \eta _{VIX_t}^{G}\\ \eta _{EPU_t}^{L}\\ \eta _{PCF_t}^{L}\\ \eta _{GDP_t}^{L}\\ \eta _{CPI_t}^{L} \end{pmatrix}. \end{aligned}$$

(F.4)

Note that the ordering of the variables in both groups is crucial for this identification scheme since its recursive nature imposes a particular causal chain. The position of the variable in the ordering shows to which variables it is restricted to react to and to which variables it is restricted to affect in period t.

To illustrate this identification scheme we create a diagrammatic flow that explains graphically the assumptions underlying the recursive identification in Fig. 13. Each arrow represents the impact of the contemporaneous structural shock of one variable on the other variables in the system. As the graph shows, each variable affect contemporaneously itself. However, the ordering of the variable in the system implies that shocks of one variable affect all variables that are placed below, and do not affect variables that are ordered before. For instance, this means that shocks in the EPU affect contemporaneously the economy, but that shocks in the economy cannot affect contemporaneously uncertainty, but do it only with a lag (as specified by the model).

Fig. 13

Graphical representation of the recursive identification à la Cholesky. Note: This diagrammatic flow represents the recursive identification à la Cholesky used to estimate the structural shocks in the structural VAR models. Each arrow represents the impact of the contemporaneous structural shock of one variable on the other variables in the system

1.7 G Robustness Results: BVAR Specifications

Fig. 14

Robustness results: IRFs of GDP to EPU shocks for the Latin American region. Note: Each panel depicts the median impulse response of GDP to a rise of one standard deviation in the EPU index in the LATAM region: the red and blue lines represent responses to EPU shocks based on the local and foreign press, respectively. Filled symbols indicate statistical significance within the 84%–16% credible set, while empty symbols represent not-significant estimates. The horizontal axis measures quarters since the shock

Fig. 15

Robustness results: IRFs of GDP to EPU shocks for the Latin American region using alternative financial variables in the model. Note: Each panel depicts the median impulse response of GDP to a rise of one standard deviation in the EPU index in the LATAM region: the red and blue lines represent responses to EPU shocks based on the local and foreign press, respectively. Filled symbols indicate statistical significance within the 84%–16% credible set, while empty symbols represent not-significant estimates. The horizontal axis measures quarters since the shock. PCF stands for portfolio capital flows

1.8 H Robustness Results: Breadth of Press Coverage

Fig. 16

Robustness exercise regarding the breadth of press coverage. Note: Each chart reports on the vertical axis the percentage of coefficients of the response based on the EPU constructed from each combination of newspapers that are close enough (i.e. not statistically different) to the credible bands of the baseline responses. The horizontal axis shows the possible combinations of sources (56 for Brazil and Mexico, 127 for Spain), ordered increasingly depending on the number of sources (the black vertical line indicates each time the number of sources increases)

1.9 I Additional Benchmark Results: BVAR Exercise for the Latin American Region

Fig. 17

Benchmark model: IRFs of GDP to EPU shocks with credible set. Note: The panels depicts the median impulse response of the specified variable to a rise of one standard deviation in the (a) local press and (b) foreign press EPU indexes, along with the corresponding 90%–10%, 84%–16% and 68%–32% credible sets (marked by decreasing colour intensity). The horizontal axis measures quarters since the shock. In both cases we estimate the model with portfolio capital flows as financial variable

Table 21 Baseline results: IRF coefficients