New Data Sources for Central Banks

A recent branch of the literature relies on newspaper articles to compute indicators of economic uncertainty. Text data are indeed a valuable new source of information since they reflect major current events that affect the decisions of economic agents and are available with no time lag.

In their leading paper, Baker et al. (see [4]) constructed an index of economic policy uncertainty (Economic Policy Uncertainty (EPU) index) for the United States, based on the volume of newspaper articles that contain words relating to the concepts of uncertainty, economy, and policy. Since this seminal paper, many researchers and economic analysts have used text-based uncertainty indicators in their analyses, providing empirical evidence of the negative effects on activity in many countries (e.g., see [50] for Germany, France, Italy, and Spain, [35] for China, or [23] for the Euro area). The authors of this chapter constructed an EPU index for Spain based on two leading Spanish newspapers: (El País and El Mundo). [38] recently developed a new Economic Policy Uncertainty index for Spain, which is based on the methodology of [4] but expands the press coverage from 2 to 7 newspapers, widens the time coverage starting from 1997 rather than from 2001, and fine-tunes the richness of the keywords used in the search expressions.⁷

The indicator shows significant increases or decreases relating to events associated, ex ante, with an increase or decrease in economic uncertainty, such as the terrorist attacks of September 11, 2001, in the United States, the collapse of Lehman Brothers in September 2008, the request for financial assistance by Greece in April 2010, the request for financial assistance to restructure the banking sector and savings banks in Spain in June 2012, the Brexit referendum in June 2016, or the episodes of political tension in the Spanish region of Catalonia in October 2017.

[38] found a significant dynamic relationship between this indicator and the main macroeconomic variables, such that unexpected increases in the economic policy uncertainty indicator have adverse macroeconomic effects. Specifically, an unexpected rise in uncertainty leads to a significant reduction of GDP, consumption, and investment. This result is in line with the findings in the empirical literature on economic uncertainty.

In addition, the authors of this chapter provide evidence on the relative role of enriching the keywords used in search expressions and widening both press and time coverage when constructing the index. Results are shown in Fig. 1, which compares macroeconomic responses to unexpected shocks in alternative EPU versions in which they vary in one of the aforementioned dimensions at a time, moving from the EPU index constructed by [4] to the new index. All of these dimensions are important since they all contribute to obtaining the expected negative sign in the responses. Expanding the time coverage is key to improving the precision of the estimates and to yielding significant results. The press coverage is also relevant.

By documenting the spillover effects of rising uncertainty across countries, the literature also demonstrates that rising economic uncertainty in one country can have global ramifications (e.g., [8, 9, 23, 59]). In this respect, [39] develop Economic Policy Uncertainty indexes for the main Latin American (LA) countries: Argentina, Brazil, Chile, Colombia, Mexico, Peru, and Venezuela. The objective of constructing these indexes is twofold: first, to measure economic policy uncertainty in LA countries in order to get a narrative of “uncertainty shocks” and their potential effects on economic activity in LA countries, and second, to explore the extent to which those LA shocks have the potential to spillover to Spain. This latter country provides an interesting case study for this type of “international spillover” given its significant economic links with the Latin American region.

The uncertainty indicators are constructed following the same methodology used for the EPU index for Spain [38], i.e., counting articles in the seven most important national Spanish newspapers that contain words related to the concepts of economy, policy, and uncertainty. In addition, however, we customize the text searches for the Latin American countries case by case.⁸ Note that these indicators are also based on the Spanish press and thereby purely reflect variation in uncertainty in LA countries that is relevant to the Spanish economy, given the importance of the region to the latter. The premise is that the Spanish press accurately reflects the political, social, and economic situation in the LA region, given the existing close economic and cultural ties—including a common language for a majority of these countries. In this respect, one may claim that the indexes provide sensible and relevant measures of policy uncertainty for those countries. This is also in line with a branch of the literature that uses the international press to compute text-based indicators for broad sets of countries (see, e.g., [2] or [53]).

To explore the extent to which LA EPU shocks have the potential to spillover to Spain, the empirical analysis relies on two exercises. A first exercise studies the impact of LA EPU shocks on the performance of Spanish companies operating in the LA region. The underlying assumption is that higher uncertainty in one LA country would affect the investment decisions of Spanish companies that have subsidiaries in this Latin American country: i.e., investment in the LA country may be postponed due to the “wait-and-see effect” and/or the local uncertainty may foster investment decisions toward other foreign countries or within Spain. To carry out this exercise, the authors consider the stock market quotations of the most important Spanish companies that are also highly exposed to LA countries, controlling for the Spanish macroeconomic cycle. Results show that an unexpected positive shock in the EPU index of an LA country generates a significant drop in the companies’ quotation growth rate in the first 2 months. This holds for all LA countries considered in the study and is confirmed by placebo tests, which consider Spanish companies that are listed in the Spanish stock market but do not have economic interests in the Latin American region. This suggests that, as expected, economic policy uncertainty in LA countries affects the quotations of Spanish companies that have economic interests in that region.

The second exercise studies the impact of Latin American EPU shocks on the following Spanish macroeconomic variables: the EPU index for Spain, exports and foreign direct investment (FDI) from Spain to Latin America, and the Spanish GDP. In this case as well, one would expect the spillover from one LA country’s EPU to the Spanish EPU to be related to commercial relationships between both countries. The higher the exposure of Spanish businesses to a given country, the higher the spillover. To the extent that the EPU reflects uncertainty about the expected future economic policy situation in the country, unexpected shocks in the EPU of one LA country may affect the export and FDI decisions of Spanish companies. Finally, the relation between Latin American EPUs and the Spanish GDP is expected to be driven by the reduction in exports (indirect effect) and by the business decisions of multinational companies that have economic interests in the region. In particular, multinational companies take into account the economic performance of their subsidiaries when deciding upon investment and hiring in Spain. This, in turn, may affect the Spanish GDP. This second exercise is carried out at the quarterly level by means of VAR models, which document the spillover effects from Latin American EPU indexes to the Spanish EPU. Unexpected shocks in Latin American EPUs significantly dampen the commercial relationship between Spain and the Latin American countries in question. In particular, Spanish firms decrease their exports and FDI toward the countries that experience negative shocks in their EPU index. As an example, Fig. 2 shows the impulse response functions of Spanish net FDI to unexpected shocks in the Mexican EPU index.

As an empirical exercise, we construct a text-based indicator that helps track economic activity, as described in [1]. It is based on a similar procedure that is used to elaborate the economic policy uncertainty indicator, i.e., it relies on counting the number of articles in the Spanish press that contain specific keywords. In this case, we carry out a dictionary analysis as in the previous section, i.e., we set up a dictionary of positive and negative words that are typically used in portions of texts related to the GDP growth rate, the target variable of interest, so as to also capture the tone of the articles and, in particular, to what extent they describe upturns or downturns. For instance, words like “increase,” “grow,” or “raise are listed among the positive terms, while “decrease” and “fall” appear in the negative list. As with the EPU indicators, this one is also based on the Factiva Dow Jones repository of Spanish press and relies on seven relevant Spanish national newspapers: El País, El Mundo, Cinco Días, Expansión, El Economista, La Vanguardia, and ABC.

We place the following restrictions on all queries: (1) the articles are in Spanish; (2) the content of the article is related to Spain, based on Factiva’s indexation; and (3) the article is about corporate or industrial news, economic news, or news about commodities or financial markets, according to Factiva’s indexation. We then perform three types of queries for each newspaper:¹²

Then, for each newspaper, we take the difference between the upturn- and downturn-related counts and scale the difference by the total number of economic articles in the same newspaper/month. Finally, we standardize the monthly series of scaled counts, average them across newspapers, rescale the resulting index to mean 0, and average it at the quarterly level.

The right panel in Fig. 4 shows the resulting textual indicator (solid blue line) against the GDP growth rate (red and dashed line).

Next, we test whether our textual indicator has some predictive power to nowcast the Spanish GDP growth rate. We perform a pseudo-real-time nowcasting exercise at the quarterly level as follows.¹³ First, we estimate a baseline nowcasting model in which the GDP growth rate is nowcasted by means of an AR(1) process. Second, we estimate an alternative nowcasting model that adds our textual indicator and its lag to the GDP AR(1) process. Finally, we compare the forecast accuracy of both models. The alternative model provides smaller mean squared errors of predictions than the baseline one, which suggests that adding textual indicators to the AR(1) process improves the predictions of the baseline model. In addition, according to the Diebold–Mariano test, the forecast accuracy of the model improves significantly in the alternative model. The null hypothesis of this test is that both competing models provide the same forecast accuracy. By comparing the baseline with the alternative model, this hypothesis is rejected at the 10% level with a p-value of 0.063.¹⁴

A major advantage of newspaper-based indicators is that they can be updated in real time and are of high frequency. This has been extremely valuable since the Covid-19 outbreak, when traditional survey-based confidence indicators failed to provide timely signals about economic activity.¹⁵ As an example, the right panel in Fig. 4 depicts the textual indicator at a weekly frequency around the Spanish lockdown (14 March 2020) and correctly captures the drastic reduction in Spanish economic activity around that time.

The latent Dirichlet allocation or LDA (see [11]) method can be used to estimate topics in text data. This is an unsupervised learning method, meaning that the data do not need to include a topic label and that the definition of the topics is not decided by the modeler but is a result of running the model over the data. It is appealing because, unlike other methods, it is grounded in a statistical framework: it assumes that the documents are generated according to a generative statistical process (the Dirichlet distribution) so that each document can be described by a distribution of topics and each topic can be described by a distribution of words. The topics are latent (unobserved), as opposed to the documents at hand and the words contained in each document.

This bag-of-words representation of the corpus is then fed to the LDA algorithm, which is used to identify 128 different topics that these texts discuss¹⁶ and to assign to each piece of news an estimate of the probability that it belongs to each one of those topics. The algorithm analyzes the texts and determines which words tend to appear together and which do not, optimally assigning them to different topics so as to minimize the distance between texts assigned to any given topic and to maximize the distance between texts assigned to different topics.

The result is a database that contains, for each quarter from 1997 to 2018, the percentage of news pieces that fall within each of the 128 topics identified by the unsupervised learning model. A dictionary of positive and negative terms is also applied to each piece of news, and the results are aggregated into quarterly series that indicate how positive or negative are the news pieces relating to each topic.

We can now turn to a machine learning model using the data resulting from the analysis of Spanish newspapers to forecast Spanish GDP.¹⁷ The term “machine learning” encompasses a very wide range of methods and algorithms used in different fields such as machine vision, recommender systems, or software that plays chess or go. In the context of economics, support vector machines, random forests, and neural networks can be used to analyze microdata about millions of consumers or firms and find correlations, patterns of behavior, and even causal relationships. CBs have incorporated machine learning techniques to enhance their operations, for instance, in the context of financial supervision, by training models to read banks’ balance sheets and raise an alert when more scrutiny is required (e.g., see [21]). For time-series forecasting, ensemble techniques, including boosting and bagging, can be used to build strong forecasting models by optimally combining a large number of weaker models. In particular, ensemble modeling is a procedure that exploits different models to predict an outcome, either by using different modeling algorithms or using different training datasets. This allows reducing the generalization error of the prediction, as long as the models are independent. [7] provides an extensive evaluation of some of these techniques. In this subsection, we present one such ensemble model: a doubly adaptive aggregation model that uses the results from the LDA exercise in the previous subsection, coined DAAM-LDA. This model has the advantage that it can adapt to changes in the relationships in the data.

The ingredients for this ensemble forecasting model are a set of 128 very simple and weakly performing time-series models that are the result of regressing quarterly Spanish GDP growth on its first lag and the weight, positiveness, and negativeness of each topic in the current quarter. In the real-time exercise, the models are estimated every quarter and their first out-of-sample forecast is recorded. Since the share of each topic in the news and its positiveness or negativeness will tend to be indicators with a relatively low signal-to-noise ratio, and since most topics identified in the LDA exercise are not actually related to economics, most of these models will display a weak out-of-sample performance: only 4 out of the 128 outperform a simple random walk. Ensemble methods are designed specifically to build strong models out of such a set of weak models. One advantage is that one does not have to decide which topics are useful and which are not: the model automatically discards any topic that did not provide good forecasts in the recent periods.

One possible way to combine these forecasts would be to construct a nonlinear weight function that translates an indicator of the recent performance of each model at time t into its optimal weight for time t + 1. We constructed such a model, using as a weight function a neural network with just three neurons in its hidden layer, in order to keep the number of parameters and hyperparameters relatively low. We used a k-fold cross-validation procedure¹⁸ to find the optimal memory parameter for the indicator of recent performance and the optimal regularization, which restricts the possibility that the neural network would overfit the data. The problem is that after all of this, even if the small neural network was able to generate all sorts of potentially very nonlinear shapes, the optimal weighting function would end up looking like a simple step function, as seen in Fig. 5.

To some extent, this was to be expected as it is already known in the forecasting literature that sophisticated weighting algorithms often have a hard time beating something less complex, like a simple average (see, e.g., [29]). In our case, though, since our weak models are really not well-performing, this would not be enough. So instead of spending the degrees of freedom on allowing for potentially highly nonlinear weights, the decision taken was to use a simple threshold function with just one parameter and then add complexity in other areas of the ensemble model, allowing the said threshold to vary over time.

Importantly, the whole exercise is carried out in real time, using only past information in order to set up the parameters that are to be used for each quarter. Figure 6 summarizes the results from this experiment and also displays the threshold that is used at each moment in time, as well as the memory parameter and speed of adjustment of the threshold that are found to be optimal each year.

As seen in Table 1, the forecasts from DAAM-LDA can outperform a random walk, even if only 4 out of the 128 weak models that it uses as ingredients actually do so. If we restrict the comparison to just the last 4 years in the sample (2015–2018), we can include other state-of-the-art GDP nowcasting models currently in use at the Bank of Spain. In this restricted sample period, the DAAM-LDA model performs better than the random walk, the simple AR(1) model, and the Spain-STING model (see [17]). Still, the Bank of Spain official forecasts display unbeatable performance compared to the statistical methods considered in this section.

The exercise presented in this section follows closely our paper [40]. In that paper, the question is whether new sources of information can help predict private household consumption. Typically, benchmark data to approximate private household spending decisions are provided by the national accounts and are available at a quarterly frequency (“hard data”). More timely data are usually available in the form of “soft” indicators, as discussed in the previous subsection of this chapter. In this case, the predictive power of new sources of data is ascertained in conjunction with the traditional, more proven, aforementioned “hard” and “soft” data.¹⁹ In particular, the following sources of monthly data are considered: (1) data collected from automated teller machines (ATMs), encompassing cash withdrawals at ATM terminals, and point-of-sale (POS) payments with debit and credit cards; (2) Google Trends indicators, which provide proxies of consumption behavior based on Internet search patterns provided by Google; and (3) economic and policy uncertainty measures,²⁰ in line with another recent strand of the literature that has highlighted the relevance of the level of uncertainty prevailing in the economy for private agents’ decision-making (e.g., see [12] and the references therein).

To exploit the data in an efficient and effective manner, [40] build models that relate data at quarterly and monthly frequencies. They follow the modeling approach of [45]. The forecasting exercise is based on pseudo-real-time data, and the target variable is private consumption measured by the national accounts. The sample for the empirical exercises starts by about 2000 and ends in 2017Q4.²¹ As ATM/POS data are not seasonally adjusted, the seasonal component is removed by means of the TRAMO-SEATS software [41].

In order to test the relevant merits of each group of indicators, we consider several models that differ in the set of indicators included in each group. The estimated models include indicators from each group at a time, several groups at a time, and different combinations of individual models. As a mechanical benchmark, [40] use a random walk model whereby they repeat in future quarters the latest quarterly growth rate observed for private consumption. They focus on the forecast performance at the nowcasting horizon (current quarter) but also explore forecasts at 1 to 4 quarters ahead of each of the current quarter forecast origins (first month of the quarter, second, and third).

The analysis yields the following findings. First, as regards models that use only indicators from each group, the ones that use quantitative indicators and payment cards (amounts) tend to perform better than the others in the nowcasting and, somewhat less so, in forecasting (1-quarter- and 4-quarters-ahead) horizons (see Panel A in Table 2). Relative root mean squared errors (RMSEs) are in almost all cases below one, even though from a statistical point of view, they are only different from quarterly random walk nowcasts and forecasts in a few instances. In general, the other models do not systematically best the quarterly random walk alternative. The two main exceptions are the model with qualitative indicators for the nowcasting horizons and the Google-Trends-based ones for the longer-horizon forecasts. The latter results might be consistent with the prior that Google-Trends-based indicators deliver information for today on steps to prepare purchases in the future.

Finally, it seems clear that the combination (average) of models with individual groups of indicators improves the forecasting performance in all cases and at all horizons (see Panel C in Table 2). Most notably, the combination of the forecasts of models including quantitative indicators with those with payment cards (amounts) delivers, in general, the best nowcasting/forecasting performance for all horizons. At the same time, adding the “soft” forecasts seems to add value in the nowcasting phase. In turn, the combination of a broad set of models produces the lowest RMSE relative to the quarterly random walk in the 4-quarters-ahead forecast horizon.

So, to conclude, this study shows that even though traditional indicators do a good job nowcasting and forecasting private consumption in real time, novel data sources add value—most notably those based on payment cards but also, to a lesser extent, Google-Trends-based and uncertainty indicators—when combined with other sources.

Second, Panel B in Table 2 shows the results of the estimation of models that include quantitative indicators while adding, in turn, variables from the other groups (qualitative, payment cards (amounts), uncertainty, Google Trends). The improvement in nowcast accuracy is not generalized when adding more indicators, with the exception of the “soft” ones. Nonetheless, there is a significant improvement for longer forecast horizons when expanding the baseline model. In particular, for the 4-quarters-ahead one, uncertainty and Google-Trends-based indicators add significant value to the core “hard”-only-based model.