Covid-19: How to build better Early Warning Systems
The COVID-19 crisis has not only brought about a double whammy on global demand and supply, but also an assault on the very assumption upon which our societies, economies, and political systems rest: that we operate in a mostly predictable environment. Consumers, investors, workers, and policymakers tend to act as if the world was calculable and the near future could be forecasted with sufficient accuracy. When this basic conviction breaks down, rational decision making becomes virtually impossible leaving many individuals all but paralysed. As a result, uncertainty shocks can lead to sharp drops of aggregate economic activity (e.g. Bloom et al. 2007, Bloom 2009), potentially even leading to prolonged periods of dismal economic performance, as Romer (1990) has shown in the context of the Great Depression of the 1930s.
This column argues that we should prepare for future shocks by designing better early warning systems, so that we have more time to make provisions to cushion the impact. What do we want from an uncertainty indicator? It should be timely and forward-looking, as Baker et al. (2020) point out. Ideally, it should capture developments before they show up in financial markets, survey data, or economic variables. It should provide reliable comparisons of uncertainty levels over time. After all, a minor shock may have no economic impact at all while a big one may present a substantial blow to the economy; a shock that occurs during a period of already elevated levels of uncertainty tends to be more harmful than one during calmer times (Mann 2020). Furthermore, an indicator should give us a clue where the shock is coming from since different sources of uncertainty may have different economic effects. Time, size, and source matter. If we could gauge the severity and origin of a shock at an early stage with some accuracy, we might be able to avoid some of its subsequent economic and social devastations. To get closer to this goal, a deeper understanding of the nature of economic uncertainty and the interactions between different kinds of uncertainty is needed. In this vein we propose a news-based indicator we call Uncertainty Perception Indicator (Müller and Hornig 2020).
The UPI is built to differentiate between various sources of uncertainty that need not be identified ex ante. While the well-known Economic Policy Uncertainty Indicator (Baker et al. 2016) is built on a predefined set of policy areas (monetary, fiscal, taxation, regulation) – and consequently responded to the COVID-19 crisis only in early April, when the March data were in and much of Asia and Europe had already been under shutdown conditions for weeks – the UPI is more open to new developments from hitherto unknown directions, whether from politics, markets, or elsewhere. Applying a topic modelling approach, Latent Dirichlet Allocation (LDA), our aim is to discover surprising unknowns at early stages as their impact on the economy starts to unfold. Such a strategy seems warranted as the world may be entering an era of ‘Green Swan events’ (Bolton et al. 2020), i.e. global uncertainty shocks that do not originate in economic policy or markets, and are truly exogeneous. The COVID-19 crisis can be interpreted as a precursor of this era (da Silva 2020).
A Taxonomy of Economic Uncertainty
In our reasoning economic uncertainty comes in three varieties:1
Market-based uncertainty encompasses events like sudden shifts in market sentiment, the bursting of bubbles, or the spreading of pessimistic expectations. The sources of this type of uncertainty are located in the market itself but cannot be forecasted due to limitations of economic models and data availability.
Economic policy uncertainty refers to unforeseen developments in the realm of politics that may have economic consequences. Surprising election outcomes, incalculable populist politics, or unforeseen effects of economic policy measures fall in this category.
Truly exogenous economic uncertainty derives from factors that are located outside of both the market and the political system. Possible sources of this type of uncertainty are plentiful: technology, natural disasters, pandemics (such as Covid-19), meteor impacts, severe weather events, to name just a few.
Figure 1 Three types of economic uncertainty and their potential (direct) economic impacts
Source: Müller and Hornig (2020)
The three types of uncertainty should not be interpreted as strictly separate but as interconnected. Figure 1 provides an overview (endogenous interactions within the market system are not depicted). Think of the COVID-19 pandemic: truly exogeneous economic uncertainty arises from a novel virus and its initially unknown infectious properties, prompting authorities and central banks to act in unproven ways with, at first, uncertain consequences (economic policy uncertainty), which, in turn, influences economic sentiment in detrimental ways, potentially leading to the bursting of bubbles and other peculiarities of market-based uncertainty.
There is a vast array of potential dangers an economy might face in the future, but most of them will never materialise. An indicator that captures all the unknowns in the world would be of little use, since it would be prone to false alarms. On the other hand, an indicator that only looks for sources of uncertainty that have materialised in the past is bound to miss the new and surprising stuff. Therefore, any uncertainty gauge should detect the known unknowns (e.g. fiscal and monetary policy, trade policy, regulation) while being open to surprising unknowns (e.g. technological, ecological, social changes with some already apparent economic impact).
The UPI combines a somewhat open query, that filters for newspaper articles containing words related to the economy and uncertainty – but does not pre-define specific policy areas as the EPU does – with the unsupervised topic modelling method Latent Dirichlet Allocation (LDA).2 LDA yields clusters of contextually related newspaper articles (‘topics’) that we subsume into uncertainty factors, i.e. thematic subsets of the UPI, each associated with different parts of the economy and different types of uncertainty.3 The frequencies of the UPI and the EPU for Germany show comparable patterns (figure 2). The EPU’s peaks are more pronounced for methodological reasons, but the UPI captures these hikes as well not missing any important event.
Figure 2 Comparing EPU and UPI for Germany*
*montly data; left axis: percentage of UPI analysis corpora relative to overall corpus, right axis: EPU index points;
Applying LDA enables a decomposition of the UPI into topics and analytical uncertainty factors (for details see endnote 3). Figure 3 shows the frequencies over time of uncertainty factors UPI Politics and UPI Real Economy. Two immediate observations are striking: a) the predominance of political uncertainty, and b) the (apparent) relationship between hikes in policy uncertainty (‘Euro crisis’, ‘Populist Surge’) and increasing uncertainty in the real economy over the period analyzed.
Figure 3 UPI Politics vs. UPI Real Economy
*shares in analysis corpus; six-month moving averages; source: authors’ calculations
Digging deeper into the model’s results reveals which sources are driving these movements. Figure 4 displays the topics behind UPI Politics. The decomposition of this uncertainty factor enables us to uncover developments in different policy areas without having to pre-define them in the query. What we find is a pattern: over the past decade, policy uncertainty was the result of a secular rise of uncertainty concerning the EU, international relations and domestic politics, while central banks played a soothing role. At the very end of the time horizon, however, this pattern changes as uncertainty concerning the unprecedented actions of central banks, that have been the first line of defense in the face of the COVID-19 crisis, shoots up.
Figure 4 Factors of political uncertainty and specific events
Source: Müller and Hornig (2020)
Early warning properties?
This column started with the question whether we were able to build a better indicator than the ones available so far, in the sense that it reacts in a timelier fashion to hitherto unknown sources of uncertainty. Would we have been able to see the enormity of the COVID-19 shock coming by, say, mid-February, if the UPI had been available then? In Müller and Hornig (2020) we propose a quantitative-qualitative routine for the analysis of the most recent past and suggest a two-step procedure:
a) How has the overall indicator behaved in recent months?
b) Which topics are on the rise? Which ones are in decline?
We find that six out of eleven topics have risen over the first quarter of 2020 relative to their levels at the end of December 2019, with ‘EU Conflicts’, ‘Central Banks’, and ‘German Economy’ showing the steepest increases. As early as February 16 we see changes in these three topics with the correct sign, i.e. a considerable rise of associated uncertainty. Thus, less than six weeks after the COVID-19 virus was first officially acknowledged in China, traces of its fallout can already be found in the German newspaper corpus’ content structure. Additionally, LDA allows us to go all the way down to the level of individual newspaper stories: recent articles that fit the model well provide insights into particular developments and dominant framings. This feature offers an economical way to analyse the characteristics of uncertainty, even in public spheres of which individual researchers may have little knowledge. It is thus conceivable to build a family of UPIs covering a host of countries.
Baker, S R, N Bloom & S J Davis (2016), “Measuring Economic Policy Uncertainty” The Quarterly Journal of Economics 131(4): 1593–1636.
Romer, C D (1990), “The Great Crash and the Onset of the Great Depression”, The Quarterly Journal of Economics 105(3): 597-624.
1 A substantial body of literature deals with the economic impact of uncertainty. But the origins of uncertainty tend to be dealt with in a rather cursory fashion. Larson (2017) is a notable exception.
2 In earlier writings (Müller et al. 2018, Müller 2020) we mimicked the EPU query for Germany exactly.
3 Since LDA is a sorting mechanism itself, aspects that are irrelevant to our analysis that show up in distinct clusters of articles (“topics”) can henceforth be ignored. We use newspaper corpora of two leading nation-wide German newspapers, Die Welt and Handelsblatt. The data was provided by the publishing companies and by LexisNexis. We review a period from January 2008 to March 2020. These two newspapers are merged into a single corpus containing 752.000 articles. Applying the EPU query (Baker et al. 2016) yields an analysis corpus of 8295 articles; the broader query, without filtering for certain policy areas, contains 15.077 articles. For each of these sub-corpora we compute LDAs with several K-values (6, 8, 10, and 12). After reviewing the results, the models with parameter values K=10 and 12 were considered the most promising and were consequently analyzed in greater detail. The analysis was conducted using tosca, an R package for statistical content analysis developed by DoCMA researchers (Koppers et al. 2020). Here, we only present results for the parameter setting K=12. LDA, the way we apply it, can be characterized as a quantitative-qualitative approach. Topics need to labeled and interpreted by human researchers. The algorithm provides three starting points for this process: each topic’s “top words” and “top articles” (the ones with best fit to model) as well as its frequency over time. Table 1 gives an overview of the results. In Müller and Hornig (2020) we provide an extensive appendix presenting each topic’s key-features.
Table 1 Overview of Topics and Labels (model A, K=12)
We combine 1 and 3, 4 and 6, 11 and 12 due to their contextual proximity.