Economic crises are notoriously hard to predict, and tracking them as they unfold is almost as difficult. Financial market data are an obvious starting point, but they do not offer a complete or objective picture of the magnitude of ongoing crises. Analysts must also consider breaking news about everything from central bank interventions to crisis-fueled goods shortages and IMF negotiations.

One elegant way to consolidate these varied news stories into a kind of crisis barometer is to combine Koto’s topic tags (introduced in a previous post) with some clever geolocation algorithms. The first ingredient of Koto’s economic crisis trend scores is our topic tag of the same name, a machine learning model that identifies when news stories have economic crises as a major theme. Tagging news stories as relevant to these crises goes beyond simple searches for keywords like “crisis”, or boolean queries like “‘crisis’ AND ‘central bank’” because economic crises are a relatively fuzzy topic: not all articles about them are guaranteed to contain obvious buzzwords, and not all articles that do contain these buzzwords actually relate to economic crises. The F1 score of our economic crises topic tag is 0.941, so it does nearly as well as humans correctly classifying articles as crisis-relevant or not.

Besides our economic crisis topic tag, the second ingredient of our economic crisis trend scores is an algorithm to determine the salience of countries mentioned in news stories. A primitive approach to this would be to identify articles as being about a country when they explicitly mention it by name. We improve upon this simple rule in two ways. First, our algorithm ensures that an article about, say, the Turkish province of Aydin -- with no mention of Turkey itself -- is recognized as being about that country. Second, it makes sure that an article entirely about Turkey would get more weight in trend scores for that country than articles about both Turkey and the United States. We do this by calculating the relative frequencies of the countries in each article, which are used as weights in our country-specific trend scores. Below, we present a few examples.

Our economic crisis trend score for Turkey nicely captures the country's recent slide after several crisis-free years with consistently poor macroeconomic fundamentals. The large spike on August 13, 2018 occurred the day that Turkey’s central bank pledged to inject liquidity into the banking system, following the lira’s 9% fall against the dollar to a multi-year low.


The economic crisis trend score for Venezuela is much busier overall, but it has picked up noticeably since the end of 2017, when that country’s economic deterioration accelerated. It peaked most recently on February 20, 2018, the day that Venezuela began selling its petro-based cryptocurrency in a bid to revive its collapsing economy.


Finally, our trend score for Italy registers its role in the Eurozone debt crisis in 2013, and its present macroeconomic difficulties. The most recent spike in Italy’s crisis score occurred in May 2018, a week after the country’s president refused to confirm Eurosceptic economist Paolo Savona as Finance Minister, leaving Italy with a caretaker government and causing a massive sell-off in the FTSE MIB and Italian government bonds. The second-highest peak, in February 2013, occurred the week of Italy’s general election, which led to a hung parliament after the center-left unexpectedly lost ground to right-wing parties, prompting a massive jump in Italy’s borrowing costs.


Besides being valuable economic crisis barometers, these trend scores are potentially valuable sources of trade ideas for investors as well. In future work, we’ll focus on whether our crisis scores for individual countries are leading indicators of prices and volatility in stock, bond and currency markets. Stay tuned.

Jason Weinreb

Quantitative Geopolitical Analyst at Koto, political science Ph.D. from Stanford, former analyst at Variant Perception Macroeconomic Research.