Over the past several years, immigration has become a hot-button political issue in Europe, as it has in other parts of the world. In 2015, the Syrian civil war spurred a massive surge in the flow of refugees across the Mediterranean Sea, a period that became known as the “European migrant crisis.” Although the flow of new asylum seekers to Europe has slowed sharply since then, the initial surge roiled the politics of many host countries in lasting ways.

If you regularly read the news, you already know the general outlines of this story, and others have done an excellent job tracking and even visualizing the flows of migrants. What’s harder to track are subtler variations across countries and over time in the political reactions to this process. Election results can be a useful metric, but they are also a clunky one, as elections only happen occasionally, and parties are rarely organized around this single issue. Some public-opinion surveys ask about attitudes toward immigration, but they are only conducted occasionally, often once a year at most.

Kensho’s topic trend scores offer another way to track the salience of political issues within and across countries, including migration in Europe. Updated daily, these trend scores are a weighted count of news articles associated with a specific country and tagged with a particular topic—here, “migration.” These topic tags are generated by an array of binary classifiers, each trained on hundreds of hand-labeled examples. The weights in those tallies represent the prominence of the country of interest in each article’s text, as determined by a separate machine-learning process.[1] The end result is a score that represents the amount of “chatter” about a given country-topic pair in a particular set of news sources.[2]

So what can these trend scores tell us about the reverberations of the so-called migrant crisis in Europe? A few examples suggest that the scores have tracked changes in the political salience of this issue in interesting ways. In particular, variation across countries suggests that the scores have been sensitive to differences in the timing and scale of immigrant flows around Europe, and in the domestic political contexts in which those immigrants landed.

Germany, for example, has Europe’s largest economy and population, and it has received more refugees over the past several years than any other EU country. The chart below shows the monthly trend score for ‘migration’ and Germany since January 2013, along with monthly counts of asylum applicants in Germany (per Eurostat) over the same period. A few points stand out.

  • After ambling along at a low level for a few years, the ‘migration’ trend score spikes dramatically in mid-2015, the very moment when refugee arrivals in Germany and other parts of Europe were accelerating rapidly.
  • What’s more, the trend score has remained elevated ever since. This is consistent with the observation that, even though the flow of refugees into Germany returned to pre-crisis levels in 2016, the crisis has caused persistent changes in German politics.
  • Spikes in the ‘migration’ trend score since 2015 correspond to major political inflection points in which political fights over immigration played a key role—for example, federal elections in September 2017, and infighting over immigration policy among members of the country’s ruling coalition in 2018.


For Italy, the data paint an different picture.

  • Although there was a noticeable increase in the ’migration’ trend score for Italy in 2015, the big jump doesn’t come until 2018, when immigration featured prominently in the country’s parliamentary elections, and in the political fights that followed over the formation of a new government and adoption of a national budget.
  • Our chart suggests that Italy’s later surge in politicking over migration may stem from the slower and later rise in the rate of refugee arrivals in that country. Eurostat data show modest bumps in asylum applications in 2014 and 2015, followed by a larger and longer rise in 2016–2017.


Almost the exact opposite happened in Hungary. As the third and final chart below shows, Hungary saw sharp early spikes in both the rate of refugee arrivals and the ‘migration’ trend score, followed by rapid declines in both.


These examples suggest that our trend scores are picking up on meaningful variations in politics across countries and over time. More generally, this exercise shows how Kensho is using machine learning to produce structured time series from large streams of unstructured text. We believe the resulting data can help analysts do their jobs better and faster by making it easier for them to spot, track, and assess important developments in cases of interest.


[1] To assess a country's prominence in a document, we use entity disambiguation to identify the people, places, and organizations in an article and then estimate the strength of their links to particular countries. For example, Xi Jinping is closely linked to China; Apple may also be linked to China, if more distantly, because of its production sites located there. By looking at the entities in an article, we can tie a story about Jinping to China, even if China is not explicitly mentioned. Instead of simply counting these entities, we weight each by the strength of the link to the country and the confidence score of the model that is performing the entity recognition. This approach is far more robust than simply looking for the text string “China” in an article, allowing us more accurately to assess the extent to which an article is really about a particular country.

[2] The results described here only use articles from Associated Press, to ensure consistency of coverage across time.

Jay Ulfelder

Political science Ph.D. (Stanford 1997), research director for the US government-funded Political Instability Task Force (2001-2011), Good Judgment Project superforcaster.