Companies Around the World Hit by Hong Kong Protests

Category: Best data-driven reporting (small and large newsrooms)

Country/area: Hong Kong S.A.R., China

Organisation: Bloomberg

Organisation size: Big

Publication date: 11 Jun 2019

Credit: Gregory Turk, Gregor Stuart Hunter

Project description:

The project aimed to track the impact of Hong Kong’s protests on multinational corporations through analysis of corporate earnings call transcripts. Through applying a natural language processing model, we were able to find many companies beyond those that had an obvious connection with Hong Kong or a brand presence in the city, such as payments processors, recruitment consultants, chicken farms and Crocs, the shoemaker.

Impact reached:

The project allowed us to cut through the euphemisms and obfuscatory language used by analysts and corporate executives to find words and phrases that were loosely associated with the protests but still relevant – for example, “the political thing” was how one described the growing unrest. By broadening the scope of companies covered with a minimum of time spent, it allowed a few reporters to quickly grasp a large swathe of the local economy in one fell swoop – and make the case to our readers that consumption patterns in Hong Kong are shifting, brands are increasingly looking to mainland Chinese markets and the former British colony is losing its allure as a shopping destination and a jumping-off point for executives travelling to the rest of the country.

Techniques/technologies used:

We used our proprietary Bloomberg Streams toolkit, which uses an AI-powered natural language processing model to search through news, Twitter accounts, and corporate filings.

What was the hardest part of this project?

Generating a meaningful AI model through our Streams tools required ingesting and tagging a lot of data – we tagged almost 3,000 entries as either relevant or not. To get the project ready in time for companies’ quarterly earnings season we had to divide up the labor, and towards this end we came up with an event we dubbed Streams Factory. This required around two dozen staff to train an AI on past examples of warnings by companies at investor days and other corporate access events for investment analysts. To encourage people to show up, we bribed them with cookies.

Natural language processing is a largely unexplored field within journalism. Bloomberg has vast amounts of data, but making sense of the deluge is nearly impossible for a single reporter. By finding a strategy that works, we now have a replicable model for future stories that we can deploy covering everything from self-driving car crashes to virus outbreaks in China.

What can others learn from this project?

Sometimes, successful application of technology is a labour-intensive exercise. It can help to break work down into batches; it can also help to have assistance from colleagues in other countries to advance the project when colleagues in another location are asleep.

Project links: