Extracting attention trends from an investment newsletter

Natural language processing can bring considerable insight to economists and investment strategists. This note describes an early example: applying topic modelling to three years of Points of Return, the daily newsletter written by Bloomberg senior editor and columnist John Authers.

We wanted to know how the newsletter’s focus had shifted since June 2020. We used Latent Dirichlet Allocation, the topic model developed by David Blei, and averaged the resulting topic distribution on a weekly basis.

The pattern was striking. Monetary policy and the Federal Reserve’s fight against inflation came to dominate the newsletter almost entirely. Attention to China and to the pandemic receded broadly in tandem. Politics resurfaced episodically rather than persistently, and brief spikes marked the invasion of Ukraine and the UK gilts crisis.

Authers responded to the analysis in Points of Return on 14 August 2023. He accepted the weight given to inflation and monetary policy as defensible, given that the consequences of the 2020 interventions remained unresolved. The decline in coverage of China and emerging markets he found harder to defend, describing himself as “remiss in neglecting” them and undertaking to return to the subject. He also noted, reasonably, that a topic model reaches only what has been written down.

That last point is the useful one. Topic modelling did not read anyone’s mind; it measured allocation of attention across a published corpus, which is a narrower and more tractable question — and one worth asking of any research process, including your own.