WP – 2026-10-23 – Severin Joho And Enzo Rossi: The Evolution Of Frequency, Language Complexity And Policy Narrative: Benchmarking Python-based NLP And ChatGPT On A Century Of SNB Rhetoric

TL;DR

Before you orderOffer from Amazon

Get smart everyday buys delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Severin Joho and Enzo Rossi’s Swiss National Bank working paper analyzes 1,063 speeches by senior officials from 1928 to 2024. It reports that speeches became more frequent, shorter and more accessible over time, while inflation gained prominence in policy discussion during the flexible-exchange-rate era.

Severin Joho and Enzo Rossi have published a Swiss National Bank working paper analyzing 1,063 speeches by senior bank officials from 1928 through 2024. The study finds that speeches became more frequent, shorter and more accessible over time, while inflation gained prominence in policy discussion during the flexible-exchange-rate era.

The researchers assembled and processed a dataset covering speeches delivered by high-ranking Swiss National Bank officials across 96 years. They examined changes in speech frequency, length, readability and thematic content, rather than treating central bank communication as a static record. The paper reports that annual speech frequency followed an upward trajectory from the mid-1950s, with a more marked acceleration after Switzerland moved to flexible exchange rates.

The analysis also identifies a long-term shift in how officials communicate. Speeches became shorter, and their language increasingly accessible from the 1960s onward, according to the authors. In the subjects discussed, inflation moved to the forefront in the flexible-rate era. By contrast, cyclical economic development and unemployment received less emphasis in the speech material examined.

To test how the texts could be analyzed, Joho and Rossi compared traditional Python-based text mining with generative artificial intelligence using ChatGPT. Their paper presents the methods as complementary and discusses algorithmic pitfalls and open research questions. The supplied summary does not report detailed performance scores or establish that one approach outperformed the other.

How SNB Communication Has Changed

The findings offer a long-run view of how a central bank’s public communication has adapted alongside changes in its operating environment. More frequent speeches can mean that officials address audiences more often, while shorter and more accessible language points to changes in the form of that communication. The paper documents these patterns in the SNB speech record; it does not, from the material provided, establish their effects on public understanding or financial markets.

The reported change in themes is also relevant to readers following monetary policy. The growing prominence of inflation in the flexible-exchange-rate era suggests that the subject occupied a larger share of official speech discourse, while cyclical conditions and unemployment featured less. That is a finding about the content of speeches, not by itself proof that the bank’s policy priorities or decisions changed in the same way.

The methods comparison has implications beyond this dataset. Researchers increasingly use language models to analyze large collections of documents, but outputs can be affected by how models interpret text. The authors’ emphasis on combining probabilistic language models with deterministic tools highlights the value of comparing methods and checking results rather than relying on a single automated approach.

A Century of SNB Speeches

The paper, titled The evolution of frequency, language complexity and policy narrative: benchmarking Python-based NLP and ChatGPT on a century of SNB rhetoric, uses speeches from 1928 to 2024. Its period spans substantial changes in monetary and exchange-rate settings. The authors connect the rise in speech frequency to the period after the mid-1950s and note a stronger acceleration following the move to flexible exchange rates.

Rather than focusing only on the policy topics discussed, the study also measures aspects of communication itself, including length and readability. It compares conventional Python-based natural-language-processing methods with ChatGPT, framing the exercise as a methodological benchmark for analysis of central-bank communication. The source summary does not specify the exact readability measures, model prompts or validation procedures used.

“The frequency of annual speeches has followed an upward trajectory since the mid-1950s, accelerating notably after the transition to flexible exchange rates.”

— Severin Joho and Enzo Rossi, authors of the working paper

Limits of the Reported Findings

The summary of the paper does not provide numerical estimates for the changes in speech frequency, length or readability, nor does it give the size of the thematic shifts. It also does not specify how the authors defined accessibility or measured the prominence of each topic. Readers therefore cannot determine from the supplied material how large those changes were or how they varied within the full period.

The reported associations do not establish why speech patterns changed. The source links the acceleration in speech frequency to the period after flexible exchange rates began, but the summary does not describe a causal test. It also does not show whether changes in speech content corresponded directly to policy decisions, audience response or market effects. Details of the ChatGPT comparison, including performance measures and safeguards against model errors, remain unspecified in the source summary.

Further Tests of Language Models

The authors say the paper identifies open questions for future research on using large language models to study central-bank communication. The supplied source does not name a follow-up study, release a timetable for further work or describe a next policy milestone tied to the findings.

For readers assessing the results, the next useful details would be the full paper’s measures, model-testing procedures and evidence on how the two analytical approaches compare. Further research could test whether the same patterns appear across other central banks or document collections, but the source does not say that such studies are planned.

Key Questions

What did the study examine?

It analyzed 1,063 speeches by senior Swiss National Bank officials delivered from 1928 to 2024, examining frequency, length, readability and themes.

What changed in SNB speeches over time?

The authors report that speeches became more frequent, particularly after the mid-1950s and following the transition to flexible exchange rates. They also found a shift toward shorter, more accessible language.

How did the themes change?

According to the paper, inflation became more prominent in the flexible-exchange-rate era, while cyclical economic development and unemployment received less emphasis in the speeches analyzed.

Did ChatGPT outperform traditional text analysis?

The supplied summary does not report that either method outperformed the other. It says the authors benchmarked ChatGPT against Python-based NLP and emphasize their complementary use, alongside attention to algorithmic pitfalls.

Does the study show that SNB policy changed because of these speech trends?

No such causal conclusion is stated in the source summary. The study reports patterns in speech frequency, language and subject matter; it does not establish that those patterns caused policy changes or affected markets.

Source: primary