CentralBank-BERT: Machine Learning Evidence on Central Bank Digital Currency Discourse 

Abstract

This paper quantifies CBDC communication in BIS-posted speeches using a domain-adapted BERT architecture. From 19,609 speeches, 5,376 CBDC sentences are identified in 588 speeches delivered by 157 authors across 57 countries, spanning 2016–2024. Sentence-level classifiers assign labels for Type (Retail, Wholesale, General/Unspecified), Stance (Pro, Wait-and-See, Anti), Sentiment (positive, neutral, negative), and Discourse (Feature, Risk-Benefit, Process). Comparable indicators, dispersion metrics (entropy, HHI), and cross-task dependence (Cramér’s V, mutual information) are constructed. Three facts emerge: (i) communication is retail-oriented and procedural. Retail is the modal type and Process the modal discourse, consistent with rulebooks, intermediated distribution, holding limits, privacy models, and offline use; (ii) tone is predominantly neutral while stance is supportive or cautious, with tight stance–sentiment alignment; (iii) activity surges from 2020, peaks in 2022, and remains elevated, reflecting a compositional pivot toward retail design and program management rather than a simple scale-up of generic statements. Geography is concentrated in the Eurosystem, with meaningful contributions from the US, UK, and Asia-Pacific. Methodologically, a reusable, replication-grade measurement layer for policy communication is provided. Substantively, the evidence indicates a shift in official narratives from whether to how, consistent with managed institutional experimentation that advances operational readiness while preserving policy flexibility.

Please cite the print version as:

Zafar, M. B. (2026). CentralBank-BERT: Machine learning evidence on central bank digital currency discourse. Journal of Economics and Business. https://doi.org/10.1016/j.jeconbus.2026.106300

Keywords: Central Bank Digital Currency, Central-Bank Communication, BIS Speeches, NLP, BERT, Stance and Sentiment, Central Bank-BERT, CBDC-BERT, Financial NLP, CBDC

JEL Classification: E42, E58, E52, D83, C55