Turning Words into Signals: How AI Maps 140 Years of Central Bank Communication

In a landmark 2025 working paper, IMF researchers Thiago Christiano Silva, Kei Moriya, and Romain Michel Veyrune unveil a transformative method to interpret how central banks talk to the world. Combining artificial intelligence with deep domain expertise, the authors harness a multilingual large language model (LLM) to analyze over 74,000 documents from 169 central banks, spanning 1884 to 2025. This immense corpus, comprising more than 21 million sentences in over 30 languages, forms the basis for the most comprehensive study of central bank communication ever attempted.

Rather than interpreting central bank language through traditional keywords or tone analysis, the study brings in a finely tuned LLM trained to classify individual sentences by topic, communication stance (forward or backward-looking), sentiment (hawkish, dovish, risk-highlighting, or confidence-building), and target audience. This powerful four-dimensional framework turns vague narratives into quantifiable insights, offering a radically new way to monitor, compare, and even forecast central bank actions.

A Fine-Tuned Model Built for Global Central Banks

To power this linguistic revolution, the team fine-tunes a BGE-m3 multilingual sentence transformer, enabling it to interpret the complex, technical, and often subtle language used in monetary policy. This approach avoids the costs and limitations of using commercial models like GPT, which are optimized for text generation rather than structured classification. Crucially, the model processes language at the sentence level, allowing it to detect shifts in policy stance or tone even within a single document.

The training data is ingeniously constructed from a mix of expert-written, synthetically generated, and real-world sentences pulled from central bank documents. This hybrid method ensures the model grasps both the stylized structure of official communication and the nuanced policy cues embedded in historical documents. It is also multilingual, preserving the semantic integrity of original texts without relying on translation, which can alter meaning, an especially important feature when analyzing decades-old reports from countries like Mexico, Chile, or Bulgaria.

What the Data Reveals: Evolution, Audience, and Asymmetry

With this model in place, the paper maps out a sweeping evolution in central bank communication over the last century. One of the clearest patterns is the rise of forward-looking language, especially since the global adoption of inflation-targeting regimes. Where once central banks focused on exchange rates and retrospective analyses, they now increasingly discuss inflation expectations, interest rate paths, and macroeconomic forecasts. This evolution reflects a broader institutional shift: communication is no longer just a mirror of past decisions, but a proactive tool used to guide market and public expectations.

Audience targeting is another standout finding. Contrary to the assumption that central banks speak uniformly to the public, the paper shows that tone is tailored with remarkable consistency. Financial markets receive neutral, technical messages; households are addressed with confidence-building language; and governments are warned of macroeconomic risks. During crises like COVID-19 or the global financial meltdown, this strategy shifts: messages to governments and banks become more reassuring, while communication with the public takes on a more cautionary tone.

The sentiment analysis uncovers a further asymmetry: dovish (accommodative) communication tends to be longer and more detailed than hawkish (tightening) statements. Advanced economies also show more risk-highlighting behavior, suggesting a sophisticated attempt to prepare markets for potential shocks without pre-committing to specific actions.

Metrics That Measure the Mood and Message

Beyond classification, the authors introduce four novel indices to capture the essence of central bank communication:

• Net Policy Sentiment (NPS) measures the balance between hawkish and dovish tones and is split into forward- and backward-looking components, allowing for a direct textual reading of forward guidance.

• Straightforwardness Index assesses how clearly a document conveys its policy stance. Clarity tends to drop during crises, when central banks hedge their messages to maintain flexibility.

• Explanation Index evaluates how much detail accompanies policy justifications. It spikes during policy tightening, when central banks must defend unpopular decisions.

• Net Confidence Index quantifies the balance between risk-highlighting and confidence-building messages, offering a dual lens for both financial stability and monetary outlook.

Empirically, the forward-looking component of the NPS predicts changes in market-based interest rates, proving its power as a real-world policy signal. Likewise, forward-looking confidence is found to correlate with future market volatility, underscoring the link between language and financial outcomes.

A Global Blueprint for Transparent Policy

The paper is not just an academic exercise; it is a toolkit for accountability and transparency. It equips central banks with a way to benchmark their communication strategies against peers, track how their messaging evolves, and assess whether their signals are being understood as intended. By making the implicit explicit, this LLM-powered framework promises to strengthen trust in central banking institutions at a time when credibility and clarity are more essential than ever.

In turning policy talk into data, the authors have done more than train a model; they’ve built a mirror for modern monetary governance. With communication now recognized as a policy instrument in its own right, this paper may well mark a turning point in how the voice of central banking is heard, measured, and understood.