Central Bank Communication Research logo
Research group

Central Bank Communication Research

We study how central banks say what they say — and what that reveals about the economy. Using natural language processing applied to monetary policy statements, minutes, and speeches, we recover sentiment, dissent, and narrative that the official stance alone does not disclose, and test whether those signals carry information for forecasts, markets, and the public.

Our current empirical focus is the Reserve Bank of India’s Monetary Policy Committee (2016–present), with comparative work on other inflation-targeting central banks.

Start here
Disciplinary scope
Economics
monetary policy, expectations, forecasting
Computer Science
natural language processing, text-as-data
Psychology
group dynamics, tone, framing effects
Political Science
institutions, accountability, independence
Method

How the work is done

The pipeline below is deliberately stated in full. Text-as-data results are only as credible as the choices made before the regression, so we set those choices out first.

Stated limitations
  • Sentiment scores are model-derived proxies for tone, not direct observations of intent.
  • The Indian MPC sample is short by macroeconometric standards; estimates are reported with that limitation stated.
  • Association between language and outcomes is not, on its own, evidence of a causal channel.
  1. 01

    Corpus construction

    Policy statements, minutes, individual member statements, and speeches are collected, dated, and matched to the corresponding decision and macroeconomic vintage. Sources and revisions are recorded so the corpus can be rebuilt from scratch.

  2. 02

    Measurement

    Sentiment is scored with several established NLP models rather than one, and reported as an average alongside its components. Disagreement between models is treated as information about measurement uncertainty, not noise to be hidden.

  3. 03

    Benchmarking against the official stance

    Recovered tone is compared with the recorded vote and the stated policy stance. Systematic divergence is the object of study — the gap between what a committee decides and how its members write.

  4. 04

    Identification

    Measured language is related to outcomes — professional forecasts, market prices, subsequent decisions — with the announced decision controlled for, so that any residual association is attributable to communication rather than to policy itself.

  5. 05

    Robustness and replication

    Results are re-estimated across alternative sentiment models, subsamples, and specifications. Derived series are published so that others can reproduce, contest, or extend the findings.

Events calendar

Seminars, conferences, and talks

Where the group is presenting, and the meetings we are convening or taking part in.

Loading events…

Engage with the work

Ideas grow best when challenged

This is an open research programme. The measures are contestable, the sample is short, and the interpretation is ours — all three benefit from scrutiny.

Contest the findings

Referee reports, replication attempts, and objections are welcome and are the fastest route to better estimates. Send comments on any paper or series and we will respond.

Write to us

Request data or code

Corpora, derived sentiment series, and estimation code are shared for academic use. Tell us the study and the intended use and we will send the current version with its documentation.

Request access

Collaborate or present

We work with researchers extending this to other central banks, and take seminar and conference invitations. Students looking for supervision on text-as-data topics may also get in touch.

Get in touch