A disagreement-weighted Bayesian dynamic measurement of how central banks talk about digital currency, estimated from 5,718 sentences in BIS central bankers' speeches, 2015 to 2025.
Stance and attention are reported separately and never combined. A central bank that says nothing is silent, not neutral, so it is absent from these means rather than counted as zero. Attention peaked in 2023; stance kept climbing after it.
Mean CSI across institutions with data, with the interquartile range
CBDC sentences per 1,000 sentences spoken, mean across institutions
The quadrant that no single tone score can isolate is the lower right: institutions that talk about CBDC a great deal and remain unpersuaded. Point area is proportional to the number of CBDC sentences behind the estimate.
Each series is a posterior path with its 90% credible band. Solid points are years with CBDC discourse; hollow points are years the institution said nothing, where the estimate is propagated by the random walk and the band widens to say so.
Cell means of the three scored dimensions, standardised across the corpus
Only observed cells are listed. Click a column heading to sort. Sentiment is the mean CBDC-Sentiment valence for that bank-year on its natural scale, from -1 to +1, and Tone is the same quantity standardised across the corpus; they are one measurement on two scales, not two pieces of evidence. Gap is tone minus stance in standard-deviation units: negative means the bank's language runs colder than the position it actually takes. None of these three enter the CSI, which is built from the stance engines alone.
| Institution | Country | Year | CSI | 90% interval | Sentiment | Tone | Gap | Salience | Sentences | Orientation |
|---|
This index measures the position a central bank communicates toward CBDC, not the affective tone of its language. The two are related but they are not the same construct, and the corpus lets us say by how much. A sentence can carry cold language while defending CBDC, and a warm sentence can decline it.
Pairwise correlation on the favourability axis, all 5,718 sentences
The two released encoders agree with each other far more closely than either agrees with an independent reading of the same sentences. They share a base model and overlapping training data, so this is method variance rather than evidence that either has captured the construct.
Mean of (tone minus position), both standardised, by rubric risk salience
Risk-dominated sentences read markedly colder than the position they actually express. The effect is strongly signed but small in aggregate, explaining under three per cent of the variance in the gap; it concentrates almost entirely in the most risk-heavy sentences.
Favourable position, negative affective tone, high risk salience
Rubric urgency inside each class of the released stance model
The whole index is here. Set the filters, choose the columns you need, and take it as CSV. Nothing is sent anywhere: the file is assembled in your browser from the data already on this page.
The index is free to use with attribution. The underlying speeches are the property of the Bank for International Settlements and the issuing central banks and are not redistributed here; the scored corpus carries a source URL for every sentence so any score can be traced back.
Use the recommended institutional citation in APA, Harvard, Chicago, BibTeX or RIS format.
The index is not an average of classifier labels. Sentence scores are treated as noisy indicators of an unobserved stance, and the variance of each indicator grows with how far the scoring engines disagree about it.
Posterior mean with 94% highest-density interval. Weights are estimated, not assigned.
Two independent passes of the rubric over a 322-unit stratified sample