CBDC-Discourse

CBDC-Discourse is a BERT-based sentence classifier fine-tuned to categorize central bank digital currency (CBDC) discourse into three conceptually distinct classes: Feature, Risk-Benefit, and Process. This model enables structured analysis of CBDC-related policy and research texts by separating design attributes, evaluative outcomes, and procedural activities.

Class	Description
Feature	A sentence that specifies a concrete design element or operational mechanism of CBDC. Examples include: wallet/card modality; programmability/smart contracts; privacy model; interoperability requirements; legal tender status; distribution via intermediaries; holding limits/caps; interest-bearing/remuneration (incl. negative rates); rulebook/scheme rules; settlement architecture (DLT/RPS/RTGS links).
Risk-Benefit	A sentence that asserts or implies outcomes, effects, or trade-offs (positive or negative) from a CBDC feature or its introduction, including policy/equilibrium impacts. Examples include: faster/cheaper/more transparent cross-border payments; financial inclusion; regional cooperation; competition/innovation; sovereignty/autonomy; efficiency/productivity gains. Also, negative concerns such as bank disintermediation; cyber/operational risk; crisis flight from deposits; privacy harms; monetary/fiscal dominance concerns; “too successful” crowd-out; legal/regulatory fragility.
Process	A sentence about research, consultations, pilots, governance, timeline, or agenda-setting, without specifying a concrete feature or claiming effects/trade-offs. Examples include: public consultations; surveys/focus groups; task forces; phases (investigation/preparation/pilot); rulebook drafting as an activity (absent specifics); reports/citations; statements of interest/attention; open questions; goal/timeline setting (e.g., “medium-term goal”).

Base Model

This classifier is built on top of bilalzafar/CentralBank-BERT, a domain-adapted BERT model pretrained on over 2 million sentences (~66M tokens) from BIS central bank speeches (1996–2024). CentralBank-BERT provides deep contextual understanding of monetary policy, financial regulation, and central banking discourse, making it an optimal foundation for downstream CBDC-related text classification.

Dataset

The model was fine-tuned on a manually annotated dataset of CBDC-related sentences extracted from Bank for International Settlements (BIS) central bank speeches (1996–2024). The dataset was balanced across three discourse classes with a total of 2,886 sentences (962 per class):

Intended Use

This model is designed for the automatic classification of CBDC discourse in policy, research, and financial communications. It enables researchers, analysts, and practitioners to distinguish whether a sentence describes procedural aspects, design features, or evaluative outcomes of central bank digital currencies. Such categorization supports policy analysis, thematic mapping of central bank communication, and structured NLP-based research in the fields of finance, monetary economics, and economic policy.

Training Details

• Tokenization: WordPiece (CentralBank-BERT tokenizer)
• Maximum sequence length: 256 tokens
• Dynamic padding (DataCollatorWithPadding)
• Train/Val/Test split: 80/10/10 stratified by label

Parameter	Value
Base model	bilalzafar/CentralBank-BERT
Epochs	6
Train batch size (per device)	8
Eval batch size (per device)	16
Gradient accumulation	2
Effective batch size	16
Learning rate	2e-5
Weight decay	0.01
Warmup ratio	0.06
Scheduler	Cosine
Mixed precision (fp16)	Enabled

• Environment: Google Colab
• GPU: Tesla T4 (16GB)
• Framework: PyTorch 2.8.0 + Hugging Face Transformers

Evaluation Results

Split	Accuracy	Macro-F1	Weighted-F1	Class	Precision	Recall	F1
Validation	0.851	0.839	0.852	–	–	–	–
Test	0.823	0.803	0.825	Feature	0.759	0.782	0.770
				Process	0.927	0.845	0.884
				Risk-Benefit	0.700	0.817	0.754

---

Other CBDC Models

This model is part of the CentralBank-BERT / CBDC model family, a suite of domain-adapted classifiers for analyzing central-bank communication.

Model	Purpose	Intended Use	Link
bilalzafar/CentralBank-BERT	Domain-adaptive masked LM trained on BIS speeches (1996–2024).	Base encoder for CBDC downstream tasks; fill-mask tasks.	CentralBank-BERT
bilalzafar/CBDC-BERT	Binary classifier: CBDC vs. Non-CBDC.	Flagging CBDC-related discourse in large corpora.	CBDC-BERT
bilalzafar/CBDC-Stance	3-class stance model (Pro, Wait-and-See, Anti).	Research on policy stances and discourse monitoring.	CBDC-Stance
bilalzafar/CBDC-Sentiment	3-class sentiment model (Positive, Neutral, Negative).	Tone analysis in central bank communications.	CBDC-Sentiment
bilalzafar/CBDC-Type	Classifies Retail, Wholesale, General CBDC mentions.	Distinguishing policy focus (retail vs wholesale).	CBDC-Type
bilalzafar/CBDC-Discourse	3-class discourse classifier (Feature, Process, Risk-Benefit).	Structured categorization of CBDC communications.	CBDC-Discourse
bilalzafar/CentralBank-NER	Named Entity Recognition (NER) model for central banking discourse.	Identifying institutions, persons, and policy entities in speeches.	CentralBank-NER

Repository and Replication Package

All training pipelines, preprocessing scripts, evaluation notebooks, and result outputs are available in the companion GitHub repository:

🔗 https://github.com/bilalezafar/CentralBank-BERT

---

How to Use

from transformers import pipeline

# Load pipeline
classifier = pipeline("text-classification", model="bilalzafar/CBDC-Discourse")

# Example sentences
sentences = [
    "The central bank launched a pilot project for CBDC cross-border settlement.", # Process
    "Programmability in CBDC allows conditional payments.", # Feature
    "CBDC may increase risks of bank disintermediation." # Risk-Benefit
]

# Predict
for s in sentences:
    result = classifier(s, return_all_scores=False)[0]
    print(f"{s}\n → {result['label']} (score={result['score']:.4f})\n")

# Example output 
# [{The central bank launched a pilot project for CBDC cross-border settlement. → Process (score=0.9989)}]
# [{Programmability in CBDC allows conditional payments. → Feature (score=0.9991)}]
# [{CBDC may increase risks of bank disintermediation. → Risk-Benefit (score=0.9986)}]

---

Citation

If you use this model, please cite as:

Zafar, M. B. (2025). CentralBank-BERT: Machine Learning Evidence on Central Bank Digital Currency Discourse. SSRN. https://papers.ssrn.com/abstract=5404456

@article{zafar2025centralbankbert,
  title={CentralBank-BERT: Machine Learning Evidence on Central Bank Digital Currency Discourse},
  author={Zafar, Muhammad Bilal},
  year={2025},
  journal={SSRN Electronic Journal},
  url={https://papers.ssrn.com/abstract=5404456}
}