What drives central bank digital currency implementation? A machine-learning analysis using support vector machines and SHAP explainability

Zhanat Khishauyeva, Ситенко Диана Анатольевна, Vitaliia Koibichuk, Arsen A. Petrosyan, Gaukhar Kodasheva, Ekaterina Dmitrieva, Kseniia Mohylna

Banks and Bank Systems, 2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.21511/bbs.21(1).2026.13
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Type of the article: Research ArticleAbstractCentral bank digital currency (CBDC) programs have rapidly shifted from experimentation to policy-critical infrastructure decisions, yet countries show strikingly uneven progress from research to pilots and implementation. This study aims to identify and explain the key structural, macroeconomic, technological, and ecosystem-related factors that differentiate CBDC initiatives advancing to pilot or implementation stages from those remaining in early research or being discontinued across countries worldwide. Using 161 CBDC projects across 109 countries (as of December 2024) and 10 project-, public interest-, technology-, and macroeconomic indicators, we estimate a Support Vector Machines classifier with GridSearchCV (5-fold) tuning and interpret the results using Shapley Additive exPlanations explainability. The raw outcome distribution was strongly imbalanced (83.85% in the early/cancelled class), so ADASYN balancing was applied, producing 270 observations with equal class shares and an 85/15 train–test split (229/41). The optimized SVM (RBF; C = 10, gamma = 10) achieved 93.90% cross-validated accuracy and 0.88 accuracy on the test set, indicating strong predictive performance on unseen data. Test-set metrics show an informative error profile: for class 1 (advanced projects), recall = 1.00 and F1 = 0.89, while for class 0 (early/cancelled), precision = 1.00 with recall = 0.75 (macro/weighted F1 = 0.88), implying that the model identifies all advanced projects but may misclassify around one-quarter of early/cancelled cases. SHAP ranks the strongest drivers as use-case direction, inflation, crypto adoption ranking, CBDC-related research output, and international participation, with mixed/wholesale projects, higher inflation, stronger scientific attention, and greater international involvement generally increasing the likelihood of advancement.