CBDC Stress Test in a Dual-Currency Setting

Battle Count: 50

The paper is primarily focused on central banking policy, financial stability, and CBDC design rather than quantitative trading strategies. However, it contains relevant information for: (1) FX trading in RON/EUR pairs through its analysis of currency substitution dynamics and exchange rate volatility; (2) Interest rate trading through its VAR analysis of monetary policy transmission; (3) Credit risk assessment for bank stocks through deposit outflow modelling; (4) Macro regime classification (tight/neutral/loose) that could inform tactical asset allocation; (5) Understanding of deposit beta effects on monetary policy. The ML models (XGBoost, Random Forest) and VAR frameworks could be adapted for financial market applications, but the paper's primary contribution is to central bank policy rather than trading strategy development.

The framework is extremely complex, integrating: (1) Synthetic agent generation with 10,000 agents and multi-dimensional behavioural profiles; (2) Multiple ML classifiers (XGBoost, Random Forest, CART, SVM, KNN, Naive Bayes, GBM) with cross-validation; (3) VAR/SVAR/MSVAR econometric models with impulse response analysis; (4) PCA for dimensionality reduction; (5) Monte Carlo simulations with behavioural overlays and confidence bands; (6) Game-theoretic models; (7) Stochastic differential equations; (8) Multi-criteria decision analysis; (9) Liquidity stress simulations across 6+ holding-limit scenarios; (10) Bank balance-sheet adjustment channel analysis for 20 banks; (11) SHAP interpretability analysis; (12) Logistic regression with marginal effects; (13) Markov switching regime analysis; (14) Bayesian game models. The paper spans 700+ pages with 40+ annexes. Implementation requires expertise in econometrics, machine learning, behavioural economics, central banking operations, and financial stability analysis. Calibration of parameters (beta weights, parity ratios, holding limits) requires domain-specific knowledge and institutional data access.

The paper explicitly emphasizes methodological replicability and transferability as its principal value. It provides detailed calibration logic, parameter settings, and sequencing of analytical stages. However, it relies heavily on synthetic datasets (10,000 agents) rather than real micro-level data, and the beta weights in the adoption function are calibrated through expert judgment rather than statistical estimation. The framework is designed to be open and reproducible, but specific synthetic data generation parameters and some calibration choices are not fully transparent. No code repository is provided. The study uses publicly available data (Eurostat, Eurobarometer, ECB) and one section uses anonymized NBR data. The author states AI tools were used for language editing and synthetic agent generation.