A SHAP-Based Evaluation of Machine Learning Models for CBRT Monetary Policy Decisions

Abstract

This study classifies the monthly interest rate decisions of the Central Bank of the Republic of Turkey between 2010 and 2024 into three categories: increase (hike), decrease (cut), and hold, using machine learning methods, and examines which macroeconomic and financial variables are associated with each decision category in a non-causal, prediction-focused manner. Data are obtained from the Federal Reserve Economic Data System and the electronic data delivery system of the Central Bank of the Republic of Turkey. The performance of the XGBoost, LightGBM, and CatBoost algorithms is compared against the Logistic Regression and Naive Bayes baseline models using a walk-forward validation design that preserves temporal order, with all explanatory variables lagged one month relative to the decision date to reduce contemporaneous information leakage. Class imbalance is addressed through class weighting and an adaptive oversampling procedure, and the contribution of each variable is examined using Shapley value-based explanation analysis. None of the three gradient boosting models achieves a Cohen's Kappa value meaningfully above zero in walk-forward out-of-fold evaluation, and the interest rate increase category is not correctly identified by any model; overall performance is comparable to, and in several cases weaker than, a simple majority class baseline. These findings indicate that interest rate decisions cannot be reliably predicted based on the macroeconomic indicators examined, and they support the role of institutional and communication-related factors that are not reflected in these variables. However, these findings alone do not conclusively prove this role.