Nonlinear Macroeconomic Granger Causality: An ANN Input Occlusion Approach on MSSA-Denoised Data

This paper introduced a novel methodology for measuring nonlinear Granger causality in macroeconomic time series by combining Multivariate Singular Spectrum Analysis (MSSA) for data denoising with Artificial Neural Network (ANN) input occlusion for causal inference. We applied this framework to five countries, analyzing key macro-financial variables, including yield curve latent factors, equity indices, exchange rates, inflation, GDP, and policy rates. MSSA enhanced data quality by maximizing signal-to-noise ratios while preserving structural patterns, resulting in more stable ΔMSE values and reduced error variability. ANNs were trained on MSSA-denoised data to predict each target variable using lagged inputs, with input occlusion evaluating the marginal predictive contribution of each input to derive causality p-values. This approach outperformed traditional VAR-based Granger causality tests, identifying 38 significant causal relationships compared to 24 for VAR. Cross-country analysis of variables revealed differences in transmission mechanisms, monetary policy effectiveness, and growth-inflation dynamics. Notably, feature importance rankings showed that policy rates and stock market indices predominantly drove macroeconomic outcomes across countries, underscoring their critical role in economic dynamics. These findings demonstrated that combining MSSA and ANN input occlusion offered a robust framework for analyzing nonlinear causality in complex macroeconomic systems.