Nonlinear spillovers in precious metals markets: A TCN-based extension of the Diebold-Yilmaz Framework

Abstract

This study proposes a nonlinear spillover framework for precious metals markets by integrating Temporal Convolutional Networks (TCNs) into the Diebold-Yilmaz approach. We simulate shocks to approximate the Generalized Forecast Error Variance Decomposition (GFEVD) without relying on linear restrictions. Using daily data from 2014-2023 on metal ETFs, futures, oil, equities, exchange rates, and sentiment indices, we uncover strong nonlinear connectedness, with a total spillover index of 95.22%. The VIX and the Shapiro-Sudhof-Wilson (SSW) sentiment index emerge as dominant net transmitters, while rolling analysis reveals shifts in spillovers around major macroeconomic and geopolitical events.