Abstract
<jats:p>The Equilibrium Real Exchange Rate (ERER) is a latent variable commonly monitored as a potential indicator of external vulnerabilities, yet its estimation is challenged by the non-linear interactions among the underlying fundamentals and by the time-varying nature of their influence. This paper proposes a Neural Network Equilibrium Real Exchange Rate (NN-ERER) model based on a Branch Directed Neural Network (BDNN) architecture that estimates the ERER for the Colombian peso by separating long- and short-run determinants of the real exchange rate (RER). The long-run component is constructed from blocks associated with terms of trade, fiscal policy, productivity, and external debt, while the short-run component is divided into blocks of variables with positive and negative expected effects. Theoretical sign restrictions are encoded into the architecture, constraining the estimation while preserving non-linear flexibility. This structure also allows the RER misalignment to be partially decomposed into short-term determinants and unexplained residuals. The empirical results indicate that the relative contribution of the fundamentals underlying the ERER has evolved over time, with terms of trade and external debt particularly relevant during the period of high commodity prices in the mid-2000s, and fiscal factors gaining weight in the post-COVID-19 period. Misalignments in 2008–2009 and after 2021 are largely attributable to short-run factors, plausibly linked to global volatility, uncertainty, and shifts in interest rate differentials.</jats:p>