A hybrid structural–machine learning framework: Conventional SVAR and random forest regression

Abstract

This study examines the transmission of copper price shocks to the Peruvian economy through a hybrid empirical framework that combines a conventional Structural Vector Autoregression model with Random Forest regression. Using quarterly and monthly macroeconomic data for 2004–2025, the analysis evaluates how copper shocks propagate through the exchange rate, foreign-exchange intervention, imported inflation, producer prices, consumer inflation, exports, and GDP growth. The SVAR results indicate that positive copper price shocks reduce depreciation pressures and activate foreign-exchange intervention, confirming the relevance of the exchange-rate policy channel. The pass-through to domestic inflation is only partial and heterogeneous across price indices, whereas imported and producer inflation reacted promptly to copper price swings while consumer inflation has remained flat and transitory. As a result, real effects have been positive but very short-lived and supportive of goods and services exports and growth, by raising cyclical terms of trade but not raising productivity and potential growth. SVAR-derived structural shocks, when incorporated into the Random Forest model, improve out-of-sample forecasting performance mainly for real variables such as exports and GDP growth. However, predictive gains for inflation variables are limited. Accumulated Local Effects allow decomposition and map conditional constant marginal effects of shocks to key endogenous variables. This helps to uncover nonlinear responses, relevance of threshold effects and unexpected interactions among copper shocks, depreciation and intervention measures on GDP. In summary, the results obtained confirm that, although copper shocks are relevant for Peru, their effects are conditional, partially absorbed by policy responses and stronger for real variables than for final consumer inflation.