El Nino, La Nina, and Forecastability of the Realized Variance of Agricultural Commodity Prices: Evidence from a Machine Learning Approach

We examine the predictive value of El Nino and La Nina weather episodes for the subsequent realized variance of 16 agricultural commodity prices. To this end, we use high-frequency data covering the period from 2009 to 2020 to estimate the realized variance along realized skewness, realized kurtosis, realized jumps, and realized upside and downside tail risks as control variables. Accounting for the impact of the control variables as well as spillover effects from the realized variances of the other agricultural commodities in our sample, we estimate an extended heterogeneous autoregressive (HAR) model by means of random forests to capture in a purely data-driven way potentially nonlinear links between El Nino and La Nina and the subsequent realized variance. We document such nonlinear links, and that El Nino and La Nina increase forecast accuracy, especially at longer forecast horizons, for several of the the agricultural commodities that we study in this research.