Shortages and machine-learning forecasting of oil returns volatility: 1900–2024

• Shortages to forecast oil returns volatility. • Oil moments estimated using Bayesian TVP-QR. • Machine learning used. • 125 years of data used. • Shortages matter more than oil market moments. The objective of this paper is to forecast the volatility of the West Texas Intermediate (WTI) oil returns over the monthly period of January 1900 to June 2024 by utilizing the information content of newspapers articles-based indexes shortages for the United States (US). We measure volatility as the inter-quantile range by fitting a Bayesian time-varying parameter quantile regression (TVP-QR) on oil returns. The TVP-QR is also used to estimate skewness, kurtosis, lower- and upper-tail risks, and we control for them in our forecasting model along with leverage. Based on the Lasso estimator to control for overparameterization, we find that the model with moments outperform the benchmark autoregressive model involving 12 lags of volatility. More importantly, the performance of the moments-based model improves further when we incorporate the aggregate metric of shortages and its sub-indexes, particularly those related to the industry and labor sectors. These findings carry significant implications for investors.