Forecasting The Volatility of Natural Gas Price using Machine Learning: Fundamentals versus Moments

We compare the ability of moments relative to fundamentals in forecasting the volatility of the spot prices of natural gas of the United States (US). In this regard, we obtain either monthly estimates of realized variance or inter-quantile range as metrics of volatility from daily data or Bayesian time-varying parameter quantile regression. The same two approaches are used to also derive corresponding moments. Given that a large number of predictors are involved in our forecasting exercise, we using linear and nonlinear machine learning approaches to efficiently use the information content of the fundamentals and the moments, over the period of 1976:01 to 2024:11. Our results show that while fundamentals matter in accurately forecasting the volatility of natural gas price relative to a benchmark autoregressive model, moments can indeed add in a statistically significant manner to the performance of the wide array of macro-finance, climate, and natural gas related fundamentals considered. This is particularly the case with the nonlinear machine learning model. Since natural gas plays a key role as a primary energy source during the transition to a cleaner energy mix and a sustainable energy future, our findings have important implications for not only investors, but also policymakers.