Short-term forecasting of natural gas prices using machine learning and feature selection algorithms

Abstract

We present the results of short-term forecasting of Henry Hub spot natural gas prices based on the performance of classical time series models and machine learning methods, specifically; neural networks (NN) and strategic seasonality-adjusted support vector regression machines(SSA-SVR). We introduce several improvements to the forecasting method based on SVR. A procedure for generation of model inputs and model input selection using feature selection (FS) algorithms is suggested. The use of FS algorithms for automatic selection of model input and the use of advanced global optimization technique PSwarm for the optimization of SVR hyper parameters reduce the subjective inputs. Our results show that the machine learning results reported in the literature often over exaggerate the successfulness of these models since, in some cases, we record only slight improvements over the time series approaches. We have to emphasize that our findings apply to Henry Hub, a market which is known among traders as the “widow maker”. We find definite advantages of using FS algorithms to preselect the variables both in NN and SVR. Machine learning models without the preselection of variables are often inferior to time-series models in forecasting spot prices and in this case FS algorithms show their usefulness and strength.