Forecasting Short-Term Crude Oil Price Movements using Futures Market Dynamics: Crude Oil Forecasting with Futures Market Properties

Abstract

This paper investigates the predictability of the crude oil spot price within a time horizon of five days, taking into account the information extracted from the term structure of oil futures prices. Two experimental cases are established and examined with historical data from 2000 to 2020: first, as the benchmark, only lagged daily cash price fluctuations are used as input to project price evolution five steps forward; in the second case with the same objective, the study considers both spot prices and the time series of factors obtained through futures curve decomposition by Principal Component Analysis (PCA). In each case, the study implements both an autoregressive time series analysis and a Long Short-Term Memory model (LSTM) to generate the predictions of interest. Simplified grid searches are performed based on validation loss to determine the latter's optimal design and hyperparameter values. With Root Mean Square Error (RMSE) and Coefficient of Determination () as performance evaluation measures, out-of-sample tests lead to a decently encouraging result for the significance of futures market properties as predictors of the spot. That is, potentially, prices of futures contracts aggregate new information that facilitates cash price forecasting. Besides, despite the fact that their accuracy for 1-Day Ahead prediction fails to surpass linear, econometric models, LSTM models demonstrate overall better predicting capability, highlighting implicit nonlinear dependencies within the dataset, and hence serve as reliable approaches to energy forecasting.