Crude oil price forecasting has received considerable attention owing to its significance in the commodity market and non-linear complexity in forecasting tasks. This study aims to develop a novel deep reinforcement learning algorithm for multi-step ahead crude oil price forecasting in three major commodity exchanges. The proposed algorithm includes two main improvements: (a) A dynamic action exploration mechanism based on the stochastic processes conforming to commodity price fluctuations is designed for accuracy and generalization. (b) A dynamic update policy of network parameters based on approximate optimization theory is developed to improve the network's learning efficiency. The algorithm's effectiveness is experimentally verified and compared with five state-of-the-art algorithms. The main findings are as follows. (a) DRL's forecasting ability is developed in crude oil price forecasting, which may be extended to the forecasting of other natural resource prices. (b) The proposed algorithm can be applied to the data of the world's three major crude oil price benchmarks with considerable universality. (c) The accuracy of the proposed algorithm declines indistinctively with the expansion of the forecasting step; however, it reflects the actual price and fluctuation. These findings have implications in accelerating the global economic recovery and exploring AI in the energy market.
Read full abstract