Forecasting International Crude Oil Price Based On LSTM-(U)MIDAS Hybrid Model

Energy is the material basis for the development of human society and civilization.As an important energy product,the price change of crude oil has gradually become an important factor affecting the stability of the world economy,the development of the national economy and the decision-making of enterprises.In the field of international crude oil price forecasting,the exploration of econometric models and machine learning methods has achieved leapfrog development,but the evolution process of the two methods is independent.Factors affecting international oil prices include differences in data frequency and non-linear relationships.Mixing models and machine learning have unique advantages in dealing with data frequency differences and nonlinear relationships,respectively.In addition,the inclusion of deep learning in the time series of mixed finance has gradually attracted widespread attention.Therefore,this paper combines the MIDAS model and LSTM to construct an LSTM-(U)MIDAS hybrid model that can extract high-frequency market information and accurately predict international crude oil prices.This paper summarizes the framework of influencing factors of international crude oil prices,analyzes the trend of oil prices and the movement of influencing factors,and makes an empirical study of crude oil price based on the LSTM-(U)MIDAS hybrid model.First,the theory and literature of crude oil price prediction are analyzed.Second,the Google search volume index and PCA method are used to extract the investor attention index of the international crude oil market.Third,this paper reviews the historical trend of crude oil price,and analyzes the fluctuation law of the influencing factors from various sources.Fourth,this paper constructs LSTM-(U)MIDAS hybrid model,uses TSCV and Grid Search CV methods to select the key hyperparameters,and applies it to the empirical study of crude oil prices.Fifth,in order to verify the effectiveness of the LSTM-(U)MIDAS model,we discuss the prediction results of ANN-(U-)MIDAS and other models,and performed DM testing on RMSE in a competitive model.Sixth,based on Permutation Feature Importance algorithm,this paper calculates the importance of each input variable in the prediction of crude oil price yield.The key findings of this paper include:(1)The prediction performance of the LSTM-(U)MIDAS model is significantly better than other competing models,and the prediction accuracy of LSTM-MIDAS is slightly better than LSTM-UMIDAS.The model can not only capture the potentially useful information in high-frequency data,but also capture complex nonlinear features and long-term dependent information,and has better applicability.Moreover,the LSTM-MIDAS model is more suitable for predicting international oil price variables with high-frequency mismatch.(2)For crude oil price projections,crude oil inventories and investor attention factors are crucial to oil price forecasts,while alternative energy prices and supply factors contribute relatively little to price forecasts.Weekly and daily high-frequency variables can provide more information for oil price forecasting and improve the ability to forecast oil prices.When analyzing the influencing factors of international crude oil prices,this paper fully considers the influence of fundamental and non-fundamental,market and non-market factors.We innovatively apply models from the fusion of frequency mixing data sampling and deep learning methods to international crude oil price forecasting.We provide certain reference significance in academic research in the field of international oil price forecasting,providing assistance for oil investors to analyze and predict international crude oil prices,and decision-makers to monitor and regulate the crude oil market.