预测全球石油需求：机器学习技术的应用

The contents of this paper are the authors’ sole responsibility. They do not necessarily represent the viewsAbstractThis study introduces a novel approach to predicting global oil demand by integrating machine learning(ML) techniques to forecast consumption across seven refined oil products and seven key regions. Byaggregating these forecasts, we offer a comprehensive view of global demand trends. The paperexamines the efficacy of ML models in providing robust and accurate demand forecasts. It also providesa transparent and repeatable process to forecast oil demand. A comparison between the extremegradient boosting (XGBoost) model and Neural Hierarchical Interpolation for Time Series Forecasting(N-HiTS) model was conducted to determine which is a more accurate model to forecast demand. Ourcomparative analysis demonstrates that N-HiTS performs better. The accuracy of global oil demandforecasts is pivotal for economic planning and policy making. of the Oxford Institute for Energy Studies or any of its Members.2 The contents of this paper are the authors’ sole responsibility. They do not necessarily represent the views1. IntroductionAccurate forecasting of oil demand is critical for strategic planning. Traditional econometric models,while useful, often struggle to capture the complex dynamics influenced by numerous economicindicators and geopolitical factors. These models typically rely on linear assumptions and may notadequately address the non-linear relationships inherent in oil markets. To address these challenges,our paper introduces a methodology that leverages advanced machine learning (ML) techniques,specifically extreme gradient boosting (XGBoost) and Neural Hierarchical Interpolation for Time SeriesForecasting (N-HiTS), to enhance the precision and reliability of oil demand forecasts.Machine learning models have demonstrated superior performance in various forecasting tasks due totheir ability to handle large datasets and uncover intricate patterns. Recent studies consistently showthat ML techniques outperform traditional econometric methods in time series forecasting by modellingcomplex, nonlinear relationships and handling large datasets. For instance, Hopp (2022)1found thatlong short-term memory (LSTM) neural networks provided better predictive accuracy than Bayesianvector autoregressions (BVAR) for nowcasting US quarterly GDP growth, especially during economiccrises. Deb (2019)2highlighted that models exploiting heterogeneity, such as finite mixture models,yielded more accurate healthcare spending forecasts compared to generalized linear models and log-linear regression. Similarly, Lukong et al. (2022)3showed that long short-term memory recurrent neuralnetworks (LSTM-RNN) models achieved significantly lower mean absolute percentage error (MAPE) inlong-term electricity load forecasting than linear regression models. In financial time series forecasting,Liu et al. (2023)4reported that ensemble methods like Random Forest and LSTM outperformedtraditional econometric models in both accuracy and interpretability. Additionally, Kontopoulou et al.(2023)5 reviewed various applications and concluded that ML algorithms generally surpassedautoregressive integrated moving average (ARIMA) models, particularly in capturing intricate datapatterns, with hybrid models proving most effective. Furthermore, comparative analyses by Oukhouyaand El Himdi (2023)6also illustrate the superior performance of support vector regression (SVR),XGBoost, LSTM, and multilayer perceptron (MLP) in stock market forecasting, with ML modelsgenerally outperforming their econometric counterparts. This superiority is attributed to ML models'ability to learn from data without relying on pre-defined assumptions, allowing them to capture morenuanced and complex relationships. These findings underscore the enhanced accuracy, efficiency, andflexibility of ML models in time series forecasting across diverse domains.In the context of oil demand forecasting, the integration of ML techniques has shown significantimprovements over traditional methods. Studies by Zhu (2023)7and Alkhammash et al. (2022)8havevalidated the effectiveness of ML models in this domain, demonstrating their superior performance incapturing complex patterns in data. Zhu (2023) conducted an AI-based analysis incorporating bothendogenous and exogenous factors, finding that machine learning models significantly improveforecasting accuracy compared to traditional models. Similarly, Alkhammash et al. (2022) usedoptimized multivariate adaptive regression splines (LR-MARS) to predict crude oil demand in SaudiArabia, showcasing the adaptability and precision of ML models in dynamic environments.XGBoost9, a gradient boosting algorithm, is known for its robustness and efficiency. It has beensuccessfullyapplied in domains such as temperature forecasting and precipitation prediction,highlighting its versatility and effectiveness across different predictive tasks. For instance, Singh an