Crude oil is an important resource that plays critical role in the United States of America (USA) energy security and economic stability ( Oladosu et al., 2022 ). This critical role makes the management of crude oil supply chain to the USA an important priority for industry leaders and policy makers. Globally, USA is one of the largest consumers of crude oil as it is used to produce essential petroleum products such as gasoline, diesel fuel, heating oil, and jet fuel. The products of crude oil are used to power various sectors of the US economy. In 2023, about 12.933 million barrels of crude oil were produced in the USA per day (b/d) while about 6.478 million b/d were also imported ( U.S. Energy Information Administration [EIA], 2023 ). While the country production has been increasing in recent years, the US still rely on foreign supply to augment what is refined in the country ( U.S. EIA, 2024 ). The reliance on foreign crude oil makes it necessary for the USA to have a sustainable and efficient crude oil supply chain that can withstand disruptions and ensure a steady flow of imports to meet the country’s energy needs.

The logistics of crude oil supply chain to the US is subjected to different complexities and challenges. For instance, globalization and interconnectivity has made crude oil supply chain to be at risk of disruptions from natural disasters, geopolitical instability, and economic fluctuations ( Golgeci, Yildiz, & Andersson, 2020 ). For example, the ongoing conflict as a result of Russia invasion of Ukraine disrupt the global crude oil supply chain ( Bagchi & Paul, 2023 ). In order to overcome the vulnerabilities of crude oil supply chain to disruption, there has been continuous advocacy for enhancing supply chain resilience and agility ( Dai et al., 2022 ). Organizations and the government are working towards having an accurate predictive method through which disruption can be predicted as to enhance decision making with respect to mitigation strategy ( Tissaoui et al., 2022 ). Such mitigation can help safeguard the continuous flow of crude oil and its derivatives ( Foroutan & Lahmiri, 2024 ). Methods that are traditionally used for management supply chain are reactive in nature. Reactive approaches to vulnerability management are no longer sufficient in today’s volatile, uncertain, complex, and ambiguous (VUCA) environment ( Bird, 2018 ). This has led to an increased interest in predictive analytics and machine learning (ML) as tools for proactive supply chain management ( Foroutan & Lahmiri, 2024 ).

Machine learning models are algorithms that have the capability to improve its predictive performance based on what it learnt from its experience. Example of these models are random forest and Support Vector Machine (SVM). Each of these models have the capability to handle large amount of data in order to generate predictive insights. SVM is defined as a supervised learning method that has the capability to carry out classification, regression and detection of outliers ( Guido et al., 2024 ). Random forest on the other hand is a machine learning approach that reaches as single result after combining the output of multiple decision trees. Both SVM and random forest have the capability to learn from historical data, identify patterns and arrive at new information which can be used to predict future trends and events ( Gunasekaran, Lai, & Cheng, 2008 ). (2) However, SVM have some setbacks, such as interpretability limitation, high sensitivity of noisy data as well as outliers and challenge of selecting the right Kernel ( Sayeed et al. 2024 ). This makes the need for comparing the predictive performance of SVM with other models.

While machine learning models such as random forest and SVM are valuable for making decision. The accuracy of their predictive performance still remains an issue that is receiving continuous attention ( Jaiswal & Samikannu, 2017 ; Mohapatra, Shreya, & Chinmay, 2020 ). The need for accurate predictive method for the logistics of crude oil supply chain cannot be overstated. The resilience and agility of the supply chain industry is dependent on making timely and effective mitigative interventions to disruptions. Traditionally, Autoregressive Integrated Moving Average (ARIMA) has received wider applications when predicting historical datasets such as that of supply chain. The simplicity and effectiveness of ARIMA in dealing with time series data makes it to be widely used for prediction ( Sonkavde et al., 2023 ). ARIMA has however been criticized for not being able to handle non-linear data and complex patterns that are characteristics of crude oil supply chains. In view of this limitations, this study aims to compare the predictive accuracy of ARIMA with that of advanced ML models such as RF and SVM for US crude oil imports.

RF and SVM were chosen over other models because; SVM has the ability to capture non-linear relationships through the use of kernel while RF models non-linearity by incorporating diverse decision trees in addition to being resistant to overfitting ( Guido et al., 2024 ). Other models such as Generalized Autoregressive Conditional Heteroskedasticity (GARCH), Vector Autoregression (VAR) and Long Short-Term Memory (LSTM) can only give moderate accuracy in forecasting. Therefore, the rationale behind the evaluation of the predictive performance of SVM and RF in relation to ARIMA ( Wang et al., 2017 ). In addition, the best performing predictive model was used to forecast U.S. crude oil imports from 2024 to 2034.

The data used for this study are secondary data of the Total US crude oil imports from Saudi Arabia, Venezuela, Iraq, Other Organization of the Petroleum Exporting Countries (OPEC), Other Non-OPEC Countries, Canada, Mexico and Russia ( Figure 1 ). The data contains crude oil supply from the countries and regions from 1973 to 2023 and it was sourced from the Energy Information Administration’s Monthly Energy Review (October 2023). Once the data was obtained, it was made to undergo several stages which includes data cleaning, model training, prediction, and model performance evaluation.

The actual trend and the predicted trend of the SVM, ARIMA and random forest model are visualized in Figure 2 . The visualized trend shows that predicted imports of SVM and random forest mimics the actual trend of the supplied crude oil to the USA than that of the ARIMA model. The evaluation of the predictive performance also shows that SVM (MAE = 265.65, RMSE = 362.91) had the best performance while random forest (MAE = 479.37; RMSE = 620.75) had the second-best performance ( Figure 3 ). The predictive performance of ARIMA was poor (MAE =1670.10; RMSE = 2195.91) as it reinforces the observed poor mimicry of the actual crude oil import to the USA.

Since SVM was the best performing model, it was used to forecast the volume of crude oil import into the USA from 2024 to 2034 and the result is visualized in Figure 4 . Based on the result, it is expected that the US would not be importing higher volumes of crude oil like what was imported between 2000 and 2020.

The predictive performance of SVM, ARIMA as well as random forest in forecasting crude oil import into the USA were compared in the study. Findings of the study suggest that random forest and SVM had better predictive ability than ARIMA model. Studies of Gajewski et al. (2023) and Zhu (2023) also found that the predictive performance of SVM and random forest for non-linear time series data were better than that of traditional statistics such as ARIMA. ARIMA has been criticized to have difficulty in capturing the dynamics that exist in intricate data sets ( Gajewski et al., 2023 ; Zhu, 2023 ). This was evident with this study where ARIMA was unable to mimic the trends of the US crude oil import compared to the performance of Random Forest and SVM. When SVM and random forest were compared, SVM had the best performance. This was in line with the findings of Zhu (2023) that observed SVM outperforms other models in handling variables that are complex such as macroeconomic factors and exogenous influences.

Furthermore, the study of Jo et al. (2023) evaluate the predictive performance of ARIMA, vector autoregression model (VAR) vector error correction (VECM), SVM, RF and k-nearest neighbors (KNN) algorithm (KNN) on oil import. Their findings revealed that VECM and SVM were the best performing model. This made the researcher to recommend that the integration of timeseries model with machine model offers potential for improve predictive performance of crude oil imports. Supporting the recommendation of Jo et al. (2023) were the study of Safari and Davallou (2018) , Ning et al. (2022) and He et al. (2022) that all found a hybrid approach that involve time series model and machine learning model to have enhanced predictive performance.

The findings of this study have several implications in the oil and gas sector where accurate forecasting is needed. Economic dependence on crude oil makes the need to ensure that there is accurate forecast that can make planning for future easy. With the accuracy provided by SVM and Random Forest, informed decision can be made on how to adapt to predicted decline in import from any source region. Furthermore, policy makers could have more information on which policy and regulation can be made to sustain future importation of crude oil. Furthermore, with the poor performance of ARIMA, performance of alternative time series model such as VECM can be examined based on Jo et al. (2023) findings. The prospect of VECM combination with SVM for prediction of oil import as a hybrid approach requires assessment in order to validate the findings of other researchers who have recommended hybrid approach ( Safari & Davallou, 2018 ; Ning et al., 2022 ; He et al., 2022 ).

In conclusion, the study demonstrates that machine learning models, particularly SVM and Random Forest, outperform traditional ARIMA models in predicting U.S. crude oil imports. SVM was the best-performing model, closely mimicking the actual import trends and providing reliable forecasts for future crude oil imports, while ARIMA struggled with accuracy. These findings reinforce the growing shift towards machine learning techniques in time-series forecasting, especially in industries with complex data patterns. The superior performance of SVM highlights its potential for enhancing decision-making, resource optimization, and risk management in the oil and gas sector and beyond. Although, it is recommended that time series models should be integrated with machine learning model such as SVM to determine how enhanced the predictive accuracy can be. The performance of the models can have generalizable implication for countries from USA is importing crude oil from. A more robust data will be needed to test how the prediction can be generalized to other countries that their data were not included in the prediction of this study.