Comparative Analysis of Machine Learning Models to Identify Key Features of Data for High Export Revenue in Pars Special Economic Energy Zone

Objective: In the national economy of Iran, special economic zones play a vital role as engines driving foreign trade and foreign exchange. Among them, the Pars Energy Special Economic Zone, focusing on the export of high-value products such as petrochemical products, gas condensates, and petroleum products, has an unparalleled position. However, not all exports from this region are equally effective in foreign exchange earnings, and identifying the factors that lead to the creation of high foreign exchange value can lead to the optimal allocation of resources, the regulation of incentive policies, and ultimately, the maximization of the country's foreign exchange earnings. Understanding this necessity, this study was designed and implemented with the aim of identifying the key factors affecting high-value exports in this strategic region in a scientific and data-driven manner.

Method: The statistical population of the study includes all export transaction data of the Pars Energy Special Economic Zone in the six-year period from 2018 to 2024, which amounts to 35,000 real records. This extensive data was used to train and test predictive models. After the data preprocessing stage, three powerful and widely used machine learning algorithms including decision tree, random forest and support vector machine (SVM) were used to analyze the data. The created models were evaluated and compared with the accuracy criterion to determine the most efficient method in predicting high foreign exchange earnings under study.

Findings: The results of running the models on real export data indicated the clear superiority of tree-based models. The decision tree model with an accuracy of 85.36% and the random forest model with an accuracy of 85.08% showed the best performance in identifying high foreign exchange earnings. This was while the support vector machine (SVM) model with an accuracy of 73.77% had a weaker performance. The reason for this can be attributed to the nature of the data and the complex distribution of classification boundaries, which linear models such as SVM face challenges in separating. In addition, the Feature Importance analysis in the decision tree and random forest models, which were among the most accurate models, identified three key factors as the most influential variables in predicting high exchange rate: the name of the product (product type) was identified as the most important factor, and the destination country of the export market was the second most important factor. Factors such as geographical distance, trade conditions and bilateral agreements, economic stability of the destination country, and ability to pay directly affect the transaction value. Exporting to specific markets can bring much greater profitability. Packaging Efficiency This finding, contrary to the popular belief that packaging is only a marginal factor, shows that the standard and optimality of packaging (in terms of maintaining quality, reducing waste, complying with international standards, and reducing logistics costs) has a direct and significant impact on the final value of exports.

Conclusion: This research empirically demonstrated that machine learning algorithms, especially tree models, have a high ability to identify hidden patterns in large export data and predict factors affecting high foreign exchange earnings. The results of this study can be a basis for designing data-driven export policies and focusing on products and target markets with the highest foreign exchange returns.