Forecasting Iran’s Saffron Export by Comparison of Machine Learning Algorithms

Abstract

Imports and exports play an integral role in the economic growth of all countries. Therefore, selecting the right products can enhance a country's competitiveness in global trade. Saffron stands out as one of Iran's most vital and unique non-oil products for export. The objective of this study was to predict saffron exports using three data mining algorithms and determine the most suitable algorithm for forecasting. The sample period for the forecasting models encompasses saffron export data from Iran for the years 2012 to 2019, gathered from the Iran Saffron Association. Following the data preparation steps, saffron export was forecasted using three data mining algorithms: artificial neural network, deep learning, and gradient boost tree. The validity of the models plays a crucial role in selecting the best forecasting model. The predictive validity of the three designed models was evaluated using the absolute error (artificial neural network = 0.036, deep learning network = 0.031, and gradient boost tree = 0.047), R-squared (artificial neural network = 0.045, deep learning network = 0.044, and gradient boost tree = 0.073), and correlation coefficients (artificial neural network = 0.95, deep learning network = 0.98, and gradient boost tree = 0.97). Based on the findings, all models demonstrate high accuracy, with very low prediction errors that are closely matched. However, the deep learning network exhibits a slightly lower, albeit statistically insignificant, error. These results can be valuable for enhancing the precision of saffron export planning.