Comparative Analysis of CNN, LSTM And Random Forest for Multivariate Agricultural Price Forecasting

Abstract

Time series forecasting is an important research topic among agriculture economics. Especially, multivariate, multi-step and multiple output prediction tasks pose a challenge in research as their nature requires the investigation of intra- and inter-series correlation. The common statistical methods like ARIMA and SARIMA fall short in this kind of tasks. Deep learning architectures like Convolutional Neural Networks and Long Short-Term Memory networks are quite good at modelling the structures of complex data relations. In this study, a new dataset is composed through manual collection of data from the Ministry of Commerce of Turkish Republic. The dataset contains daily trade volumes and prices of potato, onion and garlic, which are most commonly consumed products in Turkish cuisine. The data pertains to the period between January 1, 2018 and November 26, 2022 (1791 days). A simple CNN and LSTM architectures as well Random Forest machine learning method are used to predict the next 10-day prices of the products. Accordingly, three models provided acceptable results in the prediction tasks, while CNN yielded by far the best result (MAE: 0.047, RMSE: 0.070).