A Comparative Analysis of Machine Learning and Deep Learning Models for Forecasting Vietnam's Export Coffee Prices
Year: 2026
Type: preprint
Abstract: This study benchmarks three machine learning models (Support Vector Regression, Random Forest, and XGBoost) against three deep learning architectures (DNN, LSTM, and GRU) for forecasting the export price of Vietnamese Robusta coffee on the ICE London exchange, using a strict 300-day out-of-sample test window (20 November 2024 to 28 January 2026) covering severe structural breaks. The Gated Recurrent Unit (GRU) is the strongest architecture (MAE = 96.55 USD/ton, MAPE = 2.14%, R² = 0.967), outperforming the best machine learning model, Random Forest (R² = 0.888), and the deeper LSTM (R² = 0.856); Diebold–Mariano tests confirm that GRU's advantage over all five remaining models is significant at the 1% level. The purely feedforward DNN (R² = − 1.665) performs worse than every machine learning model, indicating that sequential inductive bias, rather than network depth, determines forecasting capability under structural breaks. Two methodological issues are reported transparently: an asymmetry in training-data scope between the machine learning and deep learning groups, and a multi-seed robustness check showing that both LSTM (CV(MAE) = 39.52%) and GRU (CV(MAE) = 24.80%) exceed the 15% stability reference threshold. Using the GRU model, the study implements a 7-day recursive forecast with an expanding 95% confidence interval, producing explicit risk-floor and risk-ceiling parameters that exporters can use to structure forward contracts.
Source: Research Square
Authors: Quang Phung Duy, Hai Dang Cao Ha
Institutions: Foreign Trade University, National Economics University
Language: English
Field: Decision Sciences
Domain: Social Sciences
SDG: Zero hunger
Open Access status: green
OpenAlex ID: W7204234080