Forecasting Sesame Price in Ethiopia Using Recurrent Neural Network Based Deep Learning Algorithms

Sesame is Ethiopia's second-most important crop after coffee, supporting farmer incomes and contributing to foreign exchange earnings. In 2018, it generated approximately 449 million USD, and in 2010, it accounted for 14% of global exports. Despite its significance, sesame price volatility driven by market fluctuations and policy changes poses challenges for stakeholders. This study forecasted sesame prices using RNN-based deep learning models: LSTM, GRU, Bidirectional LSTM, and Bidirectional GRU. Daily price data from 2012 to 2020, sourced from the Ethiopian Commodity Exchange (ECX), was split into 80% for training and 20% for testing. Model performance was evaluated using RMSE, RRSE, MAE, MAPE, and R-squared metrics. To assess the effectiveness of RNNs, their results were compared to statistical and traditional machine learning models, including the Autoregressive Integrated Moving Average (ARIMA), Support Vector Regressor (SVR), Random Forest (RF), and Time Delay Neural Network (TDNN). The ARIMA model performed poorly, with an RMSE of 487.203, a MAPE of 6.769, and an R-squared of 70.52%. SVR achieved the highest R-squared among traditional methods (97.02%) but still showed higher RMSE (77.469) and MAE (75.236) compared to the RNN models. RF and TDNN demonstrated moderate performance, but their prediction errors remained above those of the RNN models. Among all models, Bidirectional GRU achieved the best results with the lowest MAPE (7.99%) and highest R-squared (99.65%), as well as the lowest RMSE, RRSE, and MAE, demonstrating superior predictive accuracy for sesame prices. The results demonstrate that RNN-based models, especially Bi-GRU, substantially improve forecasting accuracy. This supports farmers in aligning production with favorable price periods and aids policymakers and traders in making informed decisions that promote market stability and sustainable agriculture.