Genetic Algorithm Optimized CNN-LSTM Model for Forecasting of Agriculture Commodity Prices*

Accurate price prediction of agricultural commodities is quite challenging due to their dependency on several biotic and abiotic factors that account for their nonlinearity, non-stationarity and randomness. In this study, a very prominent feature-extracting deep learning technique, that is, convolutional neural network (CNN), is integrated with an efficient recurrent neural network forecasting technique, that is, long short-term memory (LSTM), to develop a hybrid model called genetic algorithm optimized CNN-LSTM model (GCLM) to predict any non-stationary and nonlinear agricultural price series. The study compares the price prediction capabilities of the GCLM model with those of the LSTM, Time-delay Neural Network and ARIMA models. The analysis is based on the international monthly price series of palm oil, soybean oil, soybean and maize from January 1980 to November 2022. The genetic algorithm (GA) technique is utilized to investigate and select various hyperparameters of the hybrid model to develop the final optimized GCLM for each series. Several assessment metrics, including root mean square error, mean absolute percentage error, mean absolute deviation and directional statistics, indicate that the GCLM outperforms other models in the experimental data. In addition, the accuracy of the forecast is also tested for significant improvement using the Friedman test and Diebold–Mariano test, whose results validate that the GCLM outperforms other models.