Implementing a Hybrid VMD–BiTCN–BiGRU Architecture With Intelligent Optimization for Agricultural Commodity Price Prognostication in Gujarat Markets

Abstract

In the agricultural sector, the prices of pulses require adequate forecasting to plan farming activities, store and market, and guide policy-making. However, the pulse price series is highly volatile, seasonal, nonstationary, and nonlinear, and thus conventional linear time series models are not necessarily very effective. In the suggested framework, the raw price data series is decomposed using Variational Mode Decomposition (VMD) into a set of Intrinsic Mode Functions (IMFs), which helps remove noise and retain the multiscale temporal information. The disaggregated sequences are then passed through a Bidirectional Temporal Convolutional Network (BiTCN), which is trained to capture the local and distant temporal signals, and a Bidirectional Gated Recurrent Unit (BiGRU), which is trained to capture the interaction between the two directions of sequences. Key architectural and training parameters are tuned by a smart hyperparameter optimization process, leading to better forecasting quality and generalization. Furthermore, an Explainable AI (XAI) module, which calculates the effect of different price components, arrivals, and seasonal factors based on SHAP feature attributions, is embedded to increase transparency for stakeholders and quantify the contribution of the various price components, arrivals, and seasonal factors to model forecasts. The framework is evaluated on multi-year pulse wholesale price data from the mandis of Gujarat, and standard error measures of ARIMA, ANN, LSTM, and SVM models are used to compare the models. Moreover, an out-of-sample rolling-window validation is performed with a fixed window of 156 observations to test the robustness of the model when market conditions change and under different seasonal and volatility regimes. Based on the empirical results, the proposed model yields a value of 0.9794 for R², 2.1846 for RMSE, 1.1268 for MAE, and 2.347% for MAPE, which are higher than the results of other benchmark models and give more accurate and reliable predictions in the volatile market. Based on these findings, it may be concluded that the proposed hybrid framework is a practical decision-support tool for pulse price forecasting and can be used as a complementary tool to provide interpretable reasons behind the price drivers for direct implication in the risk management of farmers, alignment of the supply chain, and price stabilization policy of Gujarat. Future studies may be conducted to examine the application of this framework to other farm products and areas, and to include additional variables that would impact farms, including weather, global trade, and government policies.