MACPGANA Under Data Scarcity: Single-Commodity Validation of a GAN–Autoencoder–Q-Learning Framework for Agricultural Price Forecasting

Abstract

Accurate short-term forecasting of agricultural commodity prices is critical for protecting farmer incomes, guiding traders’ inventory decisions, and informing policy interventions. However, most deep learning forecasting frameworks are validated only on large, multi-commodity datasets, leaving their behaviour under the data-scarce, single-commodity conditions typical of regional agricultural markets largely unexamined. This study addresses that gap by evaluating MACPGANA — a hybrid framework integrating multimodal feature extraction, Generative Adversarial Network (GAN)-based feature selection, autoencoder-based price prediction, and Q-Learning-driven adaptive optimisation against three established baseline architectures (Multilayer Perceptron, Convolutional Neural Network, and Long Short-Term Memory) under a constrained single-commodity setting. All four models were trained and evaluated under identical conditions on 365 days of daily wholesale tomato price data from the Raipur Agricultural Produce Market Committee, Chhattisgarh, India (2018), using a chronological train–test split. MACPGANA achieved a Mean Absolute Percentage Error of 1.66% and a normalised RMSE of 0.0128 (₹29.38/quintal), corresponding to MSE reductions of 93.0–97.0% relative to the baselines. Under a ±10% tolerance-band classification criterion, MACPGANA achieved 94.59% accuracy and an F1-score of 97.22%, against 29.73–43.33% accuracy for the baselines, with directional accuracy of 91.67% versus 0–8.33%. These results show that a framework engineered for large-scale, multimodal price prediction retains substantial predictive advantage when applied to a single commodity with limited historical data, supporting its applicability to resource-constrained regional markets.