Masked Self-Supervised Temporal Modeling of Agricultural Prices: A Case Study of Wheat MSP in Rajasthan Region

Support Price (MSP) forecasting under sparse and irregular agricultural time-series conditions. Method: District-level wheat production and price data from all districts of Rajasthan Region, for 2016–2025 were pre-processed and segmented into temporal windows. In this work, we propose a Masked Self-Supervised Temporal learning framework for MSP forecasting that emphasizes representation learning rather than direct label dependence using masked reconstructive temporal learning. Performance was compared with ARIMA and supervised LSTM using mean squared error (MSE) on log-transformed prices. Findings: ARIMA produced the MAE (164.97), while supervised long short-term memory (LSTM) achieved 60.58. The proposed Masked Self-Supervised Temporal Learning (MSSTL) framework achieved a MAE of 32.67 and showed stable convergence and improved data efficiency. MSSTL learned temporal representations capturing seasonal patterns, production–irrigation interactions, and longer-term market memory without relying exclusively on labeled price observations. The results indicate that Masked Self-Supervised Temporal learning can provide a useful balance between forecasting performance and robustness when agricultural datasets are limited. The future-trend analysis emphasizes directional stability rather than exact long-horizon price values. Novelty: The study applies two-stage Masked Self-Supervised Temporal representation learning to wheat MSP forecasting in a data-scarce district-level agricultural setting, reducing dependence on labeled price data. Keywords: Crop Price, Minimum Support Price, Deep Learning, Masked Self-Supervised Learning, Temporal Learning