Applying economics-guided machine learning models to soybean oil price forecast

Purpose Forecasting soybean oil prices is difficult because price movements emerge from interacting supply, demand, inventory, trade and macro-financial forces, especially in import-dependent markets where external shocks can rapidly alter conditions. This study develops a comparative framework for integrating economic theory into physics-informed neural networks (PINNs) for soybean oil price forecasting. Design/methodology/approach These constraints are examined through five artificial neural network (ANN)-based models: a baseline ANN, PINN-A, PINN-B, PINN-C and a unified PINN-ABC model, evaluated not only for predictive performance but also for forecast robustness, interpretability and structural consistency. Findings Embedding economic theory into PINNs improves forecasting, but its effectiveness depends on the type of constraint imposed. Constraints that directly encode economically interpretable directional relationships contribute most, while combining multiple constraints creates trade-offs between structural richness and optimisation complexity. Research limitations/implications Economics-guided PINNs offer a promising bridge between flexible neural forecasting and theory-consistent commodity market analysis, supporting forecasts that are accurate, economically credible and decision-relevant. Originality/value Rather than treating forecasting as a purely data-driven task, the study conceptualises economic theory as the analogue of “physics”, embedding structurally meaningful economic relationships directly into the neural learning objective as soft constraints. Three forms of economic structure are incorporated: bounded short-run adjustment and temporal smoothness, equilibrium correction and monotonic responses to key fundamentals.