A hybrid framework with quantum-inspired feature selection and hyperparameter optimization for rice price forecasting

Introduction Rice price volatility in India is driven by temporal dependence, climatic variability and supply-side uncertainties, creating continuous challenges for farmers, traders and policymakers. This study proposes a hybrid forecasting framework for rice price prediction that combines quantum-inspired feature selection and hyperparameter optimization with time-series decomposition. Methods A unified dataset was constructed by integrating daily rice market prices (2001-2023) with climatic variables, agricultural supply metrics and spatial information from 23 Indian states and 602 districts. Seasonal-Trend decomposition using Loess (STL) split the price series into trend, seasonal and residual components to capture the structure of price variations. An eXtreme Gradient Boosting (XGBoost) model forecast the trend, and a Long Short-Term Memory (LSTM) network modeled the seasonal-residual components. A Quantum Genetic Algorithm (QGA) performed feature selection, and Quantum Particle Swarm Optimization (QPSO) tuned hyperparameters (Hybrid QGA-QPSO). Forecast accuracy was assessed with the Diebold-Mariano (DM) test, residual diagnostics and SHapley Additive exPlanations (SHAP). Results QGA reduced the original 47 variables to 11 optimal predictors, and QPSO ensured stable hyperparameter convergence. The Hybrid QGA-QPSO model achieved RMSEs of 164.57, 242.59 and 294.07 Rs./qtl, with R 2 of 0.9662, 0.9250 and 0.8866 at 1-week, 4-week and 12-week horizons, respectively. It recorded the lowest RMSE at all three horizons among QGA-configured models, with a relative R 2 decline of 8.24% across horizons. Light Gradient Boosting Machine (LightGBM) recorded the lowest 1-week RMSE (236.27 Rs./qtl) among full-feature models. The DM test confirmed a statistically significant improvement of the hybrid model over the baseline (DM = 5.0108, p < 0.001). Ljung-Box and Autoregressive Conditional Heteroscedasticity-Lagrange Multiplier (ARCH-LM) tests showed that forecast errors were free of autocorrelation and heteroscedasticity. SHAP analysis identified price lags as the strongest predictors, with climate influencing prices indirectly through crop supply. Discussion No single model dominates across all horizons and metrics. The proposed framework provided competitive and interpretable multi-horizon forecasts and can support market monitoring, procurement planning and post-harvest decision-making for paddy farmers in Indian rice markets.