Fewer Inputs, Better Forecasts: Can AI Help India’s Rice Farmers Make Better Choices?

A clearer view of coming prices could make those decisions less uncertain, especially when weather problems and market disruptions arrive together.

For a rice farmer deciding whether to sell a harvest or pay for storage, an unexpected price change can reshape the earnings from months of work. Traders and public agencies face related pressures when planning purchases, moving grain and keeping supplies available. A clearer view of coming prices could make those decisions less uncertain, especially when weather problems and market disruptions arrive together.

A study published in Frontiers in Artificial Intelligence explores a forecasting system built around the complicated behaviour of India's rice markets. The paper, "A hybrid framework with quantum-inspired feature selection and hyperparameter optimization for rice price forecasting," combines historical prices, climate information and agricultural supply data. Researchers Lourdeena J. S., Graceline Jasmine S. and Febin Daya J. L. developed the approach to predict prices one, four and twelve weeks ahead.

The system produced its strongest results for near-term forecasts, with errors increasing as predictions reached further into the future. Recent prices supplied the strongest clues, and weather measurements added information that helped shape individual predictions. The findings describe a promising research tool for market planning, with practical benefits for farmers still requiring testing in everyday use.

Building a forecast from India's varied rice markets

The researchers assembled records spanning 2001–2023, covering 602 districts across 23 states and territories. Their sources included AGMARKNET market prices, NASA POWER climate records and government agricultural data, bringing price, weather, supply and location information into a shared dataset. Weekly forecasts drew on this broader picture rather than treating rice prices as an isolated sequence of numbers.

Training used data from 2001–2017, model settings were checked against 2018–2020, and the final evaluation used 2021–2023. Keeping these periods in chronological order helped test whether patterns learned from earlier years could support forecasts for later conditions. The researchers compared the proposed system with statistical methods, machine-learning models and neural networks under the study's specified configurations.

A technique called seasonal-trend decomposition separated the price series into its underlying direction, repeating seasonal patterns and remaining fluctuations. XGBoost handled the trend, and a Long Short-Term Memory network, or LSTM, learned patterns in the seasonal and residual components. Combining their predictions gave the system a way to represent gradual market movement and less regular changes within one forecast.

Choosing useful information made the difference

The original dataset contained 47 candidate variables, creating a need to identify which inputs were most useful together. A Quantum Genetic Algorithm selected 11 predictors, and Quantum Particle Swarm Optimization adjusted the model's settings to improve forecasting performance. These are quantum-inspired search methods running on ordinary computer processors, with no quantum computer involved.

The selected inputs included six measures of recent prices and moving averages, four climate measures and the month of the year. An explanation method called SHAP attributed roughly 89% of the model's prediction influence to price-related inputs, with the previous week's price carrying the greatest weight. Climate inputs contributed roughly 11%, led by maximum temperature and temperature range, alongside rainfall and solar radiation.

Those contributions describe how the model formed predictions, without proving that a particular weather change caused a price movement. The authors suggest climate may influence prices through effects on production and market supply, and they acknowledge that related predictors can share or redistribute apparent importance. The findings give recent price history a central role and place weather information in a smaller supporting role.

What the forecasting results actually showed

For one-week forecasts, the proposed model recorded a root mean squared error of ₹164.57 per quintal, a measure that gives greater weight to bigger mistakes. Its mean absolute percentage error was 3.87%, and its R² score was 0.9662, indicating a close fit to observed price variation in the test data. Four-week and twelve-week errors rose to ₹242.59 and ₹294.07 per quintal, with R² scores falling to 0.9250 and 0.8866.

The hybrid model using all 47 inputs and default settings recorded a one-week error of ₹281.42 per quintal, showing that combining models alone did not deliver the best result. LightGBM led the full-feature baseline group at ₹236.27, and an alternative feature-selection approach paired with the same tuning method reached ₹223.87. The proposed system achieved the lowest root mean squared error across all three horizons among the QGA-configured models, with no model leading every metric.

A forecast-comparison test supported the improvement over the standalone XGBoost baseline, and diagnostic tests detected no clear repeating error pattern or changing conditional error variance. Regional results varied, with R² exceeding 0.95 in 21 of the 23 areas and weaker performance in Uttar Pradesh and Rajasthan. Separate historical stress tests used the untuned, full-feature hybrid architecture, finding the largest errors around overlapping disruptions such as COVID-19 and export restrictions.

Making the results useful beyond the research paper

The researchers propose one-week forecasts for market monitoring, four-week forecasts for procurement and storage planning, and twelve-week forecasts as guidance about likely direction. Occasional large mistakes during disruption periods show why a forecast should carry a clear indication of uncertainty. Actual gains in farm income, storage decisions or procurement costs were not established by this study.

The system lacked explicit indicators for support-price announcements, export restrictions and procurement volumes, limiting its ability to represent policy changes directly. Differences in reporting quality and connections between neighbouring markets could affect performance, and observational data cannot establish cause and effect. The optimization comparison used one random seed, making its reported 23.6% runtime saving over conventional particle swarm optimization a finding that needs repeated testing.

Future work could add policy indicators, real-time market arrivals, seasonal weather forecasts and models adapted to individual states or local markets. Forecast ranges that communicate uncertainty would make the results easier to use responsibly, particularly over longer periods. The study's practical promise rests on turning carefully selected information into useful estimates that can be tested against the decisions farmers and market planners actually make.