Accurate forecasting of cotton prices is crucial for farmers and stakeholders in the agricultural sector to optimize crop selection, improve profitability and mitigate market risks. This study had developed a novel weather-based deep learning model to predict cotton prices in two major cotton-growing districts of Tamil Nadu, Perambalur and Salem. Despite the non-linear relationship between weather variables and price, advanced deep learning techniques were employed to uncover hidden patterns and enhance predictive accuracy. Since the price series was non-stationary, Seasonal-Trend-Residual decomposition using Loess decomposition was done to separate trend, seasonality and residual components and distinct models were fitted to each component. Four weather parameters-maximum temperature, minimum temperature, relative humidity and rainfall-were considered as exogenous variables. Feature selection was performed based on the mutual information score. Various deep learning architectures like STL-ANN, STL-TDNN, STL-GRU and STL-LSTM were explored to assess their effectiveness in forecasting prices for each decomposed component and finally ensembled together. The results demonstrated the potential of incorporating weather data into predictive models for cotton price forecasting, with the LSTM model outperforming other three models with MAPE of 3.68 % and 4.07 % in Salem and Perambalur districts respectively. The study highlights the potential of LSTM-based models in supporting informed decision-making and improved crop planning for cotton farmers in these regions.
