Onion Price Forecasting with Decision Support System Using a Hybrid Model

Abstract

The correct and timely prediction of the prices of onions in agricultural markets like Kurnool APMC, Andhra Pradesh is necessary not just to avoid losses to the farmers, but also to make sound decisions regarding storing and selling of crops. The historical prices of onion markets available in government market repositories are the main data source in agricultural forecasting but traditional statistical models like ARIMA are not always adequate to model the intricate seasonal variations and abrupt nonlinear price changes that are observed in the onion markets. The primary theme of this paper is a hybrid time-series forecasting framework of onion price prediction with a farmer-facing decision support dashboard. The new system involves two forecasting models, Facebook Prophet and XGBoost, in which Prophet breaks the trend, yearly seasonality, and weekly seasonality into engineered features that are a part of both XGBoost and lag prices, rolling statistics and calendar variables. Several standard preprocessing and feature engineering techniques such as forward fill imputation, outlier elimination, and price momentum calculation are employed to come up with an enhanced generalization of the model. The performance assessment shows that the hybrid model will have a Mean Absolute Error of Rs.112.61 per quintal and a MAPE of 13.29 which is better than Prophet alone and XGBoost alone on all measures of evaluation. To make the hybrid model most useful and practical to assist farmers in decision-making, the most suitable model is implemented in a Flask-based web dashboard with SELL and KEEP recommendation buttons that show the projected price, anticipated earnings, and profit or loss in rupees of any target date chosen by the farmer.