Time Series and Machine Learning Approaches for Tomato Price Forecasting: Evidence from Major Markets in Karnataka, India

Abstract

This study assessed the behaviour and forecasting of monthly modal tomato prices in major markets of Karnataka using statistical and machine-learning approaches. Monthly price series from January 2010 to July 2023 were examined for Bengaluru, Kolar, Chikkaballapura, Srinivaspura, and Chintamani in the descriptive analysis, while forecasting comparisons were conducted for Bengaluru, Kolar, Chikkaballapura, and Srinivaspura. Descriptive diagnostics indicated non-normality, positive skewness, leptokurtosis, and significant autocorrelation, confirming substantial variation in the market price series. The series used for ARIMA modelling were transformed and first-differenced to attain stationarity. The selected ARIMA specifications were ARIMA (1,1,4) for Bengaluru, ARIMA (2,1,3) for Kolar, ARIMA (2,1,3) for Chikkaballapura, and ARIMA (3,1,1) for Srinivaspura. A Random Forest model was also developed from lagged price variables generated through time-delay embedding, with the first six lags used as predictors and an 80:20 training-testing division. Forecast accuracy was evaluated using RMSE, MAE, and MAPE. Random Forest produced lower RMSE and MAE values across all four forecasted markets, whereas ARIMA yielded lower MAPE values in Kolar, Chikkaballapura, and Srinivaspura. The comparison therefore demonstrates that model performance varies according to the selected forecast-error measure. The findings indicate that the two modelling approaches provide complementary evidence for forecasting volatile tomato prices across the selected Karnataka markets.