A Random Forest-Convolutional Neural Network Deep Learning Model for Predicting the Wholesale Price Index of Potato in India

The wholesale price index (WPI) is a crucial economic indicator that provides insights into the pricing dynamics of different goods within a country, especially potato commodities. In this study, we tried to build a hybrid machine learning model technique for predicting the volatile price index of potato. We introduced the Random Forest-Convolutional Neural Network (RF-CNN) model to predict agricultural volatility price index commodities. Traditional statistical time series models (Autoregressive Integrated Moving Average (ARIMA) and Generalized Autoregressive Conditional Heteroscedasticity (GARCH)) were also investigated for comparison with machine learning models (Support Vector Machine (SVM), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN)). Because the data set was volatile, the GARCH model outperformed the ARIMA model which had a lower goodness of fit value. The performance of the SVM model was comparable to that of  the statistical models. However, after utilizing an input lag selection strategy based on autocorrelation function (ACF) and RF, the machine learning models outperformed the statistical models. We used LSTM and CNN models with the appropriate input lag feature assessed by ACF and RF. Our findings indicate that the RF-CNN model beats the other models in terms of error accuracy, with improvements of 67% for root mean square error, 95% for mean absolute percentage error, 63% for mean absolute error and mean absolute squared error on the training set, and more than 90% on the testing set for all goodness of fit. Based on the error accuracy, the RF-CNN model can be utilized to better predict the potato price index in the long term. We hope our study will benefit stakeholders and policymakers by providing a realistic potato price forecast. Furthermore, our study contributes to the growing corpus of research on machine learning models for time series.