Performance comparison of wavelets-based machine learning technique for forecasting agricultural commodity prices

Abstract

Accurate forecasting of various phenomenon has got crucial importance in the scenario of Indian agriculture as this helps farmers, policy-makers and government to acquire informed decisions. Agricultural time series datasets are mostly nonlinear, nonstationary, non-normal and heteroscedastic in nature. Though the stochastic model like autoregressive integrated moving average and its component models have gained much popularity in modeling linear dynamics, they fail to capture the nonlinearity present in the series. Machine learning (ML) techniques like artificial neural network (ANN) has rapidly emerged within the area of forecasting to take care of nonlinearity in the dataset. But, the presence of high chaotic nature and sophisticated nonlinear structure of the series sometimes distorts the particular model specification. Therefore, preprocessing of the series is required to extract the actual signal in it. Wavelet transformation may be an efficient tool in this scenario. The decomposed and denoised components through wavelet transformation can be modeled using ANN to make wavelet-based hybrid models and eventually, inverse wavelet transform is carried out to obtain the prediction of original series. The incontrovertible fact is that these hybrid models handle nonstationary, nonlinear and non-normal features of datasets simultaneously. The present study discusses the above approach envisaging monthly wholesale tomato price of three major markets in India, namely Ahmedabad, Burdwan and Madanapalli. The improvement over conventional techniques is obtained to a great extent by using wavelet-based combination approach with ML technique as exhibited through empirical evidence.