Wavelets Based Artificial Neural Network Technique for Forecasting Agricultural Prices

Abstract

It has been observed that most of the agricultural time series data in general and price data in particular are non-linear, non-stationary, non-normal and heteroscedastic in nature. Therefore, application of usual linear and nonlinear parametric models like Autoregressive integrated moving average (ARIMA), Generalized autoregressive conditional heteroscedastic (GARCH) and their component models fail to capture the variability present in the series. It is also very difficult to extract actual signal from noisy time series observations. In this regard, nonparametric wavelet technique has the advantage of pre-processing the series to extract the actual signal. Optimizing level of decomposition and choosing appropriate wavelet filter can represent the series with high chaotic nature and sophisticated nonlinear structure more effectively. The decomposition can describe the useful pattern of the series from both global as well as local perspective. The wavelet decomposed components can be modeled using Machine Learning techniques like Artificial Neural Network (ANN) to result in wavelet-based hybrid models and eventually, inverse wavelet transform can be carried out to obtain the prediction of original series. The above algorithm has been applied for modeling monthly modal wholesale price of tomato for Burdwan market, West Bengal, India. Haar and D4 wavelet filters have been applied using two levels of decomposition i.e. 3 and 6. The prediction accuracy of the hybrid model is compared empirically with that of ARIMA, GARCH and ANN model and it is observed that hybrid algorithm outperformed the other models.