A Comparative Study of Advance Forecasting Models on Volatile Time Series Price Data

Abstract

 Efficient and reliable forecasting techniques for commodities with volatile price series are indispensable in agriculture dependent country like India. In this context, choosing a universally accepted model for forecasting precisely the price series of commodities like onion is one of the most challenging tasks because of the existence of seasonality, non-linearity and complexity in the data, simultaneously. Time series models like GARCH, machine learning techniques like TDNN, SVM and deep learning models like LSTM, Stacked LSTM and Bi-LSTM have been extensively studied in this research work to judge their performance on volatile weekly price series of onion for two different markets in India. The models were tuned with the training dataset to forecast the values for the next twelve horizons and eventually the forecasted values have been compared with the testing dataset. It was found that deep learning models outperformed the machine learning techniques as well as conventionally used time series models in dealing with the two volatile datasets.