An ARIMA-LSTM model for predicting volatile agricultural price series with random forest technique

Abstract

Machine learning mechanism is establishing itself as a promising area for modelling and forecasting complex time series over conventional statistical models. In this article, focus has been made on presenting a machine learning algorithm with special attention to deep learning model in form of a potential alternative to statistical models such as Autoregressive Integrated Moving Average (ARIMA) and ARIMA-Generalised Autoregressive Conditional Heteroscedasticity (GARCH) models. Further, an improved hybrid ARIMA-Long Short-Term Memory (LSTM) model based on the random forest lag selection criterion has been introduced. ARIMA model has been used to estimate the mean effect and the GARCH model is employed with the residuals obtained from the ARIMA model to estimate the volatile behaviour of the series. ARIMA-GARCH models act as superior statistical models over ARIMA models based on the lowest AIC and BIC values. LSTM model is employed on all normalised training data series. After which we built a comparison scenario independently between ARIMA, ARIMA-GARCH, LSTM and ARIMA-LSTM models on forecasting accuracy in terms of the lowest RMSE, MAPE and MASE values. The proposed random forest-based ARIMA-LSTM model proved its superiority over the conventional statistical models with an improvement to the tune of 8–25% for RMSE, 2–28% for MAPE and 2–29% for MASE. The proposed hybrid model has been successfully applied to volatile monthly price indices of pulses namely gram, moong and urad. This piece of work will enrich the literature on machine learning and further intrigue researchers to apply it to various other volatile data sets.