Comparative Analysis of ARIMA and lSTM Predictive Models: Evidence from Russian Stocks

Abstract

The article aims to find the best time series predictive model, considering the minimization of errors and high accuracy of the prediction. The authors performed the comparative analysis of the most popular “traditional” econometric model ARIMA and the deep learning model LSTM (Long short-term memory) based on a recurrent neural network. The study provides a mathematical description of these predictive models. The authors developed algorithms for predicting time series based on the “Rolling forecasting origin” approach. These are Python-based algorithms using the Keras, Theano and Statsmodels libraries. Stock quotes of Russian companies Alrosa, Gazprom, KamAZ, NLMK, Kiwi, Rosneft, VTB and Yandex for the period from June 2, 2014 to November 11, 2019, broken down by week, served as input data. The research results confirm the superiority of the LSTM model, where the RMSE error is 65% less than with the ARIMA model. Therefore, an LSTM model-based algorithm is more preferable for the better quality of time series prediction.

Keywords

• arima
• lstm
• predictive models
• stocks
• analysis
• stock quote prediction
• algorithms