Comparative analysis of the forecasting quality of the classical statistical model and the machine learning model on the data of the Russian stock market

Abstract

The main objective of this work is to compare the predictive ability of the classical machine learning model — ARIMA, as the most common and well-studied baseline model, and the ML model based on a sequential neural network — in this case, LSTM. The goal is to maximize accuracy and minimize error — selecting the most appropriate model for predicting time series with the highest accuracy. A description is given for these mathematical models. An algorithm is also proposed for forecasting time series using these models, based on the «Rolling window» approach. Practical implementation is implemented using the Python programming environment with the Pandas, Numpy, pmdarima, Keras, Statsmodels libraries. To train the models, we used stock data at the closing price per share of the leading Russian companies: Yandex, VTB, KamAZ, Kiwi, Gazprom, NLMK, Rosneft, Alrosa for the period. The studies carried out demonstrate the predictive superiority of the approach based on neural networks, while the RMSE is 71% less than the same indicator for the ARIMA model, which allows us to conclude that the use of the LSTM model is preferable for this class of problems.