Forecasting of the stock price using recurrent neural network – long short-term memory

Abstract

We employ a recurrent neural network with Long short-term memory for the task of stock price forecasting. We chose three stocks from the same sub-industry: Visa, Mastercard, and PayPal. This paper aims to test the LSTM network's prediction on stock prices and propose the best settings for selected stock price forecasting. The secondary goal is to assess how the settings differed in the case of two highly correlated stocks (Visa-Mastercard year correlation coefficient average: 0.97) and the case of only weak correlated stock (Visa-PayPal correlation coefficient average: 0.39). We tested 117 different settings of LSTM neural networks. The settings differed by the number of epochs/splits (from ten to fifty-eight by the step of four) and the range (minute, hour, and day). Our dataset was the stock price from 1.6.2020 to 15.1.2021. The best performing network has been trained on a 10-day period for Visa and 10-minute for Mastercard and PYPL. However, the differences were negligible, so we did not find the number of epochs as a key setting, unlike in the case of FOREX.