TIME SERIES FORECASTING USING A NEURAL NETWORK WITH SEQUENTIALLY CONNECTED LSTM BLOCKS

Abstract

Improving time series forecasting methods is an important task for many industries, such as finance, manufacturing, military, medicine, and energy. The article focuses on analysing recurrent neural network models with LSTM blocks for time series forecasting. The use of recurrent neural networks with LSTM blocks is particularly relevant, as they allow for effective consideration of long-term dependencies in the data. However, optimal LSTM architectures and parameters such as the number of blocks and the Dropout level remain understudied. Development and optimization of the morphology of a recurrent neural network for time series forecasting using LSTM blocks, integration of EMA (exponential moving average), RSI (relative strength index), and investigation of the influence of various model parameters on forecasting accuracy. In this work, the Nadam optimizer method is applied to optimize the morphology of LSTM, which includes comparing models with different configurations of LSTM blocks (300–350 blocks), Dropout parameters. Google stock financial data (GOOGL) collected using the yfinance library and accuracy assessment metrics such as MSE, RMSE, MAE, MAPE were used for training. An optimal LSTM architecture for time series forecasting is proposed, which considers the use of technical indicators and the Nadam optimizer. The findings highlight the efficiency of using advanced LSTM architectures for financial time series forecasting and contribute to the development of precise and robust neural network solutions for stock price prediction. The study showed that the model with 350 LSTM blocks and Dropout 0.05 achieved a minimum error of 1.64% MAPE, which is smaller than the results of previous studies. The proposed morphological and architectural solutions can be used to forecast stock prices, sales volumes and tasks in various industries where time series forecasting, and analysis are required.