Forecasting of time series using a neural network with parallel-stacked LSTM blocks
Information Technology and Computer Engineering · 20.04.2026 · (23) · 72-82
DOI: 10.31649/vitce/1.2026.72 ↗Kurz gefasst
Fachartikel in der Zeitschrift Information Technology and Computer Engineering, online erschienen am 20.04.2026. Verfasst von Yurii Futryk und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Time series forecasting is crucial for supporting decisions in financial analytics, where data is characterised by non-linearity, non-stationarity, and high noise levels. The purpose of the study was to determine the effective configuration of a recurrent neural network with a parallel combination of Long Short-Term Memory (LSTM) cell stacks to improve the accuracy of stock price forecasting, and the possibilities of applying the back-end model in industry, energy, and related domains. The study applied deep learning methods using the TensorFlow/Keras library, and used historical data from Google shares to train the model. It was established that the architecture with parallel-stacked blocks provided higher learning stability compared to standard recurrent models due to more efficient allocation of technical features of the time sequence. It has been experimentally proven that the optimal number of neurons in the hidden layers for such a task was 100-200 units, while a further increase in the power of the model lead to a retraining effect. It was found that the use of dropout regularisation in the range of 0.1-0.2 minimised the error in the validation sample, while values over 0.3 significantly slowed down the convergence of the algorithm. Feature analysis showed that integrating an exponential moving average with a short time window improved the model result, showing a higher correlation with the target index than the relative strength index. The prediction quality of the model was evaluated by the Mean Squared error (MSE), the Root Mean Squared Error (RMSE), and the Mean Absolute Percentage Error (MAPE). It was found that configurations (50-100 blocks) were characterised by increased MAPE values, while in the range of 180-400 blocks the error decreased and became stable. The most accurate result was obtained for a configuration with 325 blocks, Dropout regularisation = 0.05 and Nadam optimiser (Nesterovaccelerated Adam): MAPE=1.62%, RMSE=2.41, MSE=6.05. The practical significance of the study lied in the formulation of clear recommendations for setting up hyperparameters of LSTM models for applied short-term forecasting of financial series
Zitieren
APA
Futryk, Y., & Peleshchak, I. (2026). Forecasting of time series using a neural network with parallel-stacked LSTM blocks. Information Technology and Computer Engineering, 72-82. https://doi.org/10.31649/vitce/1.2026.72
Vancouver
Futryk Y, Peleshchak I. Forecasting of time series using a neural network with parallel-stacked LSTM blocks. Information Technology and Computer Engineering. 2026:72-82. doi:10.31649/vitce/1.2026.72
BibTeX
@article{Futryk2026,
author = {Futryk, Yurii and Peleshchak, Ivan},
title = {Forecasting of time series using a neural network with parallel-stacked LSTM blocks},
journal = {Information Technology and Computer Engineering},
year = {2026},
number = {23},
pages = {72--82},
doi = {10.31649/vitce/1.2026.72}
}
EUVIMED