Abstract:
Accurate price forecasting is a crucial focus for investors, market traders, and researchers in both academic and industrial contexts. Financial time series forecasting is inherently challenging due to its chaotic and nonlinear dynamics. These complexities often hinder the
effectiveness of traditional forecasting methods, necessitating innovative approaches. This
study investigates the effectiveness of deep learning models in addressing these complexities, proposing a stacked Bidirectional Gated Recurrent Unit (BiGRU) model for predicting one day headed prices in both stock and cryptocurrency markets. By analyzing historical price data of stock and cryptocurrency markets, the study investigates the stacked architecture of BiGRU model’s capability to capture the intricate patterns and dependencies inherent in financial time series. Comparative analyses were performed against two widely used deep learning models, Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (BiLSTM) and it has used the stacked architecture of both models. Historical price data were sourced from four datasets: Bitcoin and Cardano for cryptocurrency markets, and Google and Yahoo for traditional stock markets. Performance was quantified using robust metrics, including Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE), to provide a comprehensive assessment of predictive accuracy. The findings indicate that the BiGRU model consistently surpassed the performance of LSTM and BiLSTM models, demonstrating superior accuracy across all datasets, highlighting its ability to temporal dependencies effectively. These results underscore the BiGRU model's capacity to effectively handle the nonlinear and nonstationary characteristics of financial data. The research leveraged the Google Colab platform for model development and evaluation, ensuring both efficiency and scalability. to capture spatial data features for multivariate time series. Moreover, extending the application of BiGRU to datasets with broader characteristics, such as high-frequency trading data, could further refine its accuracy and robustness. These advancements aim to equip financial analysts and researchers with more reliable tools to navigate complex market dynamics, ultimately contributing to the evolution of financial forecasting methodologies.
