Enhancing Stock Price Forecasting on the Colombo Stock Exchange with Cluster-Based Gated Recurrent Unit Architectures

Abstract

Deep learning algorithms like Gated Recurrent Unit (GRU) networks are increasingly used in financial forecasting, particularly for stock market prediction, due to their superiority in modeling non-linear interactions. GRUs simplify internal gating processes, allowing for more computationally efficient methods for capturing long-term dependencies, enabling accurate temporal modeling, and minimizing the vanishing gradient problem. This paper introduces a novel clustering enhanced GRU model for stock price forecasting on the Colombo Stock Exchange (CSE), addressing shortcomings in existing approaches by employing clustering as a critical pre processing step. By using clustering as a crucial pre-processing step, this work fills gaps in current methods by introducing a unique clustering-enhanced GRU model for stock price forecasting on the CSE. In contrast to earlier models, our approach optimizes clustering by taking into account the noise and volatility of financial data by utilizing a number of similarity metrics, including Euclidean distance (EUD), Dynamic Time Warping (DTW), and the Logistic Weighted Dynamic Time Warping (LWDTW). The distance matrices were computed using CSE daily closing stock price data. The accuracy of stock price projections was subsequently maximized by applying clustering-augmented GRU using the three previously mentioned similarity measurements. The GRU model with clustering augmentation and the GRU model without clustering were thoroughly compared. The results show that the GRU architecture with the LWDTW methodology performs better than other approaches in terms of efficiency and forecasting ability, with the lowest root mean square error (RMSE) value of 0.0189 and the greatest R² value of 0.9545. These findings show how better financial market decision making can arise from using deep learning models and clustering approaches to assess CSE stock values more precisely.