Stock price time series forecasting using dynamic graph neural networks and attention mechanism in recurrent neural networks

Forecasting stock prices is a challenging problem in financial markets. Traditional statistical time series models often struggle to achieve satisfactory results, while deep learning models have shown promise to provide even better outcomes. These models typically focus on capturing and exploiting temporal dependencies within the time series data. In this study, we propose a novel model that incorporates both temporal and geometric patterns present in stock price time series data. Firstly, we employ the visibility graph theory to derive a graph representation of the stock price time series, enabling us to capture the inherent non-Euclidean relationships within the data, which is important because it could lead to a better representation of the observations. Subsequently, we utilize Graph Neural Networks (GNNs) to analyze and model these relationships effectively. Finally, we incorporate a Long Short-Term Memory (LSTM) network enhanced with an attention mechanism to capture the temporal dependencies. Through numerical evaluation, our proposed model demonstrates superior performance compared to state-of-the-art models in the prediction of stock prices. The integration of a graph-based representation, GNNs, LSTM, and an attention mechanism enables our model to effectively capture both the temporal and geometric patterns of the stock price time series data. This research contributes to the field of stock price prediction by introducing a comprehensive model that is able to combine the strengths of GNNs and the LSTM network to exploit non-linear dependencies and patterns inside a time series to address the limitations current Recurrent Neural Networks (RNNs).