Predictive Analytics in Financial Markets using Time-Series Data with Convolutional Neural Network and Graph Attention Network

In recent years, the financial time-series forecasting has gained much attention because of progress in machine learning and the nature of increasing complexity in market dynamics. However, the existing Autoregressive Integrated Moving Average and Long Short-Term Memory (ARIMA-LSTM) method is efficient to model linear patterns but are incapable of representing non-linear dependencies and abrupt changes in financial data characteristic. To overcome these limitations, this research presents a Convolutional Neural Network-Graph Attention Network (CNN-GAT) model which combines one-dimensional CNN and graph attention mechanisms. First, the data is gathered from financial dataset which consist of spanning stocks, cryptocurrencies, and currency exchange rates. Next, One-Hot Encoding (OHE) is utilized to convert categorical variables into numerical values and min-max normalization is used to scale numerical values. Subsequently, CNN learns high-level temporal-spatial features, where GAT improves inter-temporal dependency with multi-head self-attention. The experimental outcomes demonstrate that the CNN-GAT model outperforms current ARIMA-LSTM, ARIMA-Artificial Neural Network (ANN) and LSTM-GRU models by achieving less Mean Squared Error (MSE) of 0.098, Normalized Mean Squared Error (NMSE) of 0.312, as well as Mean Absolute Error (MAE) 0.243, respectively.