Stock Price Prediction in BIST30 Using Graph Convolutional Network, Long Short-Term Memory and Hybrid Models: A Comparative Analysis

This study investigates the effectiveness of different neural network architectures for predicting stock prices using data from the BIST30 index. Recognizing the inadequacy of traditional models in handling the volatile and interconnected nature of financial markets, our study introduces a hybrid deep learning model. We compare the performance of a Graph Convolutional Network (GCN), Long Short-Term Memory (LSTM) and a hybrid model that combines both approaches. This hybrid approach uniquely captures both the crucial relational dependencies among BIST30 stocks (via GCNs) and the sequential temporal patterns in their price movements (via LSTMs). Our methodology constructs a stock graph for GCN feature extraction, while LSTMs process the individual stock time-series. The outputs from these two distinct processing streams are then integrated to produce final predictions, incorporating additional external factors like dollar and gold parity. Our findings indicate that the hybrid model, leveraging the strengths of both GCN and LSTM, exhibits superior performance in capturing both temporal dependencies and relational structures within the stock market data.