Predicting Stock Prices in The Turkish Banking Sector with Artificial Neural Networks: A Comparison of Multi-Layered LSTM Models

In this study, the stock prices of the leading banks in the Turkish banking sector (Akbank, Garanti Bankası, İş Bankası, and Yapı Kredi Bankası) were predicted using different numbers of LSTM layers (from 1 to 5) to examine the effect of layer depth on model performance, ultimately determining the optimal LSTM architecture. Following time series decomposition of the banks’ stock prices, LSTM-based models predicted these prices using 1- to 5-layer architectures, and an experimental analysis was conducted aiming to reveal the optimal layer depth by comparing performance with RMSE, MAE, MAPE, and R² metrics. In LSTM models with different layer depths, moderately deep architectures provided the best prediction performance, while overly deep structures exhibited performance declines due to increased model complexity. Evaluating the effect of the number of LSTM layers on the stock price movements of the leading banks in the BIST30, this study demonstrates that a deep learning configuration appropriate to the complexity of financial data reveals the risk and return dynamics specific to the banking sector. By emphasizing the optimality of different layer depths in time series forecasting through an innovative method, it makes significant contributions to the literature.