Stock markets continue to be a challenging area for investors to predict due to their nonlinear and complex structure. The susceptibility of financial time series to sudden price movements and complex patterns also affect the forecasting performance of statistical methods. Therefore, the use of advanced models has become crucial to minimize financial risks, enhance investment returns, and achieve consistent predictions.

The aim of this study is to compare the performance of deep learning methods in analyzing sequential data that contains nonlinear and long-term dependencies. For this purpose, Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN) architectures were designed to forecast the daily closing prices of the BIST 100 index for the 2015–2025 period. Hyperparameter optimization was conducted using the Hyperband algorithm, and the predictive performance of the models was evaluated based on RMSE, MAE, and MAPE error metrics. The findings indicate that both models successfully capture the long-term dependencies of the BIST 100 Index. However, upon examining the error metrics, it was found that the TCN architecture produced lower error values compared to LSTM. In particular, it is found that the TCN architecture adapted more effectively to sudden market fluctuations and trend reversals, producing more accurate forecasts. These results suggest that TCN offers a powerful and innovative alternative for modelling the nonlinear dynamics in financial time series forecasting.