Assessment of an Optimal Multilayer Perceptron Model in Estimation of Stock’s Price Rates; a Sample from the Turkish Stock Exchange BIST

Stock price prediction (SPP) is critical for both investors and financial experts to make accurate choices about real estate share transactions. Nowadays, Artificial Neural Networks (ANNs) are commonly used in SPP algorithms. In this study, the performance of a Multi-Layer Perceptron (MLP), a subclass of ANN, in forecasting the closing price of the following day as well as the Direction of Change (DOC) of stock was investigated. The experiments were conducted on four randomly selected stocks from Borsa Istanbul (BIST), using Sigmoid and Rectified Linear Unit (ReLU) activation functions and optimizing with variation of the node numbers in each of the hidden layers of the ANN. For each stock, the data comprised 500 workdays recorded over April 7, 2022, to April 4, 2024 period were used for training and testing the NN model taken under analyses. The predicted values come through ANN were compared with actual data using the Mean Absolute Percentage Error (MAPE) metric. The optimal outcomes were obtained with MLP with ReLU activation function and 8-5-3-1 nodes in the subsequent layers. It was observed that the predictability of stocks changed with macroeconomic and microeconomic factors as well as different time periods, which shows that stock prices also depend on temporal parameters. The overall results showed that stocks’ price prediction may not be impossible, but not an easy business.