Investigation of the effect of different node numbers, time frames and activation functions on stock change prediction with artificial neural network models

The random walk of systematic variables, which are influenced by numerous factors, is very difficult to predict even with modern techniques. In solving such complex problems, the effectiveness of artificial intelligence-based models and which models yield better results with specific parameters is a subject of interest for researchers. This thesis investigates the prediction of future stock prices, a problem of this nature, using different artificial neural network models, testing various parameters and examining their impact on prediction performance. In the study, data containing variables with different frequencies from selected stocks of the BIST and NASDAQ stock exchanges were used, and LSTM and MLP models with varying numbers of nodes and activation functions were tested across different time frames. The experiments were carried out in two main examples. The first example examined the effect of different numbers of nodes and activation functions. In this context, predictions were made using MLP models with ReLU or sigmoid activation functions and varying numbers of nodes. In the second example, the effects of using data with variables at different frequencies, various models, and different time frames on prediction results were tested. In this section, predictions were made using LSTM and MLP models over 1, 2, 3, 4, and 5-day time frames. The generated predictions were compared with actual values using MAPE and EDOC metrics. The results suggest that the number of nodes can be determined by an approximate calculation rule, and that ReLU might be a better choice than Sigmoid for the activation function. Moreover, it was observed that using a greater number of variables with different frequencies improved prediction accuracy. Predictions made with short time frames were found to fail, consistent with the random walk hypothesis. As the length of the time frames increased, direction accuracy improved significantly with only a small increase in MAPE, and the LSTM model performed better than the MLP. These findings are consistent with Dow Theory, indicating that the past values of stocks have a relationship with their future movements.