ANALYSIS OF THE EFFECT OF HYPERPARAMETER VARIATION AND ACTIVATION FUNCTION ON MODELING CONVERSION OF SHARE PRICE PREDICTION WITH LSTM METHOD

• Main Author: Kemal Fajri, Adam
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/54933
• Institution: Institut Teknologi Bandung

Physics which aims to explain a phenomenon with modeling and theories, in its development trying to model complex systems. Physics in the most rapidly developing complex systems is econophysics which attempts to model economic systems, particularly capital markets. In the capital market, stock prices that fluctuate at any time cause risks when investing. This risk can be anticipated by performing technical analysis, which is an analysis that involves searching for patterns from historical data. One method that can be used to model and predict stock prices is the LSTM method. This paper aims to analyze the effect of LSTM hyperparameter variation and activation function variation on prediction results. The research was conducted by modeling stock prices by applying different activation functions to each closing stock price data of four companies from four different sectors. From this research, the effect of variation of epochs on modeling is that when the selected epochs value is too small, the learning process from the model cannot achieve optimal results, while the selection of a larger epochs value makes the learning process of the model redundant so that the model becomes overwhelmed and produces modeling results. which is not optimum. Then, the effect of variations in the value of batch size on modeling does not form any patterns, where each model of stock price data has a different batch size value suitability. Next, the varied learning rate values, the greater the learning rate used in pelatihan, the faster the model will converge to reach the minimum loss function target and vice versa, when the learning rate used is small, the model will tend to be slow to converge to reach the target loss function. minimum. The activation function that provides the most optimum modeling results is the Tanh activation function.