Activation Function Sensitivity in LSTM-Based Peak Stock Price Forecasting for High-Volatility Financial Time Series

Stock price prediction remains an intriguing task due to the high volatility and complex temporal dependencies present in financial time-series data. Accurate prediction of the highest stock price is particularly important for investors seeking to identify market peaks and optimize trading strategies. This study investigates the effectiveness of Long Short-Term Memory (LSTM) networks in forecasting DELL’s highest stock price by analyzing the impact of different activation functions. Historical stock price data from 2016 to 2024 were used, and several preprocessing techniques, including data normalization and chronological train-test splitting, were applied. The LSTM models were trained for 100 epochs and evaluated using Root Mean Square Error (RMSE), Mean Squared Error (MSE), and Mean Absolute Error (MAE). The main contribution of this research is a comparative analysis of the sensitivity of LSTM prediction performance to different activation functions, namely ReLU, ELU, Sigmoid, and Tanh, in the context of high-volatility financial time-series data. The experimental results show that the LSTM model using the ReLU activation function achieved the best performance, with an RMSE of 0.557942, MSE of 0.311300, and MAE of 0.338773, outperforming the other activation functions. These findings demonstrate that activation function selection significantly influences LSTM forecasting performance. The results provide practical insights for financial analysts and investors in selecting appropriate deep learning configurations for more reliable stock price prediction.