A Comparative Study on Deep Learning Models for Time-Series Forecasting

This study investigates the performance of four deep learning architectures including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional Neural Network (CNN) and Transformer for univariate time-series forecasting. To evaluate their ability in capturing different temporal dynamics, we selected two contrasting datasets: Apple Inc. (AAPL) stock prices, characterized by noise and volatility without clear seasonality and Melbourne’s daily minimum temperatures, which exhibit strong seasonal patterns. Each model was trained using consistent configurations and evaluated using standard metrics. Our results show that GRU and LSTM perform best across both domains, particularly in handling abrupt changes in financial data, while CNN and Transformer show competitive performance on smoother seasonal data. The findings highlight the importance of aligning model architecture with the underlying structure of the time series.