An Experimental Review on Deep Learning Architectures for Time Series Forecasting.

Area of Science:

• Data Mining
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning models demonstrate superior performance over traditional methods in various machine learning tasks.
• Deep neural networks are effective for time series forecasting due to their ability to capture temporal dependencies.
• Choosing the optimal deep learning architecture and parameters for time series forecasting remains a complex challenge.

Purpose of the Study:

• To conduct a comprehensive review of recent deep learning applications in time series forecasting.
• To experimentally compare the performance of popular deep learning architectures for forecasting tasks.
• To analyze the suitability of different deep learning models and configurations for diverse forecasting problems.

Main Methods:

• A thorough review of the latest deep learning techniques applied to time series forecasting.
• An experimental comparison of seven popular deep learning architectures, evaluating accuracy and efficiency.
• Analysis of model performance across over 50,000 time series datasets and 12 forecasting problems, training more than 38,000 models.

Main Results:

• Long Short-Term Memory (LSTM) networks yield the most accurate forecasts among the evaluated models.
• Convolutional Neural Networks (CNNs) demonstrate comparable accuracy to LSTMs, with less result variability across configurations.
• CNNs offer greater computational efficiency compared to other deep learning models evaluated.

Conclusions:

• LSTM networks are highly recommended for achieving the highest accuracy in time series forecasting.
• CNNs present a robust alternative for time series forecasting, balancing accuracy with efficiency and stability.
• This extensive study provides valuable insights for selecting appropriate deep learning models for various time series forecasting challenges.