Deep Adaptive Input Normalization for Time Series Forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Deep learning (DL) models excel at time series analysis but are sensitive to data normalization.
• Financial time series forecasting presents challenges due to non-stationary and multimodal data, degrading DL model performance.
• Existing normalization methods are often fixed and may not be optimal for specific tasks or data distributions.

Purpose of the Study:

• To propose a novel neural layer for adaptive time series normalization in deep learning.
• To enhance the performance of DL models in time series forecasting and analysis, particularly for financial data.
• To develop a normalization method that learns from data distributions and can be applied without retraining.

Main Methods:

• Introduction of a simple, effective neural layer for adaptive time series normalization.
• End-to-end training of the proposed layer using backpropagation.
• Evaluation on large-scale limit order book and load forecasting datasets.

Main Results:

• Significant performance improvements compared to existing normalization schemes.
• Demonstrated effectiveness of adaptive normalization in handling non-stationary and multimodal financial time series.
• The proposed method showed adaptability to new time series without requiring retraining.

Conclusions:

• The proposed adaptive normalization layer offers a simple yet powerful solution for improving DL model performance on time series tasks.
• This method addresses the limitations of fixed normalization techniques by learning task-specific normalization strategies.
• The approach is broadly applicable and effective across different time series datasets, including financial and load forecasting.