Deep Time Series Forecasting With Shape and Temporal Criteria.

Area of Science:

• Machine Learning
• Data Science
• Signal Processing

Background:

• Current deep learning forecasting methods struggle with non-stationary signals and sudden changes.
• Mean Squared Error (MSE) based training lacks precision in deterministic and probabilistic forecasting.

Purpose of the Study:

• To develop advanced deep learning models for precise multi-step time series forecasting.
• To enhance predictions for non-stationary signals exhibiting sudden changes.
• To improve both deterministic and probabilistic forecasting capabilities.

Main Methods:

• Incorporating shape and temporal criteria into deep model training objectives.
• Defining differentiable loss functions and positive semi-definite (PSD) kernels using smooth relaxations of Dynamic Time Warping (DTW) and Temporal Distortion Index (TDI).
• Introducing DILATE (DIstortion Loss including shApe and TimE) for deterministic forecasting and STRIPE++ (Shape and Time diverRsIty in Probabilistic for Ecasting) for probabilistic forecasting using determinantal point process (DPP) diversity loss.

Main Results:

• DILATE and STRIPE++ demonstrate significant improvements in forecasting accuracy.
• The proposed methods provide sharper predictions for non-stationary time series.
• Shape and temporal features effectively enhance forecasting performance.

Conclusions:

• Leveraging shape and time features in deep learning models significantly benefits time series forecasting.
• DILATE and STRIPE++ offer robust solutions for complex forecasting challenges.
• The developed techniques advance the state-of-the-art in both deterministic and probabilistic forecasting.