Causal-Guided Ultra-Long-Term Time Series Forecasting Via Anticipated Covariates

Area of Science:

• Dynamical systems
• Time series analysis
• Causal inference

Background:

• Traditional time series forecasting struggles with error accumulation over ultra-long horizons.
• Existing methods often underutilize future information, treating it as unknown.
• Seasonal covariates improve medium- to long-term accuracy but not ultra-long-term performance.

Purpose of the Study:

• To demonstrate that future effect information can enable accurate ultra-long-term forecasting of causes.
• To introduce a novel forecasting paradigm using anticipated covariates.
• To validate the effectiveness of this paradigm across benchmarks.

Main Methods:

• Investigating coupled dynamical systems where future states of effect X are provided to forecast cause Y.
• Analyzing systems with unobserved driving variables (Z → Y → X) to assess error boundedness.
• Implementing and validating the anticipated covariates paradigm on established benchmarks.

Main Results:

• Providing future effect states stabilizes forecasting and bounds errors for thousands of timesteps.
• The forecasting error remains bounded even with unobserved causal factors.
• Ultra-long-term predictions become feasible with substantially reduced errors under suitable conditions.

Conclusions:

• Anticipated covariates offer a powerful tool for ultra-long-term time series forecasting.
• This paradigm is beneficial in scenarios with high data costs, limited historical data, or unobservable causal factors.
• The findings prompt a re-evaluation of reverse time-dependent causality in forecasting.