AliO Improves Output Alignment in Long-Term Time Series Forecasting

Summary

Long-term time series forecasting models can produce inconsistent predictions for the same future timestamps when they receive lagged input sequences, reducing confidence in their forecasts. The authors propose AliO, or Align Outputs, which reduces discrepancies between such predictions in both the time and frequency domains. They also introduce the Time Alignment Metric (TAM) to measure output alignment, a property that conventional mean squared error does not capture because MSE focuses on distance from ground-truth values. Experiments show that AliO improves TAM by up to 58.2% while maintaining or improving forecasting performance by up to 27.5%. The authors present the resulting improvement in consistency as a way to make long-term forecasting models more reliable for applications such as weather forecasting and electricity-consumption planning.