Long-term time series forecasting (LTSF) uses the current sequence as input to predict future values, playing a critical role in applications such as weather forecasting and electricity consumption planning. State-of-the-art models often produce inconsistent outputs for the same timestamps when fed with lagged input sequences, resulting in low output alignment and fluctuating predictions that undermine reliability.
To tackle this issue, we introduce AliO (Align Outputs), which enforces consistency in both the time and frequency domains. The key idea is to minimize the discrepancy between predictions generated from different lagged inputs and to add an alignment term to the training loss.
We also propose a new evaluation metric, TAM (Time Alignment Metric), specifically designed to quantify the alignment of prediction outputs. Unlike the conventional mean squared error (MSE) that only measures distance to ground truth, TAM directly assesses the similarity of predictions at identical timestamps across multiple sequences.
Extensive experiments on several public LTSF benchmarks show that AliO can improve TAM by up to 58.2% while simultaneously boosting forecasting accuracy by as much as 27.5%. Enhanced alignment increases the model's trustworthiness in real-world deployments.
In summary, AliO’s alignment constraints and the TAM metric provide an effective pathway to improve the reliability of long-term time series forecasting models.
Review