AdaRDiff Boosts Time Series Forecasting Accuracy and Speed.

Researchers have developed Adaptive Reversible Differencing (AdaRDiff), a novel method designed to significantly improve long-horizon time series forecasting. Traditional differencing, which removes trends and seasonality, often relies on manual parameter selection and has been difficult to integrate into modern deep learning architectures. AdaRDiff overcomes this by using learnable weights to perform weighted differencing with past values, stabilizing the time series into residuals that are easier for models to forecast. After forecasting the residuals, AdaRDiff autoregressively reconstructs the original components, effectively capturing both trend and seasonality through a single, efficient operator. This reconstruction process is parallelizable on GPUs, yielding substantial speedups (up to 33.7x). The method also employs a two-phase training schedule to optimize temporal structure discovery and reconstruction learning. AdaRDiff has achieved state-of-the-art accuracy across eight diverse benchmarks, including electricity, weather, and traffic data. Crucially, it functions as a plug-and-play module, enhancing the performance of various backbones, from linear models to Transformers, by up to 25.9%.