AlignTime: Interperiodic phase alignment sampling for time-series forecasting

Abstract

Time-series forecasting is widely applied in various fields to provide decision-making support. However, existing forecasting methods often struggle to capture periodic behavior in nonstationary sequences. When periodic fluctuations exhibit nonstationarity over time, current forecasting architectures can face difficulties in effectively handling dynamic phase shifts across different periods. To address these challenges, we proposed AlignTime — a novel time series forecasting model that revisits forecasting from a downsampling perspective. We treated each period in a time series as an independent analytical unit and considered elements in the same phase to share similar signal properties. By sampling these phase-aligned elements, we constructed a phase subsequence rich in features from different periods, enabling a deeper understanding and utilization of intrinsic periodic patterns. We introduced a method for computing the globally dominant period and performed phase-aligned sampling based on it, effectively aggregating time points in the same phase. We then employed phase-enhanced convolution to extract features from the phase-aligned subsequences, captured dynamic phase shifts between periods, and represented the data in the feature space. Finally, we aggregated the feature subsequences from each phase window to generate the final predictive sequence, fully leveraging the inter-period dynamic phase correlations. We conducted extensive experiments on both long- and short-term forecasting tasks. For long-term forecasting, AlignTime achieves a 6.85% improvement in MSE and a 4.09% improvement in MAE compared to baseline models. For short-term forecasting, over 70% of the results achieve optimal performance, confirming the effectiveness of AlignTime in time-series forecasting tasks. Code is available at: https://github.com/xiaoxiaomiwang/AlignTime.