CAMP Enhances Time Series Forecasting with Adaptive Cycle Learning

Real-world time series data often exhibits recurring patterns, but the dominant periods can vary significantly across datasets, forecasting contexts, and even within individual input windows. Current cycle-aware forecasting models typically rely on a single, pre-defined period, which can be overly restrictive. Additionally, patch-based models often process all patches uniformly, despite the fact that patches further from the forecast boundary might need broader contextual refinement, while recent patches contain critical information that should be preserved more directly. This research introduces CAMP (Cycle-Aware Multi-Scale Patch Mixer), a novel framework designed to address these challenges. CAMP features an Adaptive Cycle Learning module that identifies dominant frequencies for each input window independently, generating both historical and future cyclic components without requiring a fixed cycle length. Furthermore, its Horizon-Guided Patch Mixer refines patches based on their position, allowing earlier patches to incorporate wider temporal context while preserving information from recent patches. CAMP also models the de-cycled residual dynamics through temporally aligned multi-resolution representations, capturing complementary dynamics across different scales within a unified framework. Extensive experiments on seven long-term forecasting benchmarks show CAMP achieving superior average Mean Squared Error (MSE) on six datasets and top Mean Absolute Error (MAE) on six, consistently outperforming existing methods across various traffic benchmarks.