CAMP forecasting model tops 6 benchmarks with adaptive cycle learning

Time series forecasting remains hard when dominant cycles shift or when multiple periods coexist. Existing cycle-aware models typically pick a single dataset-level period, missing changes within a window. Meanwhile, patch-based transformers treat all patches uniformly, even though patches far from the forecast boundary need broader context, while recent patches should be preserved nearly untouched. CAMP, from researchers Jung Min Choi, Vijaya Krishna Yalavarthi, and Lars Schmidt-Thieme, attacks both problems directly. Its Adaptive Cycle Learning extracts per-window frequencies, generating historical and future cyclic components without requiring a user-defined cycle length. A Horizon-Guided Patch Mixer then applies position-dependent refinement: earlier patches integrate wider temporal context, while near-boundary patches retain fine-grained information. The residual non-cyclical dynamics are captured via temporally aligned multi-resolution representations, letting the model express patterns that operate at different speeds within one framework.

CAMP's results are strong. In head-to-head evaluation across seven long-term forecasting benchmarks, it achieves the best average MSE on six datasets and the best or tied-best MAE on six. On four PEMS traffic benchmarks, it records the highest MSE win count across all sixteen settings. These gains matter for practical applications like energy load prediction, supply chain planning, and financial modeling, where accurate multi-horizon forecasts are critical. By removing the dependence on pre-specified cycles and adapting to each window's unique structure, CAMP offers a flexible, plug-and-play architecture that could replace rigid forecasters in production systems. The paper is available on arXiv with code and data links attached.