Adaptive Refinement of Time Series Foundation Models via Pattern and Context-Awareness

Time-series forecasting in domains such as ITOps and IoT faces two major challenges: data are non-stationary and multivariate, and state-of-the-art Time-Series Foundation Models (TSFMs) rely on fixed-size windows that miss transient phenomena (e.g., spikes, drifts) and their historical context. Prior efforts address this with seasonal-trend decompositions or frequency-aware pre-training, but these require retraining and offer limited adaptability. We propose a dynamic two-stream framework that augments any pre-trained TSFM with frequency pattern awareness and contextual retrieval as the two streams. Each input window is decomposed via Fast Fourier Transforms (FFT) and Discrete Wavelet Transforms (DWT) to extract key low- and high-frequency patterns, which are fused into a TSFM through lightweight adapters and gated embedding augmentation. In parallel, frequency pattern signatures are used to retrieve semantically similar historical sequences, enriching long-range context. This approach enhances the forecasting robustness of deployed TSFMs without retraining, achieving consistent improvements over baselines on zero-shot forecasting benchmarks, particularly with abrupt data fluctuations and complex temporal dynamics.