Retrieval-Augmented Framework with Adaptive Gating for Time Series Forecasting

Although Transformer-based models have achieved good performance in time series forecasting, existing methods still have limitations in modeling long-term dependencies. To this end, we propose a retrieval-augmented framework for time series forecasting, which models the long-range dependency explicitly. First, we consider both time-domain and frequency-domain similarities to retrieve top-K similar historical sequences for each input sequence. Subsequently, to fully exploit the retrieved information while suppressing potential noise, a dual-branch architecture is designed: the original sequence branch primarily models the internal dependencies of the current input, while the retrieval-augmented branch models the relationship between the current input and the retrieved historical sequences through a cross-attention mechanism. Furthermore, to adaptively balance the contributions of the two branches, a gating fusion mechanism is introduced to dynamically adjust the proportion of information integration from the two branches. Experiments on several public datasets show that our proposed method achieves superior and stable performance compared to existing methods, especially in long-term forecasting settings.