xLSTMPerceiver: Recurrent Temporal Modeling with Latent Cross-Channel Fusion for Multivariate Time Series Forecasting

As time series are an abundant data modality across numerous domains including healthcare, finance, or climate science, accurately and efficiently forecasting them is key for guiding decision-making. In this work, we introduce xLSTMPerceiver, a novel encoder-decoder architecture for multivariate time series forecasting. It first partitions an input series into non-overlapping patches, embeds them as tokens, and semantically augments them by adding positional and channel embeddings. A latent bottleneck mechanism then compresses these tokens into a fixed-sized global summary that captures both temporal and cross-channel relationships. For forecasting, we introduce stacked refinement blocks, each interleaving an xLSTM layer with cross-attention to the global context summary. This design couples the sequential inductive bias of xLSTM with global multivariate context, enabling progressive refinement throughout the blocks. Our comprehensive experiments across various datasets and prediction lengths demonstrate that xLSTMPerceiver achieves competitive forecasting performance compared to recent state-of-the-art baselines.