Block-Based Autoregressive Approach for Long Horizon Time Series Forecasting
Long Horizon Time Series Forecasting (LHTSF) remains a challenging research problem due to the difficulty of producing reliable predictions over long forecasting horizons. Existing approaches mainly rely on recursive forecasting or multi-input multi-output (MIMO) strategies. Recursive methods are prone to error accumulation, while MIMO approaches may fail to adequately preserve temporal dependencies across forecast steps. In this study, we propose BlockMIMO, a block-based forecasting framework that provides a balance between these approaches. BlockMIMO divides the forecasting horizon into discrete blocks and predicts each segment sequentially while updating the input window with the generated blocks. This structure aims to preserve local temporal dependencies while mitigating error propagation in long-term forecasting. The proposed framework can be applied to different neural network architectures, and demonstrates superior performance on widely used multivariate time series datasets.