OrangeTree: A Linear and Tree-based Time Series Forecasting Model Supporting Multiple Input and Output Lengths

Time series forecasting (TSF) models suffer from a critical rigidity: the inability to handle Multiple Input and Output Lengths (MIOL) within a single architecture. Current workarounds, such as padding or autoregression, invariably incur computational redundancy or error accumulation. We propose OrangeTree, a linear and tree-based architecture that decouples model parameters from sequence lengths. It employs a Segment Tree Encoder and an Inverse Tree Decoder to translate multi-length input history into multi-length output predictions within a unified framework. Crucially, a Range Weighter and Feature Fuser bridge these components, dynamically selecting and fusing relevant ranges into optimal contexts for forecasting. Experiments on six benchmarks confirm OrangeTree achieves SOTA performance, reducing MSE by 3.4% compared to the previous SOTA. In MIOL settings, it matches the accuracy of length-fixed models without the accuracy degradation and latency increase typical of heuristic methods. Code is available at https://anonymous.4open.science/r/OrangeTreeTSF/.