TimeSqueeze: Dynamic Patching for Efficient Time Series Forecasting

Recent progress in time series forecasting has produced large foundation models with strong generalization across domains. However, many of these models rely on transformer backbones, making their effectiveness constrained by the cost of processing the input context. The quadratic computational complexity with respect to sequence length imposes a fundamental trade-off on existing designs: they must either preserve high-frequency information using point-wise embeddings, which is computationally expensive for long sequences, or employ patch-based embeddings to reduce sequence length at the risk of discarding critical temporal details. To overcome this limitation, we present TimeSqueeze, a hybrid forecasting architecture that combines the strengths of both point and patch embeddings through dynamic time series compression. TimeSqueeze introduces a novel two-stage hybrid representation: (1) a lightweight state-space encoder processes the full-resolution time series with point-wise embeddings to extract fine-grained temporal features, and (2) an adaptive patching module intelligently prunes these features using variable-sized patches, assigning smaller patches to information-rich regions and larger patches to redundant segments. This hybrid approach yields a variable-resolution representation that preserves critical temporal details while reducing computational overhead. By retaining the fidelity of point embeddings and the efficiency of patch embeddings, the resulting compressed sequence enables the Transformer backbone to substantially reduce the input length without sacrificing forecasting accuracy. Extensive experiments demonstrate that TimeSqueeze achieves state-of-the-art forecasting performance while delivering substantial computational advantages, including up to $8 \times$ improvement in pretraining data efficiency and up to $20 \times$ reduction in pretraining time compared to equivalent point-embedding models.