TimePerceiver: A Unified Framework for Time-Series Forecasting
Published:Dec 27, 2025 10:34
•1 min read
•ArXiv
Analysis
This paper introduces TimePerceiver, a novel encoder-decoder framework for time-series forecasting. It addresses the limitations of prior work by focusing on a unified approach that considers encoding, decoding, and training holistically. The generalization to diverse temporal prediction objectives (extrapolation, interpolation, imputation) and the flexible architecture designed to handle arbitrary input and target segments are key contributions. The use of latent bottleneck representations and learnable queries for decoding are innovative architectural choices. The paper's significance lies in its potential to improve forecasting accuracy across various time-series datasets and its alignment with effective training strategies.
Key Takeaways
- •Proposes TimePerceiver, a novel encoder-decoder framework for time-series forecasting.
- •Generalizes the forecasting task to include extrapolation, interpolation, and imputation.
- •Introduces latent bottleneck representations for encoding and learnable queries for decoding.
- •Demonstrates superior performance compared to state-of-the-art baselines across various datasets.
Reference
“TimePerceiver is a unified encoder-decoder forecasting framework that is tightly aligned with an effective training strategy.”