Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting
Efficient transformer architecture using ProbSparse self-attention for long-sequence time series forecasting with O(L log L) complexity.
Verdict
Handles long sequences better than vanilla transformers. Useful for macro forecasting and long-term trend prediction. Implementation complexity is moderate.
Practical Use
Best suited for long-term forecasting problems (weeks to months ahead) with multivariate time series, such as macroeconomic indicators, commodity prices, or electricity demand. The ProbSparse attention mechanism reduces memory usage and speeds up training compared to standard Transformers. However, for short-horizon forecasts or univariate series, simpler models like N-BEATS may be more efficient. Available via Hugging Face Transformers for easy integration.
Paper
https://arxiv.org/abs/2012.07436Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, Wancai Zhang. AAAI, 2021
Code Repository
https://github.com/zhouhaoyi/Informer2020License
Custom. Code (zhouhaoyi/Informer2020): Apache-2.0. Paper: arXiv non-exclusive distribution license (free to read; no reuse license).
Installation
pip install transformers