LagNet uses lag-structured operators for long-term forecasting
Original title: LagNet: Lag-Structured Operators for Long-Term Time Series Forecasting
LagLayer ties history-to-future weights by relative lag and can replace temporal linear layers in existing models.
The paper studies the history-to-future operator in long-term time series forecasting and reports recurring diagonal patterns in sensitivity matrices that match relative temporal lags. LagLayer builds a Toeplitz matrix by sharing connections with the same lag, then relaxes exact sharing with horizon-wise and history-wise scaling. Stacked LagLayer modules, activations and feature maps form LagNet. On forecasting benchmarks it is competitive with strong baselines, and limited-data tests suggest it is more sample-efficient than a linear layer.
- LagLayer is designed as a drop-in replacement for temporal linear layers.
- Sensitivity diagnostics and limited-data experiments support the lag-sharing design.
- The paper does not report a product release or deployment.
- AI models
- LagNet
From the original article
1 min read at arxiv.orgLong-term time series forecasting maps a finite history window to an extended future horizon. Many existing models focus primarily on the design of input representations and model structures, while paying less attention to the structural form of the history-to-future forecasting operator itself.
By Zhenan Yu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.