Authors: Vincent Zhihao Zheng, Seongjin Choi, Lijun Sun

Published on: May 26, 2023

Impact Score: 8.45

Arxiv code: Arxiv:2305.17028

Summary

• What is new: Innovation in training method that incorporates error autocorrelation for better forecasting accuracy.
• Why this is important: Existing models often overlook serial correlation, assuming a time-independent error process.
• What the research proposes: A training method that constructs a mini-batch of consecutive time series segments and learns a time-varying covariance matrix to encode error correlation.
• Results: Improved performance of neural forecasting models across multiple datasets, resulting in notable improvements in predictive accuracy.

Technical Details

Technological frameworks used: nan

Models used: Neural forecasting models

Data used: Multiple public datasets

Potential Impact

Financial markets, weather forecasting companies, supply chain management firms

Want to implement this idea in a business?

We have generated a startup concept here: FutureSight.