Rethinking Channel Dependence for Multivariate Time Series Forecasting: Learning from Leading Indicators

Exploiting locally stationary lead-lag relationships between variates improves multivariate time series forecasting accuracy.