Quadratic Direct Forecast for Training Multi-Step Time-Series Forecast Models
This work introduces a novel quadratic-form weighted training objective to improve multi-step time-series forecasting models. It simultaneously addresses two issues: label autocorrelation and heterogeneous task weights. The proposed Quadratic Direct Forecast (QDF) algorithm adaptively learns a weighting matrix to balance these effects, achieving state-of-the-art performance across various forecast models.