USTC and Alibaba Propose CEDAR Framework, Cutting Decision-Conditioned Time Series Forecasting Error by Over 57%

Researchers from USTC and Alibaba have proposed CEDAR, a dual-decoupling modeling framework that uses an action-interleaved Transformer to explicitly model the impact of human decisions on time series and a residual correction module to fuse external intervention factors. Validated on about 32 million product trajectories from Alibaba's 1688 platform, the method reduces forecasting error by over 57% compared to the best baseline, and online A/B tests show a 13% increase in merchant LTV and a 15% increase in store ROI. The work has been accepted as an Oral paper at KDD'26.