Goal-Oriented Time-Series Forecasting: Foundation Framework Design

Conventional forecasting minimizes overall error but overlooks the differing importance of forecast ranges in downstream tasks. We propose a method that partitions predictions into segments and dynamically reweights them at inference, enabling flexible, on-demand targeted forecasting without retraining. Experiments on benchmarks and a new wireless dataset show improved accuracy in regions of interest and measurable downstream gains, highlighting the benefits of tighter integration between prediction and decision-making.