BG-CFQS: Risk-aware forecasting cuts Starlink overestimation errors by 12.6%

Starlink’s low Earth orbit network delivers highly variable throughput, making short-term forecasting critical for resource management. Traditional models optimize symmetric metrics like MAE and RMSE but ignore the asymmetric risk of overestimating future capacity—which can trigger over-admission, bandwidth overbooking, and service violations. A team of researchers from Chinese universities and Nanyang Technological University addresses this gap with BG-CFQS (Budget-Guided Coarse-to-Fine Quantile Selection), a framework that trains a family of lower-quantile predictors, identifies the quantile boundary that meets a prescribed overestimation budget, and refines that boundary to select the most accurate feasible predictor.

Tested on three real-world Starlink throughput datasets, BG-CFQS consistently satisfies the risk budget while achieving the lowest average MAE, mean positive error, and tail positive error among budget-feasible methods. In high-risk and severe-risk low-throughput regimes, it reduces harmful positive errors by 11.0% and 12.6%, respectively. An admission-control simulation further shows that the safe forecasts translate directly into fewer dropped sessions, proving that risk-aware forecasting can turn statistical safety into tangible network reliability improvements for the world’s largest LEO broadband system.