Horizon: Understanding and Predicting Global Starlink Performance

Starlink has deployed over 7,800 satellites serving millions of subscribers, yet predicting its performance remains an open challenge. Rapid orbital dynamics, frequent handovers, and weather-induced attenuation create variability that existing models, built on a handful of instrumented terminals in limited regions, cannot capture at global scale. We present Horizon, the first global-scale learning driven telemetry and prediction system for LEO satellite Internet performance. By integrating 11 months of crowdsourced measurements from M-Lab and Cloudflare spanning 90+ countries with meteorological data and orbital propagation features, Horizon achieves mean absolute errors of 17.76 ms for latency and 25.63 Mbps for throughput, outperforming state-of-the-art baselines. Our analysis reveals that geographic position dominates prediction accuracy, with latitude alone contributing 42-46%, while weather features account for 14-15%, quantifying the impact of atmospheric conditions on Ku/Ka-band links. Leave-one-location-out experiments demonstrate Horizon's robust spatial generalization to unseen regions. Horizon operates as a deployed system providing hourly forecasts across 300+ regions worldwide, bridging the gap between localized measurement studies and the need for global performance visibility.