Horizon

Understanding and Predicting Global Starlink Performance

Proc. ACM Meas. Anal. Comput. Syst. (SIGMETRICS), Vol. 10, No. 2, June 2026

What is Horizon?

Horizon is the first global-scale machine learning system for predicting LEO satellite Internet performance. By integrating crowdsourced measurements from M-Lab NDT7 and Cloudflare AIM with meteorological data and satellite orbital propagation, Horizon achieves accurate performance predictions spanning 90+ countries.

90+
Countries
7,800+
Starlink Satellites
17.76 ms
MAE Latency
11
Months of Data

Key Contributions

1

Global-Scale Prediction

First ML system predicting Starlink performance globally using crowdsourced measurements from M-Lab NDT7 and Cloudflare AIM.

2

Novel Feature Engineering

Integration of weather indices, satellite density via orbital propagation, and spatio-temporal features for comprehensive modeling.

3

Robust Generalization

Leave-one-location-out experiments validate predictions for entirely unseen regions with comparable accuracy.

4

Open Dataset & Pipeline

Publicly available dataset spanning 11 months, 90+ countries, with full reproducible pipeline.

How It Works

A five-phase pipeline from raw measurements to global predictions

1

Data Collection

Crowdsourced measurements, weather APIs, satellite orbital databases

M-Lab NDT7 Cloudflare AIM OpenMeteo TLE/SGP4
2

Pre-processing

Data cleaning, server filtering, cross-dataset validation

Unified format Server bias removal
3

Feature Enrichment

Weather Index via PLS, satellite density from orbital propagation

Weather Index Sat Density Spatio-temporal
4

Anomaly Detection

Location-aware filtering preserving natural variability

Percentile Directional MAD Isolation Forest
5

Model Training

Ensemble combining Random Forest + Gradient Boosting

RF (100 trees) GBR (100 trees) Weight optimization

Key Results

Latitude contributes 42-46% of feature importance, confirming Starlink's orbital geometry as the dominant performance factor.

The novel PLS-derived Weather Index captures 14-15% of predictive power through cloud cover, precipitation, wind speed, and temperature.

Leave-one-location-out cross-validation shows the model generalizes to entirely unseen cities with R² = 0.44 for latency.

Team

CB

Cristian Benghe

TU Delft

C.Benghe@student.tudelft.nl
VG

Vlad Graure

TU Delft (now at Google)

TS

Tanya Shreedhar

TU Delft

NM

Nitinder Mohan

TU Delft