When a satellite forecast's risk budget stops working

Satellite bandwidth allocation needs more than a good average prediction. If a system admits traffic against an optimistic capacity estimate, it can promise service the link cannot deliver. Risk-controlled forecasting instead produces a conservative bound and aims to limit how often that bound exceeds actual capacity.

An arXiv preprint submitted on November 4, 2026, by Mayan Sharma and colleagues examines whether that control survives realistic evaluation. The study reproduces BG-CFQS, a published risk-aware throughput forecasting method, and tests existing calibration alternatives. Its contribution is empirical, not a new algorithm; the supplied landing page lists no journal publication.

Random splits are not future traffic

The experiments use three StarNet Starlink traces from Chicago, Osnabruck and Victoria, plus an independent WetLinks measurement campaign. The authors also reproduce StarNet's neural forecasting backbone: a two-layer gated recurrent network with attention and an embedding of the 15-second scheduling cycle. Their main chronological evaluation uses 60/20/20 training, calibration and test blocks, with purged boundaries to prevent overlapping time steps.

Calibration uses held-out observations to choose how conservative predictions should be. Randomly mixing observations across calibration and evaluation makes those sets more alike; chronological splitting exposes changes between earlier and later traffic.

At a requested overestimation budget of 35%, Table 1 reports that BG-CFQS passes on all three StarNet traces under a random split, but only one under a temporal split. Average overestimation rises from 34.0% to 37.7%. The prose separately reports a temporal result of 39.1% and zero passes without clearly reconciling it with that table, so the precise temporal reproduction result is internally inconsistent.

Even passing globally does not protect every operating condition. Under the random split, the bound exceeds capacity on average in 84.8% of decisions within the lowest-throughput tenth. Overall risk control and protection during weak-link periods are different requirements.

A risk dial with a hard stop

BG-CFQS searches a candidate quantile range from 0.15 to 0.40. When even its most conservative candidate fails the requested calibration budget, the search collapses to that candidate rather than finding a safer one.

Across all four datasets, requested budgets of 5%, 10% and 15% return identical achieved rates within each dataset. Those rates range from 7.0% in Victoria to 18.5% on WetLinks. At the strictest 5% budget, that is 1.4 to 3.7 times the requested rate.

This is a candidate-set limitation, not a universal 15% lower bound on achieved risk: the measured floor varies by dataset. The important failure is that tightening the requested budget stops changing the output. The original method's evaluation at 35% did not expose this behavior.

Updating helps; dividing the data may not

The authors separate two familiar remedies. Online recalibration updates the correction as outcomes arrive. Group conditioning keeps different corrections for regimes defined by features such as satellite elevation, visible-satellite count or recent throughput.

Averaged across the three StarNet locations, online recalibration tracks throughput budgets from 5% to 35% within 0.8% relative error. Static calibration undershoots by 16–21%, withholding capacity relative to the chosen operating target. On WetLinks, static calibration instead overshoots: changing conditions can push the error in either direction.

At matched achieved risk, online per-regime calibration also has lower low-throughput overestimation rates and lower mean absolute error than its static counterpart. Matching achieved risk matters because extra conservatism alone can make a method look safer.

Adding groups is less dependable. Across the three StarNet locations, the mean overestimation rate in the lowest-throughput tenth rises from 66.6% without conditioning to 70.9% with all four regime axes. The authors associate the degradation with splitting feedback among more groups. Features that predict capacity levels need not improve calibration of forecast errors. A calibration-set test for differences between groups also fails to predict when grouping will help.

A latency-detection experiment adds another distinction: a single fixed additive correction cannot improve prediction ranking because it never changes the order. Online corrections can change rankings, improving precision-recall area by 51% and 43% on two links, but worsening it on Chicago.

Evidence, not a deployment guarantee

All results come from recorded-trace replay, not a live terminal. Only four datasets are studied, three from one campaign. WetLinks also restricts the combined input history and forecast horizon to 15 samples, preventing the original long-window configurations from running there.

Crucially, the observed online budget tracking is empirical, not a finite-sample guarantee, and requires outcome feedback after every decision. The practical lesson is to test risk control across time, budgets and difficult operating conditions—not to treat a calibration label as proof of deployment safety.