Predictions vs Observations

This comparison tests two independent HF propagation prediction systems against real observations from Bouvet Island — a location that neither system was specifically designed or tuned for.

VOACAP is the industry-standard HF prediction engine, developed by NTIA/ITS in the 1980s and trusted by broadcasters, military planners, and radio amateurs for four decades. It is a deterministic ray-tracing framework using semi-empirical ionospheric coefficients (CCIR/URSI) derived from decades of ionosonde measurements to predict monthly-median circuit performance.

IONIS V22-gamma is a neural network (207K parameters) trained on 175 million propagation signatures from WSPR beacons, RBN skimmers, and contest QSOs. It learns propagation patterns directly from observed data rather than from ionospheric models.

Bouvet Island at 54°S presents a genuine blind test for both systems. VOACAP's ionospheric coefficients have sparse coverage at extreme southern latitudes, and IONIS had virtually zero training data from grid JD04.

The results show that IONIS tracks observed SNR within a few dB across most bands (overall bias: −1.9 dB), while VOACAP's predictions diverge significantly at this latitude (overall bias: −11.3 dB), particularly on the low bands where it underestimates propagation and on the high bands where it overestimates.

This does not diminish VOACAP's value — it remains an essential tool and the foundation that the entire HF prediction field is built on. What it suggests is that data-driven approaches may extend prediction accuracy into regions where traditional semi-empirical models have less observational support, such as extreme polar paths. Both approaches are complementary: VOACAP's deterministic ray-tracing framework provides interpretability and reliability across well-characterized paths, while ML models can generalize from large observational datasets to capture highly dynamic, short-term propagation features.

Survivorship bias: These comparisons evaluate prediction accuracy given that a path was observed. Paths that did not propagate are not captured by RBN or PSKR, creating inherent survivorship bias. VOACAP's circuit failure predictions (< −100 dB) may be correct for unobserved paths. Training overlap: 1,426 WSPR spots from JD04 of 10.8 billion (0.00001%) — a near-blind test for IONIS.