Skip to content

Prediction Quality (Pearson Correlation)

Date: 2026-02-09 Status: COMPLETE Result: IONIS showed +0.35 correlation improvement over reference model

V20 Production

V20 Golden Master achieves Pearson +0.4879 — a substantial improvement from the +0.3675 measured in this initial test.


Objective

The IONIS vs VOACAP comparison proved IONIS correctly identifies if a band is open (Recall). This test proves IONIS tracks the quality of the opening (Correlation).

If IONIS has higher Pearson r than VOACAP against ground truth SNR, it doesn't just know the band is open—it knows how strong the signal will be.


Methodology

Test Dataset

  • Source: validation.quality_test_paths — 100K signatures from wspr.signatures_v2_terrestrial
  • Stratification: 10K per band (160m through 10m)
  • Quality filter: spot_count > 50 (high-confidence ground truth)
  • Ground truth: median_snr from signature bucket

Dual Prediction

Model Method
IONIS oracle_v13.py inference on M3 Ultra
VOACAP Method 30 (SNRxx) via voacapl on 9975WX

Metrics

  • Pearson r: Correlation between predicted SNR and actual median_snr
  • RMSE: Root mean squared error (IONIS only; VOACAP has unit incompatibility)

Results

Overall Correlation

Metric IONIS VOACAP Delta
Pearson r +0.3675 +0.0218 +0.3456
RMSE 5.00 dB
Bias -2.00 dB

IONIS Pearson r = +0.3675 vs VOACAP r = +0.0218 (delta: +0.3456).

Per-Band Breakdown

Band N IONIS r VOACAP r Delta
160m 10,000 +0.2948 -0.1950 +0.4898
80m 10,000 +0.2664 -0.2100 +0.4764
60m 10,000 +0.2449 -0.1817 +0.4265
40m 10,000 +0.4214 -0.1717 +0.5932
30m 10,000 +0.2997 -0.0892 +0.3889
20m 10,000 +0.3850 -0.0205 +0.4054
17m 10,000 +0.4993 -0.0258 +0.5251
15m 10,000 +0.4979 +0.0468 +0.4511
12m 10,000 +0.3391 +0.2691 +0.0700
10m 10,000 +0.1244 +0.1826 -0.0583

IONIS showed higher correlation on 9 of 10 bands.

Low-Band Analysis (160m-40m)

VOACAP shows negative correlation on low bands—when it predicts stronger signal, actual WSPR SNR is lower.

Model Pearson r
IONIS +0.3204
VOACAP -0.1963
Delta +0.5167

Physics Interpretation

Why VOACAP is Anti-Correlated on Low Bands

When VOACAP was developed (1960s-70s), the "noise floor" was defined by human operators with headphones. A signal at -28 dB SNR wasn't "weak"—it was non-existent.

VOACAP was calibrated for Information Throughput: when is a signal strong enough for 60 WPM Teletype or clear SSB voice? It effectively ignores the "Sub-Audit Floor."

On low bands, VOACAP overcompensates for D-layer absorption. It assumes high absorption = "dead" path.

The WSPR reality: Paths stay open deep into the noise floor. Digital modes like WSPR and FT8 decode signals that were invisible to 1970s technology.

The Resolution Difference

Model What it models
VOACAP The Mirror — can it reflect a high-power beam?
IONIS The Medium — can any energy get through?

IONIS learned the "Deep-Tissue Physics" of the ionosphere from 10.8B observations. It knows that absorption is a dimmer switch, not an off switch.


Files

File Location
Test script ionis-training/scripts/quality_test_ionis.py
Test paths validation.quality_test_paths (ClickHouse)
VOACAP results validation.quality_test_voacap (ClickHouse)

Generated: 2026-02-09 IONIS — Ionospheric Neural Inference System (V20 Production)