VERIFICATION PIPELINE
Why results may still be warming up
ROLLING ACCURACY
How the model is tracking
BY BAND
Forecast accuracy
| Band | Matches | Forecast | Observed | MAE | Bias | Correlation | Within ±10 |
|---|
CALIBRATION
Does 70 mean about 70?
| Forecast bucket | Average forecast | Average observed | Bias | Samples |
|---|
CALIBRATION EXPERIMENT
Out-of-sample WSPR evidence calibration
Evaluating historical forecasts…
| Band | Training cells | Test cells | Raw MAE | Calibrated MAE | Improvement | Decision |
|---|
Retrospective experiment: earliest 70% of observed UTC hours for training, latest 30% for testing. Measures WSPR evidence conditional on a detected path, not the probability of an SSB contact. Forecast output remains unchanged.
OVER-PREDICTIONS
Model too optimistic
UNDER-PREDICTIONS
Unexpectedly good
BY TARGET REGION
Where the model is strongest and weakest
| Region | Matches | Forecast | Observed | MAE | Bias | Correlation |
|---|
OBSERVED WSPR HISTORY
Band performance
TIME OF DAY
Strongest UTC hours by band
| Band | Best UTC hour | Average score | Evidence paths | Samples |
|---|
WSPR TARGETS
Best historical target regions
| Band | Region | Average score | Evidence paths | Samples |
|---|
RBN CORROBORATION
Broad regional CW evidence
| Band | Average score | Spots | Samples |
|---|