← Back to the Sunset & Sunrise forecast  ·  TheClimatology home

Firesky Algorithm vs Photos

Some photos courtesy of EarthCam — the world's leading live-camera network, streaming iconic skies from the Empire State Building, New York Harbor, and thousands of other locations 24/7. Their unmatched archive made this entire verification dashboard possible. Go explore earthcam.com — it's spectacular. Images © EarthCam and respective rights holders, used here for research/educational comparison only. If you are a rights holder and want anything removed, please reach out and we'll take it down immediately.
What do these numbers actually mean?
Probability %

The algorithm's predicted chance of seeing firesky over NYC at sunset. Discrete values (20 / 40 / 50 / 75 / 90) from a decision tree. ≥75% means "expected to occur".

Quality /10

The visual quality of the firesky when it occurs. Q=10 is intensely vivid; Q=4 means muted colors even when present. Only meaningful when probability ≥ 50% — if the sky is overcast anyway, Q has no use.

Ground truth

The photo source determines what actually happened each day:
★ green firesky captured on camera (72 days)
unmarked monitored, no firesky (67 days)
EarthCam Empire State Building / NY Harbor feeds serve as objective evidence.

Confusion matrix

TP algo says yes + actually yes ✓
FN algo says no + actually yes (missed prediction — worst case)
FP algo says yes + actually no (false alarm — wasted wait)
TN algo says no + actually no ✓
At ≥75% threshold → 87.1% accuracy.

Accuracy = (TP+TN) / total. Fraction of all calls that were correct. Recall = TP / (TP+FN). Of days that truly had firesky, what fraction the algo caught (avoids misses). Precision = TP / (TP+FP). Of days the algo said "yes", what fraction were actually firesky (avoids false alarms). F1 = harmonic mean of Recall and Precision. Single balanced score — closer to 1.0 is better.