↖ CPA Weather Lab

KCXY / KMDT — April 2026 Outage Diagnostic

GBR neighborhood regression · 2018–2025 base period · 5 ASOS+CWOP neighbors
KCXY outage: 2026-04-22 13:56–15:55Z (~2h) Apr 20–21: 5× >3σ cold anomaly at CXY Post-outage: elevated but NOT MDT-clone CV R² ≈ 0.994
PRE-outage resid_diff
−1.46°F
mean z = −1.20 · n=86
POST-outage resid_diff
−2.36°F
mean z = −1.95 · max |z| = 2.64
April norm (2018–2025)
−0.02°F
± 1.20°F (2σ: −2.42 to +2.37)
5-min identity rate
68%
KCXY within 0.2°F of KMDT — stable pre/post
Peak cold anomaly
−4.86°F rd
Apr 20 06Z · z = −4.03 · LEGITIMATE CAD
KCXY–KMDT correlation
0.9952
72,000 shared hours · 2018–2026

Chart 1 — Temperature Traces: KCXY · KMDT · KCXY_hat (neighbor model)

Apr 19–23 2026 (UTC) · Hourly · Yellow = outage window · Blue shading = cold drainage events
KCXY actual
KMDT actual
KCXY_hat (model)
Outage gap
Cold drainage
Cold drainage signature (Apr 20 06–08Z, Apr 21 01–02Z): KCXY runs 4–5°F below KMDT and 4°F below the neighborhood model — z scores −3.6 to −4.0. This is real physics: Camp Hill valley floor exhibiting nocturnal cold-air pooling under clear-sky radiative cooling while MDT's more exposed runway site retains more heat. Historically this happens ~0.8% of April hours. Two back-to-back nights makes it notable but meteorologically coherent.

Chart 2 — Residual Difference (KCXY − KMDT neighborhood gap)

rd = (KCXY − KCXY_hat) − (KMDT − KMDT_hat) · April 2σ envelope shown · Positive = CXY warmer than expected relative to MDT
resid_diff
Apr IQR (p25–p75)
Apr ±2σ
Apr norm (−0.02°F)
Post-outage verdict: NOT an MDT-clone event. Post-15:55Z resid_diff mean = −2.36°F (z = −1.95), which sits just below the 2σ lower bound (−2.42°F). Three of five scorable post-outage hours breach −2σ, but none breach −3σ. The pattern is consistent with a sensor that resumed normal CXY behavior but was running in a warm-advection environment where CXY traditionally runs slightly warmer relative to its neighborhood — making the negative residual look anomalous. Compare to the PRE period: the mean was already −1.46°F, so the post-outage shift is ~0.9°F colder, not the multi-degree collapse you'd expect from a true MDT-clone.

FINDINGS SUMMARY

01
The April 22 outage is 2h gap (13:56–15:55Z) — not in your outage file yet since it's 2026. The sensor went completely silent for 119 minutes then resumed cleanly. Only 1 scorable post-outage hour (15:00Z) falls inside the gap window, and its z-score is +0.32 — entirely normal.
02
The "weird MDT behavior" you noticed is real but modest. Post-outage hours 16Z–20Z have resid_diff between −1.64 and −3.19°F. The −3.19 at 20Z (z=−2.64) is the only genuine outlier. At 5-min resolution, KCXY and KMDT were tracking within 0.2°F of each other ~69% of the time — but crucially that rate was already 68% before the outage. The sensor didn't suddenly start mimicking MDT; it was already doing it.
03
The real story is April 20–21 cold-air drainage. Five hours exceed 3σ on those two nights, with the peak at Apr 20 06Z: z=−4.03. KCXY ran 4.6°F colder than KMDT while the neighborhood (KLNS/KTHV/CWOP) predicted ~43°F. That's textbook Camp Hill valley floor behavior under clear skies and light southerly flow — genuine local microclimate signal, not sensor error.
04
The 90%+ overlap statistic is a dataset artifact. KCXY and KMDT's overall 2018–2025 Pearson r = 0.9952 across 72,000 hourly pairs. That's normal for co-located mid-Atlantic ASOS stations. The "overlap" your pipeline flagged reflects the high correlation, not a data-feed contamination problem. CW0297 (East Petersburg) is the outlier at r=0.77 — and that's explained by its 35% missing-data rate and more easterly Lancaster County exposure.
05
Action items: (1) Add the Apr 22 13:56–15:55Z outage to your outage file. (2) Flag the Apr 20 06–08Z and Apr 21 01–02Z cold-drainage events as valid extreme microclimate, not sensor errors. (3) The post-outage warm-advection hours need KLNS/KTHV extended past 20Z to score fully — consider pulling a KLNS/KTHV 2026 update through April 23. (4) The 68% 5-min identity rate deserves a longer 2026 baseline audit — if it's been above ~50% for months, that's an interesting instrument drift / calibration question.
Model: GradientBoostingRegressor · n_estimators=300 · max_depth=4 · lr=0.05 · random_state=42
Features: KLNS · KTHV · CW1439 · CW7253 · ar934 · sin/cos hour · sin/cos doy
Training: 2018-01-01 – 2025-12-31 · non-outage hours only · KCXY n=19,476 · KMDT n=19,477
April seasonal norm (2018–2025): mean=−0.025°F · std=1.198°F · p5=−1.75 · p95=+2.01
← CPA Weather Lab · Learn