Dataset Mode:
Station:
Parameter:
FORECAST ENGINE
bolt
Loading...
Default: 7-day rolling bias blend
LightGBM retained as alternative model
STALENESS GUARD
shield
Checking...
Newest Obs: -- | Recent 7d: --
Automatic raw CAMS fallback on stale data
P(SEVERE) RISK INDICATOR
warning
24h
--
48h
--
72h
--
Uncalibrated indicator (joint PM2.5/PM10)
UNCERTAINTY INTERVALS
expand
p10 · p50 · p90 Spread
Lead scaling:
(1.0 + 0.004 × h)
Cones widen dynamically through 72 hours
Loading station...
--Issue Time: --
p50 Forecast
p10–p90 Range
Actual Observed
72-Hour Air Quality Index (AQI) Trajectory
Showing hourly quantiles (p10, p50, p90) with lead-dependent widening uncertainty through h=72
Unit: Index Value (CPCB NAQI)
history_edu
Ground-Truth Matched
Historical Replay Evaluation (Forecast vs Actual Ground Truth)
In Replay mode (--issue-time "2025-11-12 20:00"), the 72-hour forecast issued into held-out winter 2025–26 is evaluated against verified hourly observations recorded at the station.
VERIFIED HOURLY OBS
72 / 72 hrs
BIAS CORRECTION
Applied (145 obs)
ACTUAL AQI RANGE
380 – 500
CATEGORY MATCH
Within ±1 Tier: 83.8%
Hourly Forecast Data Table (72-Hour Horizon)
Detailed quantiles, lead hour factor, and observation status
verified
OUT-OF-SAMPLE VERIFICATION (docs/verification.json)
Six-Winter Audit & 2025–26 Benchmark Performance
PM2.5 CAMS BC MAE
50.27 µg/m³
-27.7% vs Raw CAMS (69.52)
PM10 CAMS BC MAE
107.35 µg/m³
-44.1% vs Raw CAMS (192.09)
CATEGORY ACCURACY
44.3% (±1: 83.8%)
Raw CAMS: 23.9%
BRIER SKILL (P(Severe))
+0.181 vs Persistence
Zero skill vs climatology