NOTAM, PIREP, and SIGMET as Structured Risk Signals
Pull-quote: “An advisory a machine cannot read is a hazard the forecast never sees. Distillation turns teletype prose into a signal with a geometry, a window, and a magnitude.”
Written for people, needed by machines
A NOTAM is abbreviated, all-caps text in a format that predates the computer. A PIREP is a pilot’s report squeezed into a coded string. A SIGMET is an area advisory describing hazards across thousands of square miles. All of it was designed for trained humans under time pressure, and all of it carries operational risk that a prediction system needs. The advisory layer is central enough that when the national NOTAM system failed in January 2023, departures paused nationwide.
The practitioner answer is to stop treating advisories as documents and start treating them as a feed. Four stages: parse, normalize, classify, join.
Four stages of distillation
Parse
Structure first: location identifiers, validity windows, coded fields where they exist, and the free-text field where the operational content usually lives. Parsing means extracting fields reliably, not understanding meaning yet.
Normalize
Every advisory becomes a geometry and a time window. Geometry might be an aerodrome point, a radius, a polygon, or a flight information region. Time includes recurring schedules, an advisory active daily for six hours is not the same exposure as one active continuously.
Classify
The taxonomy of operational effect. For NOTAMs: runway or taxiway closures, navaid and lighting outages, airspace restrictions, obstacles. For PIREPs: observed turbulence and icing with altitude band and severity. For SIGMETs: convective activity, turbulence, icing, and volcanic ash, each a hazard area with a class and a validity window.
Join
Exposure is the product. Intersect advisory geometry and windows with flight routes and times, and each flight acquires per-advisory risk features, while each airport acquires counts and severities by class.
Classification also has to carry magnitude. A taxiway closure and a runway closure are both closures, but they do not move a delay forecast by the same amount, and a moderate icing report on an active arrival path matters more than a severe one far from any route. The distilled signal is class, geometry, window, and weight together, which is why flattening advisories into counts per airport gives away most of their value.
raw advisory ──► parse ──► normalize ──► classify ──► join ──► risk features
NOTAM, PIREP, fields geometry operational route and per flight and
SIGMET and free and validity effect window per airport,
text window class exposure with magnitudes| Advisory | Written as | Distilled signal |
|---|---|---|
| NOTAM | Abbreviated all-caps text with coded fields | Closure, outage, or restriction with a place and a window |
| PIREP | Coded pilot report | Observed turbulence or icing with altitude band and severity |
| SIGMET | Area advisory | Hazard polygon with class and validity window |
The failure modes
The pipeline earns its keep on the ugly cases. Cancellations and replacements must retire the advisories they supersede, or the picture accumulates ghosts. Duplicates arrive through different channels. Some advisories expire in text but not in practice. Coordinates appear inside prose. Negations invert meaning in a single token. The operating rule: measure parser coverage continuously, keep unparsed advisories visible as exceptions, and never silently drop one. An advisory the pipeline could not distill should degrade confidence in the affected area, not vanish.
Where the signals land
Distilled NOTAM, PIREP, and SIGMET signals belong in the same feature space as the SWIM feeds, ADS-B surveillance, and weather, re-scored on the same cadence as the rest of the picture. They do a different job from the safety corpus: the NTSB and ASRS archives, hundreds of thousands of reports embedded for semantic search, serve retrospective retrieval and cited answers for analysts. Advisories serve the live picture. Two different jobs, one standard: text becomes structure before it becomes intelligence.
Closing
Advisories are the oldest data in aviation and the last to be modernized. A distillation pipeline, parse, normalize, classify, join, turns them into risk signals a model can actually consume, with coverage measured and exceptions visible. The alternative is a forecast that reads everything except the warnings.