Fusing SWIM, ADS-B, and Weather into One Feature Space

Pull-quote: “The model is rarely the bottleneck in disruption prediction. The picture is. Keying, alignment, and time discipline decide how good the forecast can ever be.”

The picture is the product

Ask why a disruption model underperforms and the answer is almost never the architecture. It is the operational picture underneath: which feeds it fuses, how they are keyed, and whether its features are honest about time. A system built for disruption prediction fuses the FAA SWIM flight and flow feeds, SFDPS, TBFM, and TFMS, with continuous ADS-B surveillance and METAR, TAF, and NOAA weather models into one feature space, re-scored every sixty seconds. The engineering that decides forecast quality lives in three unglamorous places: entity resolution, temporal alignment, and leakage control.

Five feeds, five vocabularies

SWIM is not one feed. It is a family of services describing the same airspace from different institutional viewpoints, at different cadences, under different keys.

The disagreements between feeds are themselves signal. A flight whose filed intent and surveillance actuals have diverged is telling you something no single feed can say. A metering time that keeps slipping against a stable flight plan says the constraint is on the airspace, not the aircraft. Fusion is what makes those comparisons possible in the first place.

One flight, three names

The same physical flight arrives keyed three different ways: a flight plan identifier in SFDPS, a metering entry in TBFM, a surveillance track in ADS-B. Fusion begins with a flight entity that survives amendments, tail swaps, and diversions, so every downstream feature attaches to the right object. Get this wrong and the model trains on chimeras: features from one flight, labels from another.

The fusion spine

SFDPS ──┐
TBFM  ──┤   parse and     resolve to       as-of join on      feature space:
TFMS  ──┼── validate ───► one flight ────► event time, ─────► flight, airport,
ADS-B ──┤   each feed     entity           with watermarks    network context
WX    ──┘                                                          │
                                                                   ▼
       ◄──────────────── re-score every sixty seconds ────── prediction engines

Temporal alignment is the second discipline. Every feature is joined as of prediction time, on event time rather than arrival time, with late and out-of-order messages handled by watermarks instead of silent overwrites. The third discipline is leakage control: a training feature must reproduce exactly what the live system knew at scoring time. A backtest built on corrected, settled data flatters every model it touches, and the flattery evaporates in production.

The feature space itself has three layers. Flight-level features come from the entity: filed route, current position, inherited delay. Airport-level features aggregate the local picture: demand against capacity, active initiatives, observed and forecast weather. Network-level features describe the system state around the flight, because the national airspace behaves like a network, not a collection of independent departures.

Sixty seconds, deliberately

The re-scoring cadence is a design decision, not a race. Sixty seconds is fast enough that a new ground stop, a metering revision, or an amended terminal forecast enters the picture before the people it affects have finished reading about it, and slow enough that each cycle produces a stable, archivable snapshot. The archive is the quiet payoff: every snapshot is a training example with point-in-time correctness built in, which is how a training corpus assembled from live snapshots and years of DOT and BTS performance archives stays honest at scale.

Calibration completes the loop. Every re-scored output ships wrapped in conformal calibration, so the interval around the forecast states how much the picture actually knows at that minute, not how confident the architecture feels.

Closing

Fusion is the unglamorous majority of aviation prediction. Resolve the entities once, align on event time, refuse leakage, and re-score on a cadence you can defend. The engines running on top of the picture, distributional prediction, causal analysis, and cascade models, inherit their ceiling from this floor. Better fusion raises what all of them can achieve.