See the return flight before it takes off
I spot the jet about to fly home empty before it even lands, and give the broker hours of lead over the entire market.
The context
In business aviation, a jet that has dropped off its client often flies back empty to its base. The industry calls it an empty leg: a seat to sell that nobody sees, and that vanishes within hours. I co-developed Empty Leg Radar with about ten US brokers who tested the tool in real conditions and calibrated the thresholds with me.
Detection 2 to 4 hours before departure. The list of operators to call for a given route comes out in about half a second, versus 30 to 45 minutes by hand. A 90-day history window, an adjustable trigger threshold, and every lead tracked in a CRM through to the close. The broker gets a qualified alert, with the operator and the approach angle already prepared.
The problem
Spotting these flights by hand takes 30 to 45 minutes per lead: cross-referencing flight plans, recalling an aircraft's history, guessing whether it will fly back. By then the signal is already cold. The window is too short for a manual process, and without structured follow-up half the leads are lost.
How it works
The pipeline has four stages. Ingestion continuously reads traffic from the monitored airports and writes each flight as a raw event. Normalization aggregates these events into one clean row per flight, keeping the most reliable time available: the actual runway time over the revised time, and the revised time over the scheduled time. Scoring reads this table over a short window and detects aircraft that arrive then quickly leave the same airport. Sending pushes signals above the threshold to the broker and logs everything, with an anti-duplicate guard so the same alert is never re-emitted.
Key decisions
- A stable flight identifier, built by priority of registration then Mode-S code then call sign, to deduplicate feeds that often contradict each other.
- A hybrid model: a rule-based core captures the simple rule, an aircraft that leaves again in under two hours is a signal worth attention, and a regression trained on history adjusts the score with behavioral variables: empty-return rate, frequency of short stopovers, distance to base, median block time.
- A bonus when the return targets the operator's base, those bases being learned automatically from history, not entered by hand.
- A materialized view so the search returns in about half a second, because beyond two seconds a broker goes back to their calls.
- A built-in CRM, Open to Contacted to Won or Lost, because a signal without follow-up is lost.
What it proves
The signal was sitting in traffic data anyone can see. My work wasn't to find it, but to know which one was worth building, in an opaque vertical nobody instruments, then to deliver it as a complete system, from capture to CRM. That's exactly what I set up for a client.