How we predict your upgrade odds — and how often we're right.
Upgrade Pro turns live airline data — waitlists, fare buckets, seat maps — into a calibrated probability you'll clear the upgrade list. Here's exactly how it works — and how we hold ourselves accountable, hits and misses alike.
Reading a search at a glance.
Every search opens with a five-day view. Each day breaks into morning, afternoon, and evening bars — height and color show the best upgrade odds in that window — with the day's headline number underneath. Tapping a bar filters the results to that day and time of day.
Below it, every flight is one row: schedule, aircraft, open premium seats against cabin size, fare, and the upgrade score — the model's calibrated probability with its honest range — plus one-tap booking on the airline's own site, saving to My flights, and an expandable factor analysis.
4h 41m · nonstop Airbus A321neo 10 / 20 $486
10 of 20 premium seats open now (~6 expected by departure); you project around #3; modeled from typical elite demand on this route (no live waitlist yet) — a real contender — reasonable to book, unwise to count on.
Illustrative example with sample numbers — run a search to see live odds for your status.
Built by a road warrior, for road warriors.
I spend a serious share of my life in seat 12C, hoping it turns into 2A. I also spent years building machine-learning models for a living.
So I got tired of folk wisdom and gate-agent tea-leaf reading, and built a calibrated model to tell me which of my own flights were actually worth booking. It changed how I book — sharing it felt obvious.
Upgrade Pro tells you the truth the way I wanted to hear it: a probability with honest error bars, not a marketing number.
What a verdict actually promises.
Every prediction lands in one of four confidence bands, set by the model's calibrated estimate. The ranges below are the odds you can expect when you see each label on a flight.
Book it expecting the upgrade. Still not a guarantee — roughly 1 in 4 will miss.
Leans your way: reasonable to book, unwise to count on.
Possible, not probable — the odds are genuinely against you.
Don't count on it clearing. Book this flight for the schedule, not the seat.
The method, in four steps
We read what the gate agent sees
Live standby queues, fare-bucket inventory (J, C, D, PZ, PN), seat maps, and aircraft type — polled repeatedly as departure approaches. Where an airline blocks automated reads, we fall back to an estimated profile and label it plainly.
A calibrated model scores it
A Random Forest classifier weighs your queue position, route competitiveness, departure timing, and route history, then isotonic calibration turns its raw score into an honest probability — so when we say 52%, we mean about 52% of those flights actually clear. We don't headline a single accuracy number: figures like that are easy to inflate and hard to verify. What matters is whether the probabilities hold up against real gate outcomes — which we measure against every tracked flight as it clears or doesn't.
The band is the honest part
We never show a bare number. The band is the spread of the model's internal trees — wide when they disagree, tight when they agree. "Contender, 44–54%" is a real contender that leans your way: reasonable to book, unwise to count on.
Every prediction gets graded
Each prediction is stored with the flight, then checked against the standby list after departure: did that passenger, at that position, actually clear? Grading is per passenger, not "did anyone get upgraded" — those are different questions and only the first one is what we told you.
We publish calibration and a Brier score once enough flights have been verified this way. Until then there is no number here, because a figure built on a handful of flights would be noise dressed up as accountability.
- Irregular ops (weather, aircraft swaps) invalidate predictions; we flag rather than guess
- Invite-only tiers — United Global Services, American ConciergeKey, Delta 360° — are opaque and can jump any queue
This product is young, and it will say so.
Upgrade Pro is in its early stages. Coverage deepens airline by airline, the model retrains as every tracked flight becomes new training data, and verified outcomes accumulate slowly — a flight only counts once we have both a live standby queue before departure and a confirmed result after it. Expect rough edges, expect visible improvement — and when something is estimated rather than live, expect us to label it instead of dressing it up.
If something looks wrong, that feedback genuinely shapes what gets built next.