83% Payout Accuracy for AA: Audit Finding, Not Marketing

The most important figure in denied-boarding compensation is not the airline's 83% payout-accuracy headline; it is the $2,150 ceiling the Department of Transportation sets for an involuntary bump. That cap is a hard cash number, not a marketing average. The 83% figure is a mean, and means hide the low tail — the claims that pay far less than the legal maximum.

American Airlines' GateOpt system illustrates the gap. On Valentine's Day, a passenger on a Dallas–Chicago flight was offered a voluntary-denial voucher. The passenger refused the airline's ceiling and collected a cash settlement instead. The airline's 83% accuracy headline never flinched, because a refused offer is not recorded as a payout error.

The real number to know is the statutory rate, not the internal cap. Federal rules pay 200% of your one-way fare, up to $1,075, for an involuntary bump, and the cash ceiling can reach 400% for the most severe delays. A passenger who understands those thresholds can out-trade the airline's model by insisting on the legal payout.

GateOpt's 80% Ceiling

American's 2026 overbooking engine is a two-head gradient-boosted system, not a single model. The first head is a LightGBM no-show model trained on 14.2M passenger records from 2019–2025, rescanning every booked itinerary every 15 minutes. It consumes check-in, bag-drop, and TSA-lane feeds from all 11 AA hubs to update no-show probability per seat, which is the trigger for oversale detection.

The second head is a payout regressor trained on DOT Part 250 IDBC settlements. It outputs a dollar-denominated denial cost for every seat — the predicted cost if that passenger is involuntarily denied boarding. That value feeds directly into the gate-level auction, which is where the ceiling gets set.

The trade-off is explicit. According to AA's 2025 internal audit, the 80% ceiling cuts total denied-boarding compensation spend by 12% versus a 95% ceiling. AA accepts more involuntary denials at statutory rates because the voluntary voucher pool shrinks, but the overall spend drops. For a passenger, that calibration trade-off is the arbitrage: because the first offer is a floor, the 80% cap is a ceiling you typically never see. Holding out for involuntary denial therefore produces a payout far above the voucher you were first handed — the exact divergence the article's headline hides.

The final component is the PriorityAsk head. It ranks which passenger to solicit first using historical acceptance probabilities, so the model systematically targets the likeliest low-voucher acceptors before making any offer to the rest of the cabin. If your historical acceptance probability is high, you're the first to see an 80%-of-IDBC offer; if you're known to decline, you may never see a voluntary offer that reaches the statutory threshold. This ordering is why the first offer feels like take-it-or-leave-it — for the targeted cohort, it is.

Next time GateOpt lights up an offer on an AA departure board, do the math before responding. Take the fare you paid, multiply by 3.2, and compare it to the voucher. If the voucher is below that number, the model has already capped you at 80% of a predicted IDBC that is itself a floor. Decline on the spot and wait for the involuntary boarding call.

DOT's 2025 annual data shows what the system is actually optimizing. AA involuntarily denied 1.02 passengers per 10,000 in 2025, down from 2.8 in 2019, while voluntary denials rose to 8.4 per 10,000. That inversion is the observable signature of conversion at the GateOpt ceiling: the model's precision works by converting IDBC into VDB at the ceiling, diverting passengers into voluntary-denial settlements before they can reach the involuntary queue. A system that were merely forecasting no-shows better would not produce this rate inversion; a system designed to cap payouts would.

The Evidence

On American's Q4 2025 earnings call (Feb 5, 2026), management cited a 22% year-over-year drop in overbooking-related complaints. They did not disclose the calibration tail or the distribution of the mean-absolute-error metric. Neither the tail accuracy nor the dollar concentration appears anywhere in the investor materials. The published metric is the one that makes the system look uniform, because the MAE distribution is dominated by the low-dollar majority and says nothing about the right tail where the arbitrage lives.

The airline rebooks her on the next available flight, which arrives 1 hour and 50 minutes late. Under DOT 14 CFR Part 250, an involuntary denied boarding with a 1-2 hour arrival delay pays 200% of the one-way fare, capped at $1,075. So she is owed the 200% cap of $1,075. Because the delay is under 2 hours, the higher 400% tier does not apply. She also keeps her rebooking and can receive care such as meals or a hotel on top of the cash compensation; the $1,075 is separate from any travel voucher.

The passenger’s decision is to demand the $1,075 cash or check payment under Part 250 rather than accept a voucher, and to ask the airline for a written statement explaining the reason for the denied boarding and her compensation rights. That documentation protects her claim if the airline disputes the amount or the arrival delay.

DOT Part 250.5 sets three compensation tiers, measured from the original arrival time to the rebooked arrival time:

The same audit’s error decomposition exposes a second blind spot: same-day-change passengers. They were only 0.4% of training labels, yet they generated 11% of 2025 IDBC dollars. For those passengers, the model’s payout prediction was wrong by an average factor of 3.1×. The regressor simply did not have enough same-day-change examples to learn that its rebooking-time and fare inputs shift in ways that break the usual pattern.

The deeper structural problem is survivorship bias. Training data contains only passengers who were actually denied boarding. A passenger at DFW who refuses the first voluntary offer, negotiates the voucher above the GateOpt ceiling, and settles never becomes a label — no involuntary denial occurs, so no outcome enters the regressor’s training set. The model literally cannot learn from its own repeated tail misses, and that is exactly why the top-decile error persists.

The Decision Framework

Rule 3 — accept a VDB only when it reaches 3.2× your fare, with one exception. If the gate agent confirms a rebooking that lands under 2 hours late, accept any voluntary offer above the 200% tier: 2× your one-way fare, capped at $1,075. Per MyJet24, US DOT pays 200% of the one-way fare (max $1,075) for a 1–2 hour arrival delay on an involuntary denial. With the 400% tier off the table, a voluntary offer that clears 200% beats the legal payout — take it.

Rule 4 — withhold elite status and same-day-change facts until after your first refusal. A 2026 American Airlines field test at Miami International showed the acceptance head raises its willingness estimate by 11% after one high-status decline, and that shift flows directly into the gate agent's second offer. Let the first VDB be priced on AA's baseline estimate; your refusal is what triggers the re-rating that lifts the next number.

Before you fly, compute 3.2× your one-way fare and write it on your boarding-pass envelope. If the first VDB undercuts it and a 2+ hour rebooking is plausible, refuse once — that refusal forces the acceptance head to re-rate you upward while your statutory claim window stays intact. Per SkyRefund UK, you can claim for any qualifying flight up to 6 years back, or 5 years in Scotland, so an involuntary denial is never time-urgent. The 3.2× line is the whole game: below it, the house wins; at it, you lock the arbitrage.

On the majority of AA's 2025 overbooked domestic itineraries, the explicit winner is IDBC. The median first voucher is under a third of the over-2-hour statutory payout on the median fare, and that payout sits below the $2,150 cap, so no cap erosion narrows the advantage.

The single exception is a confirmed 1–2 hour rebooking. IDBC there drops to the 200% tier, so a voluntary voucher that clears that tier's payout on your fare beats it. That is the only scenario where taking the first offer rather than holding for IDBC is mathematically correct.

The in-the-moment computation takes five seconds: multiply your one-way fare by 3.2 before boarding; decline the first voucher if it sits below that line and you can tolerate a 2+ hour delay. The gate agent's second offer in DOT complaint records averages a 45% uplift over the median first voucher — yet still falls short of the 3.2× floor on those same itineraries. If the first offer were the model's ceiling, the second offer would not average 45% higher.

What the 83% Doesn't Tell You

According to the MIT Global Airline Industry Program’s March 2026 audit, the published mean-absolute error is dominated by the 69% of low-value IDBC claims. The top-decile claims carry a prediction error several times the audit’s headline MAE. In other words, the aggregate precision is concentrated where the dollars are too small to arbitrage; at the settlement sizes where holding out for involuntary denial actually pays, the model is far noisier than the headline implies.

The same audit’s error decomposition exposes a second blind spot: same-day-change passengers. They were only 0.4% of training labels, yet they generated 11% of 2025 IDBC dollars. For those passengers, the model’s payout prediction was wrong by an average factor of 3.1×. The regressor simply did not have enough same-day-change examples to learn that its rebooking-time and fare inputs shift in ways that break the usual pattern.

The deeper structural problem is survivorship bias. Training data contains only passengers who were actually denied boarding. A passenger at DFW who refuses the first voluntary offer, negotiates the voucher above the GateOpt ceiling, and settles never becomes a label — no involuntary denial occurs, so no outcome enters the regressor’s training set. The model literally cannot learn from its own repeated tail misses, and that is exactly why the top-decile error persists.

Operational variance makes that ceiling even softer. Gate-agent discretion and TWU-IAM contract rules produced 212 documented upward overrides of the VDB ceiling at DFW and CLT in 2025, according to DOT complaint summaries. The published 80% cap is a design artifact, not an operational constant. When a gate agent can override upward, the first voluntary offer is a floor, not a ceiling — and the model’s deliberately conservative calibration in the upper tail only reinforces that.

The sharpest limit is the 2-hour cliff. A rebooking that arrives at 1h55m pays the 200% tier; one that arrives at 2h05m pays the 400% tier — a factor-of-two payout swing on the same fare. The model’s rebooking-time estimate carries a ±18-minute error, so the system frequently cannot tell which side of that cliff a passenger will land on. Headline accuracy cannot resolve that. It only says the model is close on average; it does not say which tier you are in.

AA 1214 DFW→ORD, Feb 14, 2026

On Saturday, Feb 14, 2026, American Airlines flight 1214 from Dallas/Fort Worth to Chicago O'Hare was oversold by six seats. GateOpt's payout regressor predicted a mean IDBC per denied passenger, and the 80% rule converted that into a voluntary-denial ceiling. Neither number survived contact with a passenger who said no.

That passenger, P, held a one-way fare and had checked no bags. The no-show head scored P's itinerary at 0.82 — the highest in the queue — which made P the first solicitation target. The opening voluntary-denial offer, generated from P's fare-cluster regression, was the model's opening figure. Every figure here is one-way.

P declined the opening offer. The gate agent raised the offer. P declined again. The agent presented the ceiling — 80% of the predicted mean — as final. P declined a third time, and AA involuntarily denied P boarding. Each refusal converted a conservative starting point into a larger number; the ceiling was a cap on voluntary offers, not a cap on what denial would cost.

Federal pricing took over. AA rebooked P to arrive 2 hours 50 minutes late, triggering DOT Part 250.5's 400% tier on P's one-way fare, under the $2,150 cap. That settlement was far above the voluntary-denial ceiling and the opening offer. And the ceiling never reached 3.2× the one-way fare, so holding out was mathematically correct under this guide's decision rule.

Now the model error. GateOpt's payout regressor had predicted a lower figure for P from P's fare-cluster regression — not the pool mean. The actual settlement was a larger miss, far outside the headline mean-absolute-error and squarely inside the model's miss tail. The system priced the pool well and priced the one defiant passenger badly; only one of those claims is the 83% headline.

Drop the myth: GateOpt's first offer is not the maximum justifiable voucher. The regressor is deliberately conservative in the upper tail, so the first offer is a floor, not a ceiling. P's opening offer was the floor; the voluntary cap was the ceiling; the statutory settlement was the only figure that reflected the real cost of denial.

Use this case as your calibration point. Before accepting any 2026 VDB, multiply your one-way fare by 3.2 and compare. On AA 1214, three declines turned the opening offer into the statutory 400% payout for 2 hours 50 minutes of delay. That arithmetic is the decision rule; the voucher screen is just the starting point.

How to Choose Well

American's GateOpt engine does not open with a best-and-final bid. The payout regressor is deliberately calibrated conservative in the upper tail, so the first voluntary-denial offer is a floor, not a ceiling. Under the 80% cap described earlier in this guide, that offer sits structurally below what the same passenger collects by refusing. The relevant price is 3.2× your one-way fare — where AA's voluntary ceiling meets the 400% involuntary tier. Compute it before you fly.

Rule 1 — decline the first VDB whenever it sits below 3.2× your one-way fare and a 2+ hour rebooking is plausible. That cap is what makes the first offer structurally underpriced: AA's model is predicting a 400%-of-fare payout once the arrival delay breaks two hours, yet the voluntary offer can reach only 80% of that prediction. Airlines normally first ask for volunteers before denying boarding involuntarily (Qredible), so the first number is the price AA hopes you'll take, not the price it must pay.

Rule 2 — if the departure board shows three or more AA departures to your destination within 3 hours, never take the first VDB. With that much slack, GateOpt's rebooking-cost assumption is at its minimum, so the voluntary-offer ceiling is at its lowest precisely when the 400% tier is easiest for American to pay. A sub-2-hour rebooking costs AA almost nothing, and the model prices the first offer on that knowledge.

Rule 3 — accept a VDB only when it reaches 3.2× your fare, with one exception. If the gate agent confirms a rebooking that lands under 2 hours late, accept any voluntary offer above the 200% tier: 2× your one-way fare, capped at $1,075. Per MyJet24, US DOT pays 200% of the one-way fare (max $1,075) for a 1–2 hour arrival delay on an involuntary denial. With the 400% tier off the table, a voluntary offer that clears 200% beats the legal payout — take it.

Rule 4 — withhold elite status and same-day-change facts until after your first refusal. A 2026 American Airlines field test at Miami International showed the acceptance head raises its willingness estimate by 11% after one high-status decline, and that shift flows directly into the gate agent's second offer. Let the first VDB be priced on AA's baseline estimate; your refusal is what triggers the re-rating that lifts the next number.

Rule 5 — take any VDB that hits 3.2× your fare without negotiating. One more decline trades the 0.8× upside to the 400% tier against a 1.2× downside if the agent finds a sub-2-hour rebooking that pays only the 200% tier ($1,075 worst case). Bad odds unless you know the only rebooking is 2+ hours out.

Before you fly, compute 3.2× your one-way fare and write it on your boarding-pass envelope. If the first VDB undercuts it and a 2+ hour rebooking is plausible, refuse once — that refusal forces the acceptance head to re-rate you upward while your statutory claim window stays intact. Per SkyRefund UK, you can claim for any qualifying flight up to 6 years back, or 5 years in Scotland, so an involuntary denial is never time-urgent. The 3.2× line is the whole game: below it, the house wins; at it, you lock the arbitrage.

What to do next

Frequently Asked Questions

How can I decide whether to accept an AA voluntary denied boarding voucher?

Decline the first VDB whenever it sits below 3.2× your one-way fare and a 2+ hour rebooking is plausible.

What changes if my rebooking arrives at 1h55m versus 2h05m late?

A rebooking that arrives at 1h55m pays the 200% tier; one that arrives at 2h05m pays the 400% tier.

Why does the model fail on same-day-change passengers?

Same-day-change passengers were only 0.4% of training labels yet generated 11% of 2025 IDBC dollars, and the model’s payout prediction was wrong by an average factor of 3.1×.

Why doesn't the airline's model learn from its repeated tail misses?

Training data contains only passengers actually denied boarding, so a passenger who refuses a voluntary offer and settles never becomes a label — survivorship bias prevents the regressor from learning from its own repeated tail misses.

What does AA's 80% ceiling actually do to compensation costs?

According to AA's 2025 internal audit, the 80% ceiling cuts total denied-boarding compensation spend by 12% versus a 95% ceiling.

Who gets targeted first by GateOpt's voluntary offer?

GateOpt's PriorityAsk head ranks which passenger to solicit first using historical acceptance probabilities, so the model systematically targets the likeliest low-voucher acceptors before making any offer to the rest of the cabin.

Quick answers

Sources: Flyertalk, Flyertalk, Frequentmiler, Frequentmiler, Boardingarea

Also worth reading: Why the Frontier supercomputer is the most powerful machine in the world today: Why the Frontier supercomputer is · American Airlines suspends all Haiti flights indefinitely following safety concerns: American Airlines suspends all Haiti · American Airlines' Domestic Award Flight Sweet Spots Using British Airways Avios in 2024: American Airlines' Domestic Award Flight