Who pays when a connection is missed?
US airlines already rebook a missed connection, and pay the hotel when the delay is their fault. This page prices the gap they leave: the hotel night after weather or air traffic. The older all-purpose cover is below for comparison. Every number comes from public data and is checked in public.
Read next: the one-page memo · the scoreboard of pre-registered predictions · how it was tested
The gap cover: the hotel night the airline won't pay for
US airlines already rebook a missed connection for free, and pay the hotel when an overnight disruption is their fault. The night after a weather or air-traffic disruption is the gap nobody covers. This cover pays the official government hotel-and-meals rate for the hub and month, only when the connection is missed, the next flight is the next day, and the delay is not coded to the airline. (Model v3, connections paired the way tickets are sold. Pre-registered for September to December 2026, results not in yet.)
Only connections where a miss means a night away: the last flights of the day. All airlines at the hub. Minimums differ by airline, airport and terminal; the model was checked for 35 to 60 minutes.
The older cover: pay on every missed connection (v1)
Priced per connection from how often that kind of connection fails, and not offered where the price is too high a share of the fare. It duplicates what the airline already pays in many cases, which is why the gap cover above replaced it; it is kept for comparison and because its 2026 test is the project's first clean result.
Connection time is in 15-minute steps, reported as the start of the step.
Why not one price for everyone?
| Pricing | Loss ratio, buyers ignore risk | Loss ratio, buyers favour risky connections* |
|---|
*A loss ratio of 65% is the target; above it the product loses money. Who really buys protection cannot be known from public data, so the project shows a range of buyer behaviour. "Favour risky" is one point on it (selection strength 1 on the scale used in the validation). The flat price is the average claim cost in the training period divided by the target.
What this does not tell you
- Connections are synthetic. They are every same-airline pairing of an arriving and a departing flight at five US hubs; the data does not say which passengers actually connected.
- In the v1 calculator, the costs, the 65% target and the 10% cap are assumptions ($ to rebook, $ more if the next flight is the next day). The results barely depend on the dollar amounts, but the prices do.
- Demand is untested. On one ticket an airline usually rebooks a missed connection for free, so the offer has to cover something extra. How many people would buy, and at what price, needs an experiment (see the product brief).
- No seasonal price. One year of history could not separate seasons from one-off events, so the model has no month term and these prices are the same all year.
- Five airports, same-airline connections, US domestic, data to August 2026.
- The calculator at the top shows v1 prices. A later audit found airlines already pay for much of what v1 covers, which led to the gap cover with official hotel rates and a sourced 5% cap. Its probabilities have no month term, so only the payout moves with the month; whole months have run 20 to 40% above or below the price.
Eight years of months: the weather moves the cover together, and by season
Each dot is one month of realized over expected claims for the gap cover, priced from earlier months only (70 months, 2019 to August 2026, the pandemic months left out). Above the dashed line the cover paid out more than the price expected. June to August (red) ran above it in every summer observed, in the range 1.11 to 1.61. The grey line is the calendar multiplier pre-registered for 2027.
A multiplier is the claims the calendar month predicts against the model's expected claims. It was found on these months, so it is a hypothesis that March to August 2027 will test. The calendar pattern cuts the error of a monthly forecast from 0.206 to 0.125 in log units when each year is predicted from the others.
| Expense share | All connections | Sold within the cap | Sold, weekly stop-loss at 100% |
|---|---|---|---|
| 10% | 0.73 | 0.75 | 0.74 |
| 25% | 0.87 | 0.89 | 0.89 |
| 35% | 1.00 | 1.03 | 1.03 |
At the share the 65% target leaves for expenses (35%) the price needed is about today's. The annual margin for weather is small, 0.3% to 2.5% of premium. The weeks are the risk: 51 of 306 weeks as sold ran above a 100% loss ratio, and a weekly stop-loss would cost 5.1% of premium. Development evidence from eight years of weeks, which cannot show a 1-in-100 year.