The test plan, with every change made to it before the opening, is on a separate page. In this test, a missed connection means either flight was cancelled, or the real gap after delays was shorter than our estimated transfer time. ANAC's records hold no passenger data, so this is an estimate and not a count of real passengers.
Summary
| What was graded | Cases tested | Result | On the card |
|---|---|---|---|
| Domestic connections (São Paulo Guarulhos, Campinas, Brasília, Belo Horizonte Confins, São Paulo Congonhas) | 2,025,340 connection-days | Pass (all five bars; bar 3 used Brazil's own figures, and against our US figures grades C, D and F would have failed it) | Grade, with Brazil's own per-grade figures |
| Domestic flights with one end at those hubs | 386,917 flight-days | Pass (bars 1, 4 and 5) | Grade, no percentage (none is published for flights) |
| International flights (Guarulhos, Campinas) | 50,020 flight-days | Fail: grades out of order (B worse than C) | Grade marked Provisional, no percentage |
| International arrival connecting to a domestic flight (Guarulhos, Campinas) | 144,375 connection-days | Fail: predictions 15.3 points too high. Also not a clean test | Grade marked Provisional, no percentage |
Under the owner's decision of 25 October, the result decides the label and what we publish. It does not decide whether grades launch.
Domestic connections: pass
Missed-connection rate by grade, whole test period (September 2025 to May 2026):
| Grade | Missed-connection rate | 95% interval | Model predicted | Connection-days |
|---|---|---|---|---|
| A | 5.1% | 4.3–6.1% | 3.0% | 864,533 |
| B | 7.8% | 6.9–8.9% | 7.0% | 378,973 |
| C | 8.8% | 8.0–9.7% | 8.9% | 304,868 |
| D | 9.0% | 8.1–10.0% | 9.5% | 246,469 |
| F | 16.8% | 15.7–18.2% | 20.7% | 230,497 |
- Order: no grade out of order.
- Calibration: predicted and realised rates differ by 1.6 points on average (the limit is 3).
- Brazil's own figures: on one half of the test weeks the rates were A 5.2%, B 8.0%, C 9.1%, D 9.2% and F 17.1%. The other half came in within 0.6 points at every grade (the limit is 5). Cards quote these, rounded: A 5%, B 8%, C 9%, D 9%, F 17%. They are lower than our US figures: in the test, Brazilian connections at each grade ended in a missed connection less often.
- Ranking: the correlation between predicted and realised outcome is positive (0.19 per connection-day, interval 0.18–0.20). Per connection with 15 or more days it is 0.63, against 0.96 in the US.
- Coverage: every hub has far more than 2,000 graded connection-days.
- C and D are about the same. A C and a D connection in Brazil had about the same chance of a missed connection (D minus C: +0.2 points, interval −0.1 to +0.4). A/B, B/C and D/F do differ. The grades still put connections in the right order. Why C and D don't separate is not yet known.
- Many connections have no grade yet: from 33% (Campinas) to 57% (Congonhas) of domestic connection-days. We think the main cause is flights without enough history in the build months, because schedules changed after August 2025, but this is not yet confirmed. They show "not graded yet".
Connections of 40 to 50 minutes, by hub
Released because domestic connections passed. Missed-connection rate for graded connections with a scheduled gap of 40 to 50 minutes, September 2025 to May 2026. Graded connections are 46% to 66% of each hub's connection-days in this band. Our US backtest of 23 September used one fixed transfer time for every connection; this table uses the model's estimate for each pair, so the two methods differ in that respect.
| Hub | Missed-connection rate | 95% interval | Missed / graded connection-days |
|---|---|---|---|
| Brasília (BSB) | 9.3% | 8.1–10.6% | 3,415 / 36,593 |
| São Paulo Congonhas (CGH) | 11.5% | 10.2–13.4% | 5,530 / 47,929 |
| São Paulo Guarulhos (GRU) | 12.9% | 11.5–14.5% | 10,158 / 78,574 |
| Campinas (VCP) | 13.1% | 11.7–14.6% | 6,729 / 51,557 |
| Belo Horizonte Confins (CNF) | 13.5% | 12.3–14.8% | 4,611 / 34,139 |
Reported after the opening, outside the plan: Brasília was lower than each of the other four hubs, by 2.2 to 4.2 points.
Domestic flights: pass, with weak ranking
"Missed" here means the flight arrived 60 minutes or more late, or was cancelled.
| Grade | A | B | C | D | F |
|---|---|---|---|---|---|
| Rate | 4.4% | 5.5% | 5.6% | 6.3% | 10.1% |
- The rates rise from A to F, and ranking is positive but weak (0.056 per flight-day; 0.46 per flight with 15 or more days).
- B and C, and C and D, are about the same.
- The model over-predicts at F: 19.0% predicted against 10.1% realised.
International flights: fail
| Grade | A | B | C | D | F |
|---|---|---|---|---|---|
| Rate | 6.5% | 10.6% | 7.9% | 12.1% | 13.3% |
- B did worse than C, and their intervals don't overlap, so the order bar fails. Ranking is still positive (0.10 per flight-day; 0.57 per flight with 15 or more days).
- Campinas has only 2,499 flight-days, a thin sample.
- The model is too optimistic at A (2.9% predicted, 6.5% realised) and too pessimistic at F (26.5% predicted, 13.3% realised).
International arrival connecting to a domestic flight: fail, and not a clean test
| Grade | A (39 cases) | B | C | D | F |
|---|---|---|---|---|---|
| Rate | 2.6% | 8.1% | 14.9% | 26.6% | 61.3% |
| Predicted | 13.7% | 23.6% | 38.7% | 50.6% | 74.4% |
- Calibration fails by 15.3 points (the limit is 3). The order and ranking bars hold, but they prove little here, because the grade and the "missed" outcome rest on the same estimated arrival time.
- Why we think the predictions run high: the predictions draw on the full spread of the estimated arrival time, whose slow end reaches 256 minutes at Guarulhos. A connection was judged missed against the typical time (124 minutes). The slow end pushes the predicted rates far above the realised ones.
- Not a clean test: two inputs broke the plan's dated-sources rule. One was a customs time resting on an undated account. The other was a terminal-move time with no dated archive copy recorded when the model was committed. Dated sources for both were found the same day and can be used in the next test.
- Ungraded share: 53% at Guarulhos and 43% at Campinas.
- Since the test (owner's decision): the launched model uses revised, measured arrival steps. That model was not tested here, and its grades are marked Provisional.
Checks on the test itself
- A separate count of test cases matched the prediction file exactly: 4,528,088.
- No actual times were missing at any hub. ANAC records each flight as either flown or cancelled.
- A second, separate reading of the raw records for Guarulhos in January 2026, chosen before any outcome was seen, matched exactly (24,487 flights).
- A hand check of 100 cases, worked out by hand from the printed times, gave 0 disagreements.
Disclosures
- Brazil's own figures were chosen as the bar-3 basis after a build month, August 2025, had failed against the US figures. No test-month result had been seen (owner's decision).
- The prediction script changed twice during the run, once to write in chunks and once to resume after a memory stall. Each was shown beforehand to give a byte-identical file on a build month.
- One row of the sealed test table was printed early. It was used nowhere.
- Grades ship whatever this result says, and a failed test is labelled Provisional (owner's decision).
- International-to-domestic connections were reinstated after a build-month calibration miss, and this was not a clean test.
- The 15-minute boarding cut-off and several arrival steps are modelled judgements with no dated source.
What we are fixing next
- International arrival to domestic: measured arrival steps, the wait for bags overlapping passport control and never added on top of it, the outcome judged and predicted on the same basis, and a new sealed test on fresh months.
- Domestic connections: find out why C and D don't separate, and reduce the share that has no grade. We'll measure first and assume no cause.
- International flights: the B/C inversion and calibration at both ends, once there is more data (Campinas is thin).
- Domestic flights: over-prediction at F.
Flight records: Agência Nacional de Aviação Civil (ANAC), Voo Regular Ativo (VRA), January 2024 to May 2026, downloaded 23 October 2026, published under a Creative Commons Attribution licence on dados.gov.br. ANAC had no part in GateCast and does not endorse it; grades and figures are Wingmark Analytics' own estimates.
Dados de voos: Agência Nacional de Aviação Civil (ANAC), Voo Regular Ativo (VRA), janeiro de 2024 a maio de 2026, obtidos em 23 de outubro de 2026, licença Creative Commons Atribuição (dados.gov.br). A ANAC não participa do GateCast nem o endossa; as notas e os números são estimativas da Wingmark Analytics.
Checksums (SHA-256): test flight table 5fd945a2…, sealed predictions 76faaf3f…846d, outcomes 3e00a73d….
Results opened 25 October 2026. Public version prepared 2 November 2026. Grades are estimates.