AI-dragon-zakura challenge 1: US Flight Delay Prediction
In the United States, one in five domestic flights arrives more than 15 minutes late. Delays do not occur in isolation; a delay in an early morning flight can propagate to subsequent flights operated by the same aircraft, and adverse weather conditions at one airport can ripple through national routes passing through that hub.
For airlines, knowing which flights are likely to be delayed before departure has significant practical value. It allows for scheduling buffer time, preparing alternative aircraft, or proactively informing connecting passengers. However, these measures must be taken before departure to be effective.
Goal: Using only information confirmed before a flight departs, predict whether the arrival will be delayed by more than 15 minutes.
Why "15 Minutes"? This threshold corresponds exactly to the definition of on-time performance established by U.S. federal regulation 14 CFR Part 234: "On-time means a flight that arrives less than 15 minutes after its published arrival time." The Department of Transportation calculates each airline's on-time performance rate monthly using this standard, based on gate arrival times (excluding cancelled and diverted flights). Thus, this competition is a task to predict the industry's official metric directly.
Time-Series Split: The dataset is split chronologically: - Train: July 2022 to June 2023 - Test: July 2023 to October 2023
Regarding Cancelled Flights in Test: The test set includes schedules for cancelled and diverted flights. Since regulations exclude these from on-time performance calculations, they are also excluded from scoring. However, it is impossible to distinguish which rows are subject to scoring. Therefore, predictions must be submitted for all rows.
Prohibited Actions: The data source is the BTS (Bureau of Transportation Statistics). It is technically possible to obtain actual performance data for the test period (July–October 2023) from external sources. Obtaining the ground truth by referencing original data is strictly prohibited. Similarly, using external datasets containing delay records for U.S. domestic flights during this period is forbidden. If you wish to use external data (such as geographic information), you must consult the discussion forum first to ensure fair play and verify licensing.
Evaluation: Submissions are evaluated based on the Area Under the Receiver Operating Characteristic Curve (ROC AUC) between the predicted probabilities and actual labels.
Submission File:
Submit a file predicting the probability that a flight will arrive more than 15 minutes late for every row in test.csv. The format must include headers:
flight_id,arr_del15
0,0.184
1,0.372
...
The arr_del15 column must contain probability values between 0 and 1. Values exceeding 1, negative values, or missing values will result in scoring errors.