The cost of flight delays—whether inflicted by Mother Nature, man-made bottlenecks, or congressional budget cuts—add up to a more than $31 billion annual hit on the U.S. economy. Even minor delays are costly. Flights that miss their scheduled gate arrival time by just a few minutes rack up millions of dollars in annual cost overruns for airlines and airport operators. Ahem, Newark Liberty International Airport.
That’s why Alaska Airlines (ALK), which boasts an admirable on-time track record (PDF), and General Electric (GE) took the conundrum of how to improve on-time arrivals out of the clouds and into the crowd, calling on Big Data analysts from around the world to devise an industry-beating algorithm for more accurately predicting when U.S. domestic flights can be expected to arrive at their destination.
The winner of the contest, a team of analysts at Singapore’s Institute for Infocomm Research (I2R), collected $100,000 for their data-crunching work, which predicts flight arrival times far more precisely than current industry benchmarks. According to Xavier Conort, a specialist in machine learning and a member of the I2R team, when airlines make an error in calculating the gate arrival time—what’s referred to as “average error time” in the industry—the flight is off schedule by seven minutes. It should be noted that a majority of flights do arrive at their gates on time, Conort says, but knowing exactly when is key to avoiding a domino effect of bottlenecks on the ground that quickly spread to the air.
The I2R team’s algorithm was able to trim an average of three minutes off the industry’s “average error time” benchmark. It may still not be down-to-the-second accurate, but trimming off even one minute saves plenty. Every minute in flight-reduction time could save, over the course of a year, more than $6 million in annual crew costs and fuel consumption.
Making a pinpoint-accurate prediction on gate arrival times is notoriously tricky, as a plethora of factors alter flight times. Weather and wind are the most common factors, but there are also seemingly avoidable ground issues, like that knucklehead passenger who neglects to board his flight on time, causing his bags to get pulled from the plane’s hull. The data crunchers had to take this all into account—weather reports along the route, airport arrival statistics, even airplane models and gate numbers were considered—to build their predictive model based on more than 20,000 U.S. daily domestic flights traveling across the contiguous 48 states from November to February.
“It’s not bad,” Hon Nian Chua, who was part of the five-person I2R data team, says of the group’s modeling handiwork. “We can still do better. Still, I don’t know if you could get it all the way down to zero.” The team believes its model will help pilots and air traffic controllers to one day use a more rigorous approach that combines historical departure/arrival data with current weather and traffic snarls to determine the quickest and most fuel-efficient flight path. That would be one way to trim a few more minutes off the journey, he says.