Airfare prediction algorithms go haywire
But the difficult years of the pandemic made it all more complicated. Oren Etzioni is now CEO of the Allen Institute for AI, but in the early 2000s he built – and sold to Microsoft – one of first airfare forecasting tools. Prediction algorithms are pretty good at re-weighting the importance of different factors as the world changes, and, he says, “they have a chance to adjust automatically by having the most recent data available.” But that can take time, according to Etzioni: days or even weeks.
Google Flights helps customers find the cheapest tickets for their preferred routes and dates. But since spring 2020, the search engine has drastically reduced the amount of “predictive information” – forecasts of when prices are likely to rise or fall – that it offers searchers. In general, Flights aims for 90% prediction accuracy, says Eric Zimmerman, director of travel products at Google. “With the increased volatility in airfares, it has become more difficult to achieve that high level of confidence,” he says. The pandemic and its effects on air travel also prompted the company to halt an experiment launched in the summer of 2019, in which it guaranteed fares for specific routes and sent refunds to travelers if the price dropped before takeoff. . That could bring the project back soon, Zimmerman says, as the industry begins to stabilize.
Giorgos Zacharia, president of online travel agency and search engine Kayak, says he has a team of MIT PhDs who spend their working lives tending to the website’s price prediction tool . While the prediction algorithm, first launched in 2013, typically needs to be adjusted every few years, he says, the latter two have seen “serious retraining” every few months, and sometimes every few weeks. . He says the accuracy of prediction tools, which is typically around 85%, may have periodically dropped over the past few years, perhaps closer to 83%. This means that at certain low times, waiting or buying when the website told you to was less likely to have led to the lowest possible price – and could have led, instead, to a slight shake of the fist toward the sky.
“Machine learning likes to learn old and past repeatable patterns, and make predictions based on the likelihood that those patterns will work again,” says Zacharia. “So the pandemic, which brings many unexpected outliers, also affects the input data of models like this and makes it a more challenging environment.”
Hayley Berg, chief economist at Hopper, says the company’s predictive tool is trained on 75 trillion routes and eight years of historical price data. But today, the algorithm weighs more heavily on what it has seen over the past three years, which has helped the tool maintain 95% accuracy throughout the pandemic, according to the company. Even in the early days of the Covid-related shutdowns, she says, Hopper got his plane ticket price predictions correct 90% of the time. Still, customers shouldn’t be shocked by the price volatility – Hopper found that the average domestic flight changes price 17 times in two days, and 12 times if it’s an international flight.
All of these changes lead to many conspiracy theories among ticket buyers, even those who don’t care about price prediction platforms. No, the executives say, airlines don’t track cookies or raise prices if they see you’re interested in a certain route. (Zacharia, president of Kayak, says rates are sometimes higher or lower depending on your location when you search because the systems take “point of sale” into account.) No, there isn’t. no reason for flights to be cheaper. a Tuesday than any other day, a persistent rumor among bargain hunters. “The best time to book will depend on your trip, particularly origin, destination, departure and return,” says Berg. “And that can be very different depending on where you go.”
Today, however, it is not always necessary to use a sophisticated machine learning algorithm to choose the best time to buy – there is no such thing as a good time. The prices are so high, says Kayak spokeswoman Victoria Hart, that there aren’t “many ‘waiting’ indicators these days.”