Machine learning can improve the forecast, but only the fares you file can be sold

Most claims about machine learning in revenue management are about the forecast: how many travellers will want each price on each flight, and when. A better forecast is worth having. It still works inside the fare structure, the booking classes and the channels you have, and its value shows only when it is measured against flights that did not change.

Machine learning helps most where the forecast needs more inputs

A classic forecast uses the airline's own booking history by flight, class and days before departure. Machine learning methods can take in more: searches on the airline's website, rival fares, events and schedule changes. Which of these help depends on the airline's own data, so a review tests them on its own routes before anyone pays for them.

Some research replaces the separate forecast

Bondoux, Nguyen, Fiig and Acuna-Agost, from Amadeus, describe a revenue management system based on reinforcement learning that does not need a separate demand forecaster: it learns which booking controls earn the most from the results of its own decisions. It is published research. An airline considering a system of this kind should ask how it was tested, on what kind of network, and how analysts can see why it chose a control.

A forecast cannot open a price that is not filed

Whatever the method, the forecast and the optimiser decide which of the filed prices are open. GDS travel agents see the fares you file and the classes you open. If the ladder has six price points, a better forecast chooses better among six. This is why our reviews start with the fare structure and then the forecast. That order is our judgement.

Five questions to ask about a machine learning forecast

• Which inputs does it use beyond your own bookings, and which of them does your airline have?
• How does it treat bookings in classes that were closed, and travellers who buy the cheapest open fare?
• Which forecast error is reported, and is it measured against the classes that were open at the time? Fiig, Weatherford and Wittman show why that matters.
• Can an analyst see why a recommendation changed from yesterday?
• Will the result be measured against flights that did not change before the contract is renewed? Seven questions show how.