Using Synthetic Search Data for Flights Price Forecasting

Andrew Reuben5 Oct 2023

Problem

Expedia Group shows travelers a chart of historical and forecasted flight prices next to the search results. The ideal forecasting dataset has at least one search a day for every combination of route, trip date and passenger count. Traveler searches do not provide it. Many combinations are searched rarely, so their price series have holes, and popular routes have gaps too once every trip date is considered. A model trained on such series sees only part of how prices move from day to day.

Example from the article: for the LAX to JFK route, a set of trip dates built from organic searches shows many missing days of pricing data between the points of the series.

Solution in brief

Run automated searches on a fixed daily schedule with predefined parameters, using the same search path as travelers. This guarantees at least one price observation per day for each chosen combination. The scope stays limited to the most popular routes, because the searches use system capacity and storage.

Questions this case answers
  • How to build a complete daily time series per entity when passively logged data covers only some combinations of entity, date and options
  • How to choose which combinations to sample actively when each probe costs system capacity and storage
  • How to get a gap-free daily price history for forecasting when organic search logs are sparse?
  • How to decide which routes and trip dates to collect price data for under a limited search budget?

Tools and methods

  • Synthetic search data — searches made by an automated process with predefined parameters; the output has the same form as organic search data.
  • Daily price time series — the dataset for forecasting: one price per day for each route, trip date and passenger-count combination.
Continue reading — it's free

Google or a one-time email code. No password.