Our technology and data

Bing travel's predictive technology evolved from a University of Washington research project led by computer science professor and Internet search expert, Oren Etzioni. We've spent several years developing and refining our state-of-the-art predictive, statistical and data management algorithms with the objective of providing consumers with unprecedented, actionable information to help decide “when is the best time to buy?”

Data collection

Each day Bing travel processes millions of round-trip, priced flight itineraries from several complete and partial airfare information sources. These itineraries extend up to a 180-day period, encompassing trip lengths up to 21 nights and span over 2500 combinations of U.S. origins to domestic and international destinations. For each origin and destination pair, this is equivalent to performing thousands of searches on a typical travel search site. The data is stored and forms the basis of our airfare price predictions, deals, and flexible travel features.

Data aggregation and analysis

The huge volume of airfare and hotel rate data we process every day is aggregated and transformed using a variety of statistical measures. These statistical measures allow us to intelligently filter air and hotel data to reveal where the best prices and deals reside. For our airfare price predictions, we also calculate special features from our historical data that are predictive in nature. This data is used to train our predictive models which ultimately power our arrows and recommendations on our web site.

Visualizations

Given the multi-dimensional nature of the raw data we process, as well as the derived data sets we create, we rely heavily on data visualizations to help us understand airfare and rate trends and patterns. Being able to visualize time-series information in many dimensions can be a challenge with traditional visualization schemes. Oftentimes creative and unconventional forms of data images have provided the most potent snapshot of phenomena we are interested in researching. Further, data animations have been a useful tool for studying variables that change over time such as price.

Modeling and prediction

Our Research and Development team is comprised of experts in the fields of statistics, data mining and machine learning. We have developed algorithms that can identify patterns and conditions from our history of accumulated airfare data associated with significant price changes. These patterns, learned from our historical data, are represented and stored in predictive models. Once trained, we use these models to predict the future of airfares in conjunction with current market conditions.

Measuring Price Predictor performance - simulation technology

In order to test our predictive accuracy we have developed computer programs that simulate passengers buying airfares based upon prices from real data we process and store every day (sometimes referred to as “back testing”). We build sequences of models for each day we have data and simulate a distribution of passengers with different economic constraints purchasing tickets, similar to those found in the real world. Each simulated passenger receives a recommendation for the day and market they are shopping from the predictive models. We tally up the results of the outcomes of recommendations made to thousands of simulated passengers. This data tells us how accurate we are and how much money was saved or lost by our simulated customers. Although the simulations do not mimic real life exactly, they provide a close enough approximation for us to understand the efficacy of our technology.

Technology platforms

Since we have so much data, we use a lot of very fast, very large storage solutions. We also crunch a lot of numbers creating statistical measures and predictive models. As such, we use grid technology to cluster racks of dual-processor, 64-bit commodity hardware boxes to make our daily amount of massive calculations more manageable.

Patent protection

Bing travel's prediction technology and its proprietary features, functionality, and business processes are patented (U.S. Patent No. 7,010,494) or have a patent pending in both the U.S. and multiple international markets.