The airline industry was at the forefront of many of the significant innovations of the last century. Besides for all the advances related directly to aviation, airlines were also pioneering in developing computer systems that could be accessed around the globe to book and reserve airline tickets. Airlines also established revenue management systems to optimize revenue from ticket sales. Inventions by the airlines in this area have informed revenue management models for multiple other industries as well, most notably for hotels and car rental companies.
Yet, in my view the current state of the art as it relates to pricing and revenue management in the airline industry can best be described as an evolution rather than a revolution. Here is what I mean. In the not too distant past when access to massive amounts of data and compute was prohibitively expensive, scientists who wanted to optimize a process would hypothesize about the mathematical makeup of the problem. Based on that postulation they would build a mathematical model that represented the problem to be solved. They then validated the mathematical model by checking if it properly explained the underlying real world (or — in absence of real world data — simulated) data.
In the present world where we have easy access to massive amounts of compute and big data there is an alternative approach to this process. When we want to predict a certain outcome, instead of hypothesizing what a mathematical representation of the problem would be, we hypothesize what the underlying causes of the predicted outcome is. We then use data to represent those causes. This is done using data transformations, modeling variables and by joining multiple dataset together. Data scientists call this process Feature Engineering. We then choose machine learning algorithms that can build us a solid predictive model. We split the data into at least three parts, one to train the model, one to validate the model and one to test it. The optimal result is a model that can predict an event or a state on new data that was not used in the model building process.
Key to the difference between these approaches is that in the latter there is no need to for the human to actually build the mathematical model rather the algorithm builds it by finding patterns in the massive amounts of data it trains on. Another key difference is that while the former starts with a hypothesis with regards the underlying distribution of the problem from a mathematical perspective the latter starts with a hypothesis of what features cause the predicted outcome and what algorithm might build the optimal model — finding the underlying distribution and patterns is the work that the machine learning algorithm figures out.
It must be pointed out that there are also many similarities to these two approaches, for example both approaches will use classical Exploratory Data Analysis and data visualization techniques and many of the underlying statistical and scientific methods are similar if not identical. In addition, often data scientists will use mathematical modeling in the feature engineering stage of the model building process.
In my perusal of the airline revenue management literature, however, it is clear that the main focus of finding solutions to airline revenue problems such as dynamic pricing and bundling uses the former older approach. While companies such as Uber, Lyft and Aibnb are now leapfrogging the airline industry in terms of innovation in this area because they are using the more contemporary Machine Learning and AI approach.
This is understandable given that the practical use of Machine Learning is relatively new and airline revenue management professionals mostly were doing their work when the older methods were all we had. And people are often slow in catching up to the fast-changing technological landscape. Furthermore, the airline industry is saddled with formats and streamlined processes that are hard to change quickly. Despite all of this, if current airline players are to thrive they must outgrow the past and move into the machine learning and AI era.