Machine learning is an area of computer science that focuses on the development of programs that can teach themselves to grow and change when exposed to new data. It’s a field that will bring significant improvements to Revenue Management practices and systems. The concept is probably new to the typical RM user, so let’s explore machine learning via football strategy.
Three Teams, Three Strategies to Win
Let’s consider three teams: the Amarillo Armadillos, the Bangor Badgers, and the Chattanooga Chinchillas. All three teams collect and analyze data from their previous games. That could include the mix of plays they have run in those games, the outcome of the games, and other data points. The three teams use the data to develop a strategy to use in upcoming games. By combining their past strategy and performance, they develop a future strategy that gives them the greatest chance to win. How each team chooses to use the data they collect will ultimately decide which strategy is best.
The Amarillo Armadillos only distinguish between running and passing plays in their data and game plans. They also decide to use a single game plan that they will use for each game over the entire season. For instance, they may uncover the fact that they are most likely to win games when their play mix is comprised of 70 percent passing plays and 30 percent running plays. Despite their limited data set, the Armadillos are still able to determine a strategy that will result in their best possible performance. However, there are advantages to adopting a more sophisticated approach.
The Bangor Badgers analyze their historical data at a much more detailed level. They compile a history of the specific plays they have run from their playbook, as well as the plays run by their opponents. They also prepare game plans by deciding which plays they want to include against the specific opponent they are to face. The Badgers are able to adapt their strategy for each game based on what they know about their adversary. Though it’s more labor intensive, the Bangor Badgers will probably outperform the Amarillo Armadillos given the increased level of precision in their data and decision-making.
The Chattanooga Chinchillas take a similar approach, but they can also adapt their strategy in real-time during the course of a game. In other words, each play that is run and its outcome becomes a new data point that they immediately incorporate into their game plan. They are able to “re-optimize” their strategy for the remainder of the game after each and every play. So each time they make a decision, they know the result will be optimal given the scenario at hand.
The Chattanooga Chinchillas are utilizing a form of machine learning. They are making an optimal decision based on continuously evolving new data. Then they add that new data into their strategy. There are many possible applications of machine learning in Revenue Management, especially with forecasting and optimizing forecast accuracy. A great example is setting the parameters in a demand forecast.
Suppose you are deciding how much to weight two separate forecast approaches (such as a long- and short-horizon forecast) within a blended forecast. Like the Amarillo Armadillos, you can decide upon static weights for each data point, and optimize them for the best possible accuracy over time.
Now suppose you are able to set these parameters at a more granular level. For example, you could use different weights for each business segment, room types, days of week, and so on. Like the Bangor Badgers, you can still find the set of parameters that optimizes forecast accuracy, but the amount of data you’re using has increased exponentially.
What’s Best for Your Business? A Sophisticated, Agile Approach
Ideally your forecast would adapt to each new observation you collect like the Chattanooga Chinchillas do with their in-game strategy. Just one additional day of actualized demand for one segment of your business could render your current parameters sub-optimal. A machine learning approach will optimize your weightings with each new data point you collect. Your forecasting model will adopt a “self-learning” approach towards continually ensuring that the forecast’s outputs are as accurate as possible.
There are many more sophisticated applications of machine learning within Revenue Management. Picture a forecast that goes beyond simply setting parameters that could actually choose from a set of completely different algorithms at any given time. Or perhaps it could know to include or exclude entire datasets from the forecast model all based on the data available at the time. But for now, think about how your favorite football team goes about optimizing its performance, and you’ll have a pretty good understanding of what machine learning is all about.