One of the greatest unknowns which rental car companies currently face, is how to respond to competitor pricing. When renting a car, customers tend to choose among the cheapest options available. This means setting your price above competitors’ prices can be a risky move. But in this price war, it doesn’t have to be race to the bottom.
While price-matching competitors is the right strategy in some situations, there are other scenarios in which the firm has considerable pricing power, giving them the option of pricing above the market. However, distinguishing between these two scenarios can be challenging, given that it’s hard to isolate what the key drivers of competitive price sensitivity are.
The competitor price sensitivity of a rental offer depends on some combination of car attributes (e.g. size, economy/luxury), location attributes (e.g. airport, city/state), seasonal attributes (e.g. weekend/weekday, summer) and customer attributes (e.g. loyalty member). However, this relationship is most likely nonlinear. For example, convertibles might be less price sensitive than other car types in summer months, especially on weekends when there are typically more leisure travelers and in coastal and mountainous areas that have scenic roads.
In a situation like this, random forests are a great tool to use due to their ability to effectively model outcomes that depend on potentially complex nonlinear interactions of variables.
But first, what are random forests?
Random forests are a supervised machine learning technique in the class of ensemble methods. Ensemble methods use multiple machine learning models which, when used together, give better predictive performance than any component model on its own.
Specifically, random forests are ensembles of classification and regression trees (CART). Essentially, you train multiple CART models and blend the predictions of each tree to determine the combined prediction.
Random forests are considered “black box” models because their internal workings can be complex and hard to interpret by humans. However, a random forest is a collection of regression trees, which each have their own logical structure. So, it is possible to draw out more human-interpretable insights from random forests than a typical “black box” algorithm, which can be helpful when eliminating unknowns.
What’s random about them, and why is randomness important?
CART is a deterministic algorithm, therefore, for any given set of input data, the algorithm will always output the same best tree. So, if you ran CART on the same data 30 consecutive times, you’d end up with 30 identical trees.
However, randomness can help produce an ensemble of trees that are each somewhat different. There are two ways in which random forests use randomness:
1. Random sampling of data – rather than using all available input data to train each CART model, each model uses a random sample (say 60-80 percent) of the total data set. This allows each tree to fit to slightly different patterns that appear in the data samples.
2. Random sampling of possible split variables – as the CART algorithm builds a tree, at each node it must choose which variable to split to get to the next level in the tree. In random forests, rather than considering all possible variables to split on, only a random sample of variables are considered at each step.
To understand why randomization is powerful, let’s recall how CART works. It chooses the variable to split in a “greedy” way – by determining which variable will maximize the fit of the next level down. But the variable that maximizes the fit one level down may lead to a sub-optimal result a few more levels down. By randomly limiting the split variable options, and thus disrupting some of the “greediness” of the algorithm, it is possible to build a more accurate tree. Of course, it is also possible for randomness to lead to errors and inaccuracy. That is why a random forest contains many tree models that are blended, leading to greater accuracy than a single CART model.
So, why are random forests useful in rental car pricing?
For rental car pricing, random forests can be used to predict the difference between utilization-adjusted price and competitor price, based on numerous predictor attributes. These predictor attributes can include any available data on car attributes, location attributes, seasonal attributes and customer attributes.
For all future rental car offers, the random forest model will tell you how closely you should mirror competitors’ pricing based on the attributes of that offer. For example, for an economy car weekend rental in Atlanta in August, it might tell you to price at 3 percent below competitor price. This means the model has detected that in comparable situations in the past, prices needed to be slightly lower than competitors to reach typical utilization. In another situation, it might tell you to price at 5 percent above competitor price, implying you have the power to successfully price above the competition without sacrificing sales.
While the random forest model may not tell you exactly why it is recommending certain prices, it yields actionable rules based on historical data. This makes it a powerful supervised machine learning technique that can be used to improve how rental car companies, and others, respond to competitor prices.
Many thanks to Gabrielle Shklovsky, Operations Research, Senior Consultant, for co-authoring this thought leadership piece with me.