M

any airline revenue management teams take advantage of heuristics in the form of optimization profiles to efficiently boost RASK performance. To make this more manageable, they associate flights into groupings that share similar commercial dynamics.

Management efficiency is enhanced because analysts can more often apply a common set of business rules when optimizing multiple routes and departure dates. The key to this working well is to use a robust classification method that minimizes behavioral dissimilarities within groupings.

When the world runs smoothly and historical patterns remain consistent, the homogeneity of flight groupings can often stay solid for many months or even years. But what happens when something like a pandemic interrupts those prior patterns?  

As we have seen in countless client airline situations the past few months, old patterns are giving way to new. Booking curves are compressed. Willingness to pay has decreased. Seasonality effects have tapered. Price elasticities have fallen. Business travel has disappeared. And competitive aggression has shifted gears.

Revisit Your Flight Groupings

In aggregate, it should surprise no one that an airline’s 2019 schema used for grouping flights into optimization profiles will start to fail as the “new normal” unfolds in 2020 and beyond. For that reason, Kambr is reminding its clients to revisit their flight groupings and to redefine the associations that align best under today’s environment.

It sounds like the RM version of mundane housekeeping, but sharpening these arrangements can significantly improve the impact of their heuristics. We estimate a 1.5%-2.5% revenue loss can be reclaimed when homogeneity is restored.

"Whatever approach best fits your organization, at its core the objective should be consistently straight forward… minimize the commercial behavior dissimilarities within flight groupings."

So, how should your RM team determine its new and improved flight groupings and optimization profiles?  

Like many business problems, there are multiple pathways to an answer. Some carriers are well served by using a simple k-means cluster analysis. Others may invest in deep neural networks to obtain the best classifications. Or myriad other methods for clustering may be tried.  

Whatever approach best fits your organization, at its core the objective should be consistently straight forward… minimize the commercial behavior dissimilarities within flight groupings.

Booking Curve Comparison

Usually, booking curve shapes are a key characteristic to examine during the process.  As Figure 1 illustrates, two curves can often start with nearly identical build rates.

Figure 1: Comparing two booking curves

And they might even finish at the same load factor. But during intermediate stages of the curve, say 90-30 NDO (number of days out), some degree of divergence may occur.  

And those distinct build rates may reflect different sets of competitors, inventory strategies, or consumer propensities.

So how similar are they? Regardless of the underlying cause, a means for quantifying dissimilarity is needed.

Quantifying Dissimilarity Between Flights

Whether it is gross bookings, yield, unit revenue or some other attribute descriptive of a flight’s behavior, an array of metrics that score dissimilarity are computed.

In the example shown in Figure 2, we consider the simple dimension of booked load factor by NDO.  Once dissimilarity has been quantified, criteria for clustering can be evaluated by the algorithm of your choice.

Figure 2: Quantifying dissimilarity between flights

With a robust methodology, even networks with seemingly disparate and wide-ranging behavioral dynamics will eventually reveal a set of commonalities that define logical clusters.

Flight 3: Widespread behavioral disparity in a network is not uncommon

More Clusters Does Not Equal More Gains

There is a natural temptation to choose a large number of clusters. After all, the more groupings one allows, the easier it is to establish homogeneity. But with more clusters comes more management complexity which is a real factor to keep in mind.  

And, while the dissimilarity continues to fall as more groups are allowed, there is a diminishing marginal gain here.

Figure 4: Optimal number of clusters

A practical balance can be had without having dozens of primary clusters. Indeed, Kambr has seen networks with high order dissimilarities converge around 7-9 primary clusters as a pragmatic set after algorithmic tuning and data anomaly detection are applied.

Ideally, a schema with 4-6 primary clusters will emerge and then sub-clusters within those can be delineated. But there is no magic number in this regard.  Each network tends to classify and cluster in ways that are uniquely reflective of its own environment.

Figure 5: Primary clusters will ideally converge around 4-6 topologies

Describe Clusters in Terms of Feature Importance

A final but important step is to describe the emerging clusters in terms of feature importance. This is helpful in a few ways. First, it becomes more meaningful to humans when sub-clusters represent sets they can conceptualize.

Figure 6: Primary clusters may be sub-clustered by sets of distinguishing attributes

So, rather than referencing something esoteric like “Cluster 9/Sub-Cluster d”, the analyst can, for example, think instead about low-season, peak-day VFR markets facing ULCC competition.

Such profiling also lends itself to the task of categorizing a new route where prior history is likely absent and would otherwise hamper cluster analysis. By evaluating attribute descriptors for their importance within sub-clusters, the introduction of a new route can be supported by a meaningful classification.

There are many classifier systems that may be considered including regression, k-nearest neighbors, random forests and others. The most appropriate method for your network depends upon the diversity of market characteristics and the resources supporting your RM team.

"By evaluating attribute descriptors for their importance within sub-clusters, the introduction of a new route can be supported by a meaningful classification."

In closing, cluster analysis and classification schemas are not the sexiest parts of revenue management. It just isn’t the kind of job content that hiring managers advertise when recruiting MBAs into their RM groups. But they are vital parts of optimizing RM performance. Necessary housekeeping to say the least.  

And the vagaries of this current pandemic environment will only exacerbate the value of doing them well and timely. The main thing to remember is that housekeeping of this nature is critical to effective RM outcomes.

Those who accomplish this will find their optimization profiles and business rules doing the best job possible under the circumstances. Happy housekeeping!