Lessons in pricing with precision from United Airlines

One innovation that is being discussed in airline revenue management is more granular pricing and inventory management. Tom Bacon finds that this can deliver positive results

Current RM systems often limit the number of ‘inventory buckets’; by this I mean ranges of fare values treated as a single demand grouping. If an airline is limited to 26 buckets – with some set aside for ‘group’ or ‘government’ or ‘fare sale’ or other special categories, they tend to define narrow ranges on the low end ($59 - $74; $75 - $89; $99 - $114, for example), where most of the demand exists. At the higher end, where demand tends to be much lower, the ranges are bigger ($200 - $249; $250-$324). When the ranges get large, this inhibits fare changes: the airline may be reluctant to move abruptly from $200 fares to $250. Presumably then, more precision, more buckets can be exploited for higher total revenue. Let’s get more precise! In fact, at the extreme, let’s allow the system to move fares in $1 increments!

But wait a minute!

United Airlines, however, recently announced a problem with their sophisticated revenue management system that should give all airlines pause when considering more precision. The airline stated that they could potentially improve their revenue performance by $1 billion by adopting less forecast precision, effectively by forecasting at more aggregate levels.

United found that by forecasting at a very granular level they had inadvertently forecast at an unforecastable level. The granularity came with extremely high variance – one example given was a standard deviation at 11 times the mean! United is now hurrying to re-forecast at a more aggregate level, reverting to granularity only where it makes sense, or where the standard deviation implies a more robust solution. Many airlines have come to the same conclusion: forecasting at too granular a level is counter-productive.

Many airlines have come to the same conclusion: forecasting at too granular a level is counter-productive.

For example, the variance in demand for $100 tickets versus $120 tickets can overwhelm the $20 differential. Demand is interdependent – potentially demand is forecast more aggregately for the two fares together. RM systems factor in the uncertainty in demand in any forecast and, generally, opt for more inventory at the lower level rather than ‘risk’ inventory for highly uncertain, even if more valuable, passengers. A more reliable forecast for a wider fare range, say $100 – $139, could translate into a more robust RM solution; the system could better allocate inventory between each of sub $100, $100-$139, and over $139 tickets.  Refining demand to $1 or $5 increments will often translate into buckets of only one or two passengers on average with variances from zero to ten – not a good basis for inventory controls.

One airline I worked with ended up developing more robust demand forecasts at three distinct levels, rather than independently forecasting each of the 20+ levels they had previously had. They still maintained their fare classes but the detailed forecasts were formed by disaggregating the three more aggregate fare ranges. This revised approach improved forecast results dramatically.


Thus, developing the optimal fare ranges for bucketing potentially has multiple steps. First, one might use the standard routines recommended by some RM suppliers. These routines typically analyse the distribution of demand across all fare levels and identify clusters of demand among all of the fare level demand data points. The objective is to divide up demand into 20 or so such clusters where:

  • All the fares within a cluster are within a narrow range; the distance among such fares is minimised

  • The differences between the clusters is maximised      

Even if this tool is used, however, the clusters may not lend themselves to independent demand forecasting. Thus, there needs to be another process for grouping multiple fare ranges for forecast purposes. Specifically, this second step tests clusters for cross-correlation. ‘Low fare’ modules, also offered by some RM suppliers, similarly attempt to segment demand within a defined ‘bucket’ between ‘price-able’ (demand associated with the price point) and ‘yieldable’ demand (demand that would pay more if this fare weren’t available). These modules explicitly focus on the inter-dependence of demand across ‘buckets’.

United’s announcement is in line with experience at other airlines – forecasting demand at too granular a level, assuming independence, is counter-productive. Although conceptually, more detail – tapping into bigger and bigger data – can drive improved business results, it must be pursued with caution. Forecasting at more aggregate levels can actually result in more robust forecasts, giving the RM systems a better chance to drive maximum revenue results.

Tom Bacon has been in the business for 25 years as an airline veteran and now industry consultant in revenue optimisation. He leads audit teams for airline commercial activities including revenue management, scheduling and fleet planning.

Questions? Email Tom at or visit his website

Related Reads

comments powered by Disqus