Dynamic pricing has gotten extremely sophisticated – and complicated. For airlines, dynamic pricing relies on a series of algorithms to determine the optimal number of seats to be allocated to different fares or fare classes. The algorithms are generally not completely transparent, and there are literally millions of calculations that are updated nightly to insure allocations remain up-to-date.
Dynamic pricing, of course, has been adopted in a multitude of industries besides that of airlines, although the version for airlines may be much more mature. Due to the complexity of models, including, in some cases the implementation of artificial intelligence, insuring systems aren’t simply a “black box” becomes increasingly challenging.
There are three broad steps in the process of dynamic pricing: construction of a database for modeling customer behavior, forecasting customer behavior or changes in demand (often in highly granular market segments), and optimization (offering the right price at the right time). There are often many choices, assumptions, and calculations at each step, making it difficult to follow directly from input-to-output. Let’s take a closer look:
Data Collection and Categorization
Historic data needs to be assembled in a useful way for modeling. What data is collected, and how it’s gathered, directly impacts the dynamic pricing outputs.
- What Data is Tracked?: To support dynamic pricing, airlines track net flight bookings – gross bookings less cancellations for each flight by O&D by fare level by day. Ancillary purchases have not been tracked in the same detail but is of increasing importance. The booking database often does not identify highly localized events, and forecasting demand for such events is often left for analysts to handle outside the system.
- Categorization/Classification: Definition of various segments of demand is a fundamental modeling decision. For airlines, a key input into the typical model is the fare structure and the grouping of fares for the purposes of forecasting. The model doesn’t work well if there are too many such groupings and the model won’t be able to differentiate among groupings that don’t vary significantly from each other. On the other hand, if there are too few groupings, the forecast model can break down from trying to impose one demand model on highly diverse sub-groups.
- Unconstrained Demand: The database must estimate unmet demand when the observed demand fills all of the inventory allowed it. When the plane went out full, what was the total demand? All systems have various procedures for doing this and none of them are perfect, given that they are estimating demand that wasn’t actually observed.
- Elimination of Outliers / Assessment of One-time Blips (that should not be included in the model): Group bookings, for example, are often excluded in demand forecasts given that they are not inherent in base demand.
Customer Behavior/Demand Forecasts
Dynamic pricing relies heavily on model forecasts of customer behavior – or demand for each product and each fare. The forecast is never simply an average of historic demand. It is the result of a series of transformations which are not always transparent and drive optimization results.
- Time-to-Adjust: A coefficient on historic demand governs how quickly a demand change is incorporated in the forecast. Frequently, airlines use a coefficient that dampens out sudden increases in favor of more gradual demand adjustment over time. On the other hand, during a period of rapid change, airlines should use a coefficient that weights recent history more heavily so as to more quickly capture the change in demand.
- Adjustment for Inelastic Demand: Some pricing systems attempt to account for that portion of customers at every fare level who bought a lower fare but would have paid more had that fare not been available - reducing the forecast demand for the low fare and increasing the demand forecast for the higher fare. The proportion is related to estimated price elasticity among observed purchasers of the lower fare.
- Seasonal and Day-of-Week Patterns or Adjustments: Just as the demand forecast is adjusted by flight by day as more information comes in, the patterns of demand – over a week or seasonally – are also adjusted. These adjustments and patterns drive each flight’s daily demand forecast.
- Linear and Nonlinear Extrapolations: Finally, dynamic pricing builds regressions on the historic, adjusted data to create forecasts. The forecast model captures trends and linear or nonlinear patterns. These trends are best-fit equations, often not based on any fundamental drivers like overall market growth or a competitor’s fare initiative. Thus, they may well not be robust, and they are highly likely to be replaced by a different forecast equation or updated coefficients in the next period or the season.
Optimization (Right Price at the Right Time)
Similarly, there are many variables in the optimization algorithm. Other than the demand forecasts, factors may include:
- Fare Differences Between Market Segments (Fare Classes): Even if the market segments are unique and non-overlapping, a small fare difference translates into less upside to holding out seats for higher fare passengers. Understanding how the model treats fare differences is critical to truly understanding how optimization works.
- Variance in the Forecast Model: Volatile demand for a high fare class drives the model to reduce the seats set aside for that class. High variance suggests demand is potentially “unforecastable,” undermining the optimization algorithm.
- Competitive Pricing Algorithms: Some airlines tap into automated systems to “match” competitive fares, which can override revenue management system recommendations.
Dynamic pricing is reliant on the three steps of data collection and categorization, development of models of customer behavior or demand, and optimization of pricing based on those forecasts. Transparency into the “black box” offers insight into how each of those steps ultimately drive pricing.