Capturing Patterns And Trends In Passenger Demand
For decades, determining the actual passenger numbers for a past time period was difficult, which made it nearly impossible to accurately forecast future demand. Now that the historical picture of airline passenger travel can be more precisely constructed, established modeling methods yield more accurate predictions.
One of the fundamental inputs to any commercial planning study at an airline is the demand expected in each of the origin-and-destination (O&D) markets involved. For example, a long-term fleet acquisition plan requires an estimate on the growth outlook of airline demand in the region. A network plan involving the introduction of a new route requires a demand estimate on all key O&D markets flowing on the proposed route.
Knowing whether a particular market is going to grow at 5 percent or 10 percent per year can lead to completely different strategic decisions.
There are several dimensions to the required forecast, which depend on the nature of the study. For example, in a typical network planning study, the question asked may be, “How many passengers will fly from city A to city B in December 2014?” Or possibly, “What is the minimum year-over-year growth rate that is confidently expected for passenger demand from country A to country B in winter 2014?” A few dimensions to consider include:
- Time Horizon: Are the numbers required for the next season? The next year? How about five years ahead?
- Geography: In which part of the world will the numbers be used?
- Granularity: Are the numbers needed for a region as a whole? For a specific pair of countries? Or even a specific O&D market?
- Confidence: Is an expected forecast or a conservative forecast required?
In terms of ordering a study, it is common to ask a leading question followed by a slightly different follow-up question to make an informed business decision. For example, perhaps the growth forecast for the next season is the leading question, and a longer-term outlook is the follow-up question. However, the data set and forecasting methodology can change considerably when one of these dimensions is altered, resulting in long lead times between ordering the study and receiving the results.
Modern Data Sources
Traditionally, Marketing Information Data Tapes (MIDT) provide the primary data source for an O&D view of any market covering all airlines. However, MIDT only covers airline bookings done through a global distribution system (GDS). With the growing focus on direct distribution channels and the rise of low-cost carriers (LCCs), the percentage of global airline bookings that go through GDS channels has been steadily dropping at roughly 1.2 percent each year for the last three years.
Therefore, if someone is only looking for a growth percentage number, forecasting using MIDT will yield the wrong answer. Applying a correction factor to account for the drop in MIDT penetration is not straightforward, because the drop in penetration is not uniform. Short-haul markets, for instance, may see a greater drop than long-haul markets, and the entrance of an LCC can result in an abrupt drop that is difficult to model.
This leads to an even more fundamental question: Do we know the true market demand at any given point in the past?
Fortunately, this question is important enough that several concerted efforts have been made in the industry to generate the right answer. For example, Global Demand Data, offered by Sabre Airline Solutions®, uses proprietary algorithms to estimate the true picture of airline passenger demand using a variety of data sources collected from around the world.
As a result, solving the problem of market-size forecasting becomes a lot more palatable. Given the historical market-demand pattern over a period of time, is it possible to forecast the demand for a future period? This fits neatly into an established field of analytics called time-series forecasting.
A time series is simply a series of data points measured at regular time intervals. Time-series forecasting encapsulates a number of techniques well explored in business and academic literature, which help identify the underlying pattern in the series and extrapolate it.
In a typical scenario involving time-series forecasting of O&D passenger demand, the seasonality in the demand — an underlying, repeating pattern of increase or decrease in demand that is characteristic of a region at a particular time of the year — is first identified. Once the seasonality in the demand has been identified, the demand can be adjusted to account for this particular season.
The demand series that remains — known as the deseasonalized data — is highly revealing. The trend in the demand over the years will now be a lot more visible.
The final step involves separating the trend from the noise and modeling it in a way that can be forecasted.
This, in a nutshell, is time-series forecasting, and it is very popular because of its intuitive and visual nature. Both the seasonality and the trend reveal insights into the market, and analysts and decision makers, as a result, tend to be on the same page.
Time-Series Forecasting Challenges
Time-series forecasting techniques have been researched for decades, as have advances in handling huge amounts of data and the algorithms that provide a true picture of historical demand. However, there are a number of complications that need to be addressed — made all the more difficult if the intention is to automate time-series forecasting to work on several thousand country pairs or thousands of markets in the world, with differing properties.
One type of complication that can arise is multiplicative seasonality, a situation in which the degree of seasonal variation in the data increases (and decreases) with the trend. The more typical seasonal variation that remains static while the trend increases or decreases is called additive seasonality.
Another example is the handling of unusual events that skew the travel pattern temporarily. For instance, the eruption of the volcano in Iceland in April 2010 left a sizeable dent in traffic in Europe. If the seasonality for Europe is naively extracted without accounting for the disruption in April 2010, demand for that month will inevitably be underestimated.
By now, it should be evident that demand patterns vary dramatically for different markets around the world. Developing automated techniques that are flexible and advanced enough to accurately forecast demand can present quite a challenge.
The operations research team at Sabre Airline Solutions is developing an automated market-demand forecasting tool and in the process has gleaned valuable insights and promising preliminary results. The team processed demand data at various levels of aggregation, fitted the demand series against Global Demand Data up to December 2012, forecasted 10 months into the future and validated it against actual data from the Global Demand Data system. The metrics used were:
- MAPE = “Mean Absolute Percentage Error” = Weighted Average of Absolute Value of [(Forecast – Actual) / Actual ]
- MPE = “Mean Percentage Error” = Weighted Average of [(Forecast – Actual) / Actual ]
As expected, the greater the degree of aggregation, the better the forecast. Also, the forecast is more accurate for the top markets (or countries or region pairs). Both observations indicate that while there are high-level statistical patterns that are predictable, there is some volatility at a deeper level that is driven by daily business decisions and the economic forces of demand and supply. Some examples include:
- An airline introduces a non-stop flight in a market historically served only by connections, and the market size jumps.
- An airline slashes fares in a market in response to competition. This starts a price war in the market, and the market size increases as air transport pulls demand away from other modes of transport.
- A country experiences a recession and domestic market demand plummets.
Given the various possibilities by which a market can change unexpectedly, these metrics are reasonably good. The path of improvement, however, leads into challenging and uncharted territory.
To account for the volatility in smaller markets, it is possible to introduce input variables or features into the model. In other words, forecast the market demand as a function of some factors that could potentially be incorporated into the model, including:
- Economic variables, such as gross domestic product, which may present challenges including:
- Availability for the countries in question,
- Sufficient number of quarterly data points,
- Reliable forecast of the economic variable is needed to forecast the market demand.
- Published airline schedules, which may present challenges such as:
- Schedules represent the leg view and no easy way is available to convert the schedule to an O&D model,
- Future airline schedules are reliable only up to a certain number of months.
The forecast quality can also be improved by taking into account factors that humans intuitively know, but the computer does not. For example, markets involving Libya were under-forecasted because the model didn’t know the country had recovered from the war.
Thanks to advances in data management and processing, the industry has come a long way from the days when no one knew how many passengers flew which airline on which route at a given point in the past. While the question of, “What happened?” has been satisfactorily answered, “What will happen?” is now a priority.
There is a need in the industry for a framework to forecast O&D market sizes under different input parameters, and it is an active area of research that is picking up momentum. As with several other advances in the industry, this effort is enabled by data quality, quantity and ease of computation, which was previously difficult to obtain.