Consumer product companies, like airlines, regularly develop demand forecasts for thousands of products. Retailers, for example, must forecast not just how many men’s shirts will be purchased but what styles, colors, and sizes, and at which locations. In the case of airlines, we not only forecast demand for each of a thousand or so flights a day for the next year, but we often forecast 15 price points and 20 or more O&Ds on each flight. The numbers of individual forecasts can become overwhelming.
Demand forecast analysts for consumer product companies speak of certain items being “unforecastable.” They are unable to find any identifiable demand pattern based on history, and the variance in observed demand overwhelms a single point forecast.
Certainly, for airlines, high fare demand on certain flights can be deemed “unforecastable.” Such demand may often be 0, but it can be 5 or 10 or more when there is a key business meeting or important conference.
There are two basic methodologies for dealing with unforecastable demand:
Demand generally becomes more statistically relevant in larger groupings. Rather than forecast demand for $100 - $120 fare levels, forecasting demand for $100 - $150 may drive a more robust forecast. “Leg-based” forecast systems often treat all connect traffic in aggregate, rather than forecasting each connect O&D separately.
Often we deceive ourselves when we think we are being more precise with a more granular demand forecast; the greater granularity may result in “unforecastable” demand that defies statistical modeling. I worked with one airline that concluded its extremely granular forecasting process was adding significant overall forecast error. The recommendation was to apply some logical aggregation.
Optimal inventory allocations are the result of both statistical “averages” and “forecast variances.” The predictability of demand is as important as the “most likely case” in any optimization process - a large forecast variance for only a slightly higher fare may not justify an allocation of any inventory. In fact, the “optimal” result could be to ignore the high-variance demand. This is the practical implication of demand being “unforecastable.”
Group travel for airlines often represents “unforecastable” demand. Since “groups” represent demand that comes in “lumps”, the “average” forecast is meaningless. A group of 25 passengers will need 25 seats, not their probability-weighted average of 3.5. Thus airlines often separate group demand from their base forecasting processes. Most airlines apply their forecast algorithms to base demand -- excluding groups -- and then add group demand to the base forecast as it actually materializes.
“Group” demand blips may be thrown out by some statistical models as “outliers,” that do not fit in the historic pattern. However, it is better to identify these demand segments separately and explicitly exclude them from the base forecast.
Prioritizing Product Demand Forecasting
To deal with “unforecastable” products, consumer-oriented companies often prioritize their thousands of products. Those with high unit sales are the most important and merit greater attention. Low volume products may be more “unforecastable” but also may not be as critical for planning. Airlines that rely on huge computer models don’t necessarily do this prioritization explicitly – they often let their forecast systems apply statistical analysis to all fare levels across all O&Ds. In fact, many airlines I have worked with pride themselves on the detail with which their revenue management systems forecast demand. It is important for airlines to become more conscious of the “forecastability” of their products or price points. Specific recommendations for airlines include
- Monitoring forecast accuracy for their key products: Airlines with O&D revenue management systems direct most attention to the largest O&Ds (both local and connect). They need to monitor their ability to accurately forecast demand at this level – and periodically adjust their forecast process accordingly
- Understanding the implication of narrower fare ranges or O&D-level forecasting: Greater granularity does not necessarily drive greater accuracy. Consider aggregation as a means of gaining greater accuracy.
- Separation of “lumpy” demand:“Group” demand is a terrific example of “unforecastable” demand that is generally handled well outside the base forecasting system. There may be other such “lumpy” demand segments in your network that are better separated from the base demand forecast.
“Unforecastability” is a term that needs to be added to airline revenue management lexicons. Airline revenue managers too often overstate the value of greater granularity and, consequently, let their revenue management systems develop forecasts without proper oversight.