Surely for FF's they could profile based on the individuals no show history etc, and for non-FF's use statistical averages. I would be very surprised if Qantas did not analyse individual travel patterns for forecasting when it comes to frequent flyers.
Perhaps. Although you would hit at a theoretical problem. You would need a sufficiently large sample size (large number of flights
per individual) to be able to estimate your required parameters (rate of no shows, etc) at a practically useful level of confidence to take that approach. Thus would only work well for very frequent FFs.
The alternative is to model based on a population of individuals (your pax and would be pax) adopting two or more possible strategies (eg. buy flexible whatever the cost, buy cheapest and rebuy if necessary, etc, etc).
Your system features a population with certain % of population adopting each given strategy. You run a model based on optimising revenue (yield) and run in silico (on computer) experiments to determine the best intervention of combination of parameters, which you can control (fare levels, number of seats, etc).
Now the smarts of the model increases when you can incorporate dynamic positive or negative feedback between variables (this is where I question how smart this yield management game really is). For example, offering a sale fare not only resets a fare level parameter, but has a knock on effect on the difference between lowest restricted fare and fully flexible and thus perceived value - this in turn might shift the proportion of population adopting each fare buying strategy. In other words your parameters are inter-dependent and your system dynamic. You have moved from a relatively simple "supply and demand" model to a "dynamic" model.
The validity of your model now depends on how well you capture the
inter-relationships between model elements, particularly the human element of how perceptions shift given changes in context (price change, economy, etc), not just how well you estimate the probability of events (no shows, etc).
Just to make it more fun your basic model will probably have just one optimal solution allowing you to adjust your yield parameters (seat price/availability) under the assumption that you are aiming at one stable optimal solution. On the other hand, the dynamic system model may tend towards more than one stable state, each of varying value to you in terms of yield. Encouraging a system to go in the direction you want it to go (more yield in this case) is a fine art and one with which we struggle in a number of obvious pratical examples (including the economy, climate and genetic engineering!!!).
Can I add yield management to this list - do folk feel that QF has got its maths deployed at the cutting edge???
One final thought. When estimating parameters QF for their models have a lot of data - loadings on all flights broken down by fare type, when ticket purchased, some travel profile info, etc. Plus their market research. BUT most (if not all?) is based on existing customers. If they really weant to increase their yield, they must also increase their market share in terms of individuals' airline preferences - however, the always or mainly DJ folk may or may not be effectively represented by their QF customer based data...
