Tom Bacon does not dispute the value of analysts but argues that there is definitely a central role for machine learning in RM

Great revenue management analysts are invaluable. The 5+% increase in revenue attributed to a sophisticated RM system cannot be achieved without well-trained, highly experienced analysts who can properly maintain the system and intervene when appropriate. This means that, potentially, groups of 50 individuals could be responsible for $1-$2 billion in incremental earnings a year at each of the larger airlines - $20-$40 million per analyst. Give these guys a raise!

Some airlines are experimenting with machine learning. Can the computer ‘learn’ from experienced analysts and replicate their interventions? At an Eyefortravel conference, easyJet, for example, explained how machine learning reverses the normal modelling process.

So the normal flow of big data analytics used in most RM systems goes like this.

• Data => Computer => Programme => Output

Here, the system takes the data, builds a model around the data, and produces output (demand forecasts by fare and recommended inventory allocations by fare).

With machine learning, on the other hand we have:

• [Data + Output (Analyst Intervention)] => Computer => Program

In this case the big data analytics approach is applied to analyst interventions themselves. So, the system can learn from the analysts and automatically adjust the demand forecasts/allocations without human intervention.

First, of course, we need to acknowledge that all analyst interventions – overriding the big data analytics modelling and recommendations – need to be performed with caution. Academic studies have measured a negative value-add from many analyst interventions. Even experienced analysts can intervene too often – or their adjustments may be too large or too small. The new programme must have a way to assess the actual value/benefit of the interventions or the machine could be learning poor practices! Ultimately, machine learning needs to have a way to evaluate each intervention; in many cases, the machine will need to learn to ignore or offset interventions that don’t add value.

Let’s review some standard analyst interventions and identify positive learning opportunities –interventions that may add value and which, in a machine-learning context, can be replicated by the machine.

1. Overall demand adjustment 

When market demand changes dramatically, RM systems may be slow to react – they are designed to change the forecast only when an increase/decrease persists for a period of time; they don’t want to adjust the future based on a one-time ‘blip’. An analyst may be able – based on external factors like a schedule change, fare changes, or market-specific news – to determine earlier whether a change is likely to persist.  Intervention can thus speed the model to adapt to the new conditions.

Of course, the model could take schedule and fare changes as inputs and, through machine learning, apply appropriate adjustments. In fact, the model may be better able to quantify the effect of a new competing connect service versus a non-stop and the effect of a three-day sale versus a longer-term sale.

Sometimes, however, the analyst is not actually bringing external market-specific information to the model. Sometimes, he simply judges that the recurrent period of strong bookings represents a significant break from history and that the model is too slow to adjust its forecast.  Machine learning – finding a consistent pattern in such interventions - can increase the sensitivity of forecasts to demand changes. Such an adjustment may apply on a macro basis or a micro, flight-specific basis.

2. Inventory Adjustment

Another common intervention involves directly modifying recommended inventory allocations. Analysts effectively apply business rules to their markets – either implicitly or explicitly. Example business rules are:

• If a flight is booked at 80% 14 days out, close ‘S’ (a low fare revenue ‘bucket’)

• If flight is not booked at 50% 21 days out, open ‘S’

Although not recommended by most academicians – since the model has an explicit and sophisticated algorithm-based rationale for opening and closing buckets – this intervention can add value when price fences are not effective or when forecasting is particularly difficult. Machines, of course, can definitely ‘learn’ these – they can make such rules explicit and make the adjustments automatically.

To conclude, I would argue that fundamentally, machine learning can replicate most analyst interventions and, in fact, make such adjustments more disciplined and consistent. It can ensure that changes to the model are appropriate and based on a track record of improved results. So there is, without doubt, definitely a role for machine learning in airline revenue management.

Tom Bacon has been in the airline business for 25 years and is now an industry consultant in revenue optimisation. Email Tom or visit his website for more insights