AI-Fueled Innovation: Revenue Management
As the industry continues to evolve, airline revenue management is evolving right along with it. This session, co-led by PROS Senior Director of Product Management Justin Jander and key innovation customer, Lufthansa Group’s Head of Commercial Offer Methods & Automation Florian Martin, reviews industry-leading innovation in AI and forecasting that are set to further evolve revenue management practice and strategies to maximize revenue as the industry moves to embrace modern airline retailing with offer and order management.
Full Transcript
Justin Jander: All right, I think we'll get started. All right. So first I'll take a minute and introduce ourselves. We'll start with Florian. Go ahead.
Florian Martin: Hi, I'm Florian, Florian Martin. I'm from Lufthansa Group, responsible for business development, so anything kind of research development and systems and business intelligence. And I think we have a long history together with PROS, developing a lot of systems together, RTDP, forecasts, group systems, continuous pricing, you name it. I think we have a history for collaborating together also successfully, I would say. And what we're going to present today is just another example I think of that very fruitful collaboration....
Justin Jander: Indeed, yeah. And so PROS is very happy to have Florian co-present with us today and really talk about some of the innovation that we've done together, and as the... First off, I'm Justin Jander, I lead our product management team for RM, RTDP, GSO and our dynamic offer solutions as well. And so as the title says, we're going to talk about disentangling, which I've always found a bit of an odd word, because wouldn't that just be untangle, not disentangle? But then I googled it and disentangle is actually a real word. And this word was brought into our space on the revenue management side in about 2016 or so, a group of Lufthansa Group and a couple other folks wrote a paper called Disentangling Revenue Management and Pricing. And I'll be honest, I read the paper multiple times, didn't understand almost a word of it because it's really, really theoretical.
Justin Jander: And after some time, we've worked through some of the details of that, working with the Lufthansa Group to understand what was meant by disentangling, understand what was the objective and what the goals were with it. And that's what brought us where we are today with a little minor pandemic in the middle, where we had to slow down a bit of progress in-between, but that really gets us to where we are today. And so what we want to do is jointly introduce the concept to you and then Florian's going to give you a bit more detail on why disentangling is the approach that Lufthansa, the Lufthansa Group is taking. So with that, as we dive in, a bit of detail. So when we... We saw, if you happen to attend Michael Wu's session this morning, he talked about pricing being a factor of supply and demand on the airline side. That's of course true, but we're going to dive one level deeper into that, and really focus on the elements that make up the price.
Justin Jander: And we see it as there's two elements that really determine the price. There's the pricing perspective from the airline, and there's the pricing perspective from the passenger. And those two things are, one of those is controlled by the airline, that's the airline perspective. They control the components of that. The other is the passenger perspective. And as much as the airline might try, you don't control what the passenger does, right? You try, but sometimes you're successful, maybe in other times, not, but really, what we're looking at is those two components. From the airline perspective side, what we're really talking about that you have control of is the capacity and the schedule. So those are the things you really control from a pricing perspective.
Justin Jander: First, making sure that you put the planes in the right place, so that people will want to fly and value that product at a certain amount. And so when we're considering that, there's things that you're interested in as the airline from that side, you want to see how many seats can you sell? Do I have more people that want to fly than there are seats to sell? And if there are more people who want to fly, how much do they value those seats? So we're looking at the value and demand that make up that component. On the flip side is the passenger's perspective. This is thinking of it more of a retail problem, and Florian really used that terminology when we were discussing this, that with the passenger side, this is an e-commerce problem, how much is the passenger willing to pay for a particular product? And so the idea here is we really want to focus between the airline perspective and the passenger perspective, okay?
Justin Jander: So this is how we've kind of framed the problem. And then as we move forward, when we look at how the airline industry has evolved, what do we look at as RM and pricing in a fenced world? So fencing is filed fares with rules in place that break up the exact kind of product definition inside of the RM space, and the pricing space. So first the way that worked was you forecast class-based demand. So you can see this is a great chart. If you've got five classes, you forecast the demand there. And the idea here is that the demand for Q-class for example, is I know that the passenger wants the letter Q because I put conditions in the fare that dictate that they want the letter Q. So they've got a Q stamped on their forehead. I'm trying to sell as many Qs as possible, et cetera, and maximize my revenue under the assumption of these independent fare classes.
Justin Jander: We take that demand, we run it through a dynamic program to get a bid price. So now I'm holding a bid price which represents the value of the next available seat on the plane. So if there's not enough demand to fill the plane, your bid price is zero. If there is enough demand to fill the plane, then the amount of the size of the bid price represents the value of the next seat. So how valuable is that demand? So those components mixed together, and then you go into a pricing decision. And on the pricing decision, what we're really talking about here is understanding availability, but it's contingent on some type of rules already being in place that designates the fare classes themselves. And so that's important here because, those fare rules are either in the filed fares themselves, or they've been replaced inside of RTDP or an equivalent or subpar availability engine. Come on, somebody gets a little subtle hint there, subpar availability. All right, never mind. All right. I'm going to move on. We'll think about it a little bit there. There's one clear winner.
Florian Martin: Let it sink in.
Justin Jander: All right. All right, thank you. I got like one laugh out of that, so I'm going to at least move on from there. All right, so the whole idea here is that we were in this scenario, there's a disentangled concept in place, because you focused on the airline perspective on the left two pieces, and you focused on the passenger's perspective on the right hand side, because your filed fares, a pricing analyst was going through, setting the prices and setting the rules inside of that. It could have been through an AP rule (advanced purchase), so you set a 21-day AP, dictated that I want the passenger to have that segmentation. I want to focus on that piece, and I set a price for a passenger that's willing to pay the Q-class fare, because they want the product associated with Q, which is a 21-day AP.
Justin Jander: Now some airlines started peeling out those AP rules so that you had fewer fences in place in the filed fares, but most airlines just replace those inside of your availability engine, whether it be through day's prior rules or even through load factor strategies, for example, which not always the best decision. So what we said was, well, let's actually model that behavior together. Let's entangle those two problems. So that's what we moved into the entangled world here, where the first thing you need to do is forecast the relationship between price and demand. So that relationship between price and demand is of course elasticity. When you understand the relationship between price and demand, you can take that component and perform a fare transformation. The fare transformation is saying, given the likelihood a passenger is going to buy down, how do I prevent that with the fare transformation, closing class, adjust the marginal revenue for that, and force the passengers to buy up into higher classes to maximize my revenue.
Justin Jander: Great, so far so good, we take those transformed fares that we got out of the fare transformation step, put those into our optimization, and we still get a bid price out of that component. So now, we're holding this... Almost fell down. That was good. We are holding the bid prices, we're holding those transformed fares, and now we can make an availability decision. But now instead of the availability decision being based off of the daily fare versus the bid price, it's now the transformed fare versus the bid price. So the entangled part that we're talking about here is that we've mixed together, we mix together the transformed fare in the bid price determination and the transformed fare in the availability determination. So I take one calculation of elasticity, and use that for both bid price generation and availability generation. And that same transformed fare can be used in a continuous pricing algorithm. So just like if you were in the previous talk, we talked about continuous pricing, that same transformation concept can be used inside of the continuous pricing algorithm where now you're able to determine a price, instead of an availability.
Justin Jander: So of course, as you just kind of heard, those are all entangled together. So now the airline perspective and the passenger perspective are combined together, and there's nothing inherently wrong with this. In fact, the theory would tell you that there's a lot right with this, but there's also a different perspective and that's where I'll turn it over to Florian and hand him carefully so I don't fall down.
Florian Martin: Yeah, I mean as you said, there's absolutely nothing wrong with that per se. It's absolutely valid. I think it often also comes out of an enhanced focus on an O&D (origin & destination) perspective, right? When we had originally leg level controls perfectly fenced, it was low transversal shares, you didn't have much control over single passenger streams. Then it came the whole OD perspective, where you wanted to know a lot more about how different passenger streams react to changes in price to your bid price control. And of course, if you clump everything together, you end up in this entangled setup. But that's fine, unless you want to have a very sophisticated OD steering in place, I would say. The more you put into this passenger perspective, the more it becomes a burden to have this fully entangled, because there's also like, I like to describe it to the people in our organization from the market perspective, from the passenger, the OD perspective. It's about the passengers, it's about their attributes, what they want to buy, their willingness to pay their preferences, it's about their purchase decision.
Florian Martin: But this is completely independent from all the complexity that you face as an airline, that the fact that you are operating a huge network of thousands of flights that all interconnected because they share the same passenger streams. The fact that we sell these seats over a long timeframe of up to a year, which creates all these temporal dynamics, you as a passenger don't care about all this complexity. You are having a certain willingness to pay towards a certain product to get from A to B in a certain condition, and that's what you are interested in, and that determines what you're willing to pay. So for us, it became apparent. If you want to get more granular on this passenger perspective, re-run into issues on some of the other complexity that you have in a network world.
Florian Martin: So the most obvious thing I think is that you just get smaller and smaller OD buckets and at some point there's just zeros and you don't have any data that you can actually base your predictions on. But I think there's also like a lot more less technical arguments for that. So that kind of got our people thinking back then in like 2016 I think is when it was published. And it says novel here, because I think it is still quite novel, but it's not a moonshot. It's important to me because first of all, we have implemented our version of this already, and we're working right now together with PROS on products that are exactly doing that. So we are not talking about like an idea 10 years plus down the road, we are talking about something that is happening now. And basically our idea is to say, well, if you would separate this passenger perspective very explicitly, let's assume that for a second. We can do this completely independently.
Florian Martin: Well what is actually then the challenge that is left on the capacity side? Well, we don't need an OD perspective anymore, because that is a market passenger thing. We actually need a bid price, and the bid price lives on leg level. It's a value valid for a single flight. So actually, when you follow this thought, it actually it frees you up on the bid price calculation side to go essentially back to leg level, which is a positive thing, because you lose a lot of the complexity that the OD perspective brings in the bid price calculation. Suddenly, you can clump all this data together you need to know to calculate a bid price is exactly one thing essentially. You need to know if you are, let's say 100 days before departure, if you've charged $100, how many people are going to come on average? That's what you need to know. And you need to know the same for 99 days before departure, and 98 days before departure, and you need to know it not just for $100, but for other values as well.
Florian Martin: But this is all information that you can easily have on leg level. So the idea is to say let's divide and concur. Let's have a specialized forecast on leg level that can be class based or class independent. We do class independent, and you basically get a price demand relationship on leg level, and that has some very neat properties because... And I'm first going to show the second side and then show you how it all holds together, because if I can kind of simplify this idea, I get a lot more like statistical accuracy on that side. I still can cater for all the network effects on that side, and I'm going to talk about this in a minute. But what it now enables me to do is I have a completely independent and free, baggage free if you want, no network problem, no temporal dynamics problem on the passenger side.
Florian Martin: And I can use whatever fancy machine learning method that you guys can think of on that side, because it doesn't have to deal with all that network complexity anymore. So what I'm going to do essentially is have a very sophisticated OD level model, where I'm going to go very much granular into the passenger type, their preferences, how they react to changes in price, like doing willingness to pay effects, super detailed. I can even go into the product dimension if you please. And I can do this with a wide variety of off-the-shelf methods. That's kind of what we say, that's the e-commerce problem essentially. Pricing without capacity constraints. Nice. No, not airline specific anymore. And you can do this, get your transformed fares if you please, and then essentially bring these two things together in your availability mechanism. And in this case, in this example essentially, what we have is that if you look at this, we have bid price of 25, but if you see it's only Q-class and up available, so B and V are not available, even though... Then they would not be available even if the bid price is zero, because the pricing side said that essentially these two classes are inefficient, because all these people up here, they would buy down here, so we're going to close these two classes irrespective of capacity constraints. It's the 25 bid price that closes another class. It also closes Q-class, and we end up at M availability.
Florian Martin: Now that's all fare and nice, but how does this whole thing hold together? And there's one nice property I think that is often undervalued, which is yes, this list on leg level, but how do you ensure that these two things kind of work in sync? You have two models now. Isn't that kind of a duplication of effort? Isn't that bad? What I would say AI has learned as one thing, sophisticated strategies don't have to be one monolithic model. It's in the smart orchestration of models where you can really, as Michael Wu said beautifully today, get the best of both worlds. It's just where the art of AI comes in is in finding setups where this works, and how this works is essentially by you saying, well, I'm not the... I said we're building a price demand relationship on leg level, which is true, but it's strictly speaking not demand over price, it's demand over bid price. And I'm just going to make one statement there to illustrate why this is important. What that does is bid price is your control. It tells you, what this model tells you on leg level is, if you have a bid price of a 100, you get this much demand. If you had a bid price of 150, you get a bit less demand.
Florian Martin: What it basically doesn't tell you what happens on the customer end price side, it just tells you what happens between your control and the customer reaction. And in-between, there's the whole pricing side. There is this model in-between, oh, sorry, that was actually one step. Wow, that was quick. That was Michael Wu's experience today, but it was my mistake obviously. I just want to say that what you have here essentially is, this whole side is treated as a black box on that level. And that's good, because essentially what this model tells you is saying, I'm setting a certain action with my control variable and I'm getting a certain reward, which is a certain number of bookings. See the language already? Action-reward. That's kind of a machine learning paradigm, right? That means that whatever happens in between can be enormously complex, highly dimensional, can be neural net, whatever.
Florian Martin: You don't care on this side. You just need a relationship between action and reward, and that can be very simple. And this is what this model does. So it is not really aware of this model here. It just says, if I do this, something will happen over that on that side. I'm just learning what happens to my reward. That's it. And this side says, well, I actually don't know too much about this side, because how people react to price is irrespective of whether you fly with a 777 or with a 330. It's irrespective of capacity, it's irrespective of volume. Whether 100 people fly or 1000 people fly doesn't change the willingness to pay off a single passenger.
Florian Martin: So you see, by smart orchestration, you can specialize both sides of the problem to have a leg level forecast, optimize a combo that gives you a bid price, and you have a highly dimensional, complex model potentially on the pricing side, which is on OD level. And if you combine them, it still works in any classic availability mechanism. And maybe to wrap up, some of the advantages that we see and experience, because we have to set up is that, separating kind of these airline impulse constraints from the passenger focus frees you up enormously on the willingness to the pay side. You gain so much flexibility there. It also makes the whole organizational split and task split among analysts easier.
Florian Martin: For example, we only have two user roles. Not three, for example, as most have. We only have people setting a bid price, doing a leg level demand forecast ideally, and setting a bid price. Then there is on the other side, there is someone who determines the willingness-to-pay, sets the fare restrictions, determines the fare structure, and that determines the price. It's very simple. I think this has a lot of potential also going forward, because you can use, again, many fancy AI models suddenly become accessible because you don't have the complexity of network optimization. And last but not least, we also, not just claim this, we also test this and we are quite thorough in our testing of our solutions there. We put quite some effort into this. We say here, the revenue uplift, by only that sophisticated OD steering side, the OD sophisticated willingness-to-pay pricing side, that's only one part, the lower part of the slide that you saw. That thing alone, we have already rolled out with PROS, it's called Request-Specific Pricing. You might have heard it in several talks already around. Michael Wu also mentioned it today.
Florian Martin: The first version of a predecessor of this system, which was in-house, Lufthansa-developed, gave us a 2.3% physically significant revenue uplift. With the new PROS solution, we upped this to 5.2%. This is net of continuous pricing that's even coming on top. We have high confidence that this is really worthwhile to do. We're not saying it's strictly better in all scenarios than an entangled approach. I think, it depends on what kind of carrier you are and what set up you have and how advanced you already are on the willingness-to-pay side. We think it has a lot of potential for many carriers, already now, and even more so going forward.
Justin Jander: Great, thank you very much, Florian. That was really insightful.