As we emerge from the pandemic, both disciplines of airline revenue management (RM) and B2B pricing are posed with great challenges. Airline revenue managers must now work with shorter relevant histories for demand forecasting due to changes in people’s travel behaviors. Meanwhile, B2B pricing managers are often faced with supply shortages and volatile shipping costs due to the disrupted supply chain. As our teams of scientists help our customers address these new challenges, we’ve created the blueprints for many innovating solutions. These innovations are pointing to an interesting phenomenon—although the nuances of the airline vs. B2B world have always called for different approaches to pricing, they are now converging. Consequently, airline revenue managers and B2B pricers can learn a lot from each other. PROS Chief AI Strategist, Dr. Michael Wu, will unveil the innovations we made in RM and B2B pricing over the past year while discussing the greater economic forces behind this convergence. This understanding of the “why,” will reveal a strategy for moving forward in a world of greater economic uncertainty while protecting us from future black swan events.
About the Speakers
Dr. Michael Wu is Chief AI Strategist at PROS, a Thought Leader and Author on AI, Machine Learning and Data Science. His research spans many areas, including customer experience, CRM, online influence, gamification, digital transformation, AI, and more. Wu's R&D won him recognition as an Influential Leader by CRM Magazine along with Mark Zuckerberg and other industry giants. Wu believes in knowledge dissemination, and speaks internationally at universities, conferences, and enterprises.
Ezgi Can Eren serves as Lead Scientist at PROS specializing in revenue management and pricing optimization, analytics, Stochastic and Quantitative modeling, applied probability, and operations management.
Guangrui Xie is a data scientist with more than six years of solid background in building predictive models and optimization algorithms using statistics, machine learning, and optimization techniques.
Jiabing Li is currently a Scientist on the machine learning team at PROS. She has a Ph.D. in Electrical and Computer Engineering, focusing on Image Processing, GPU computing, and Machine Learning, and is proficient in programming in Python and C++.
Jonas Rauch serves as Lead Scientist at PROS and is an expert in Revenue Management (RM) and Pricing science and research, focusing on the next-generation RM and pricing methods for travel and non-travel industries. Rauch has 10 years of experience in Revenue Management strategy and research at Lufthansa Group where he was responsible for developing state-of-the art forecasting and optimization methods and integrating them with business processes. There he developed the long-term commercial strategy of advanced offer management through modern distribution channels (NDC).
Kyle Schoener has over seven years of experience using operations research and revenue management techniques to solve business problems throughout the life cycle of a product. His background includes sales engagements, identifying the optimal solution, building prototypes, product management, and implementation of the solution at customers across the hotel, cruise, and business to business service providers. Schoener currently serves as Science Manager at PROS leading a team of scientists in implementing, configuring, and innovating pricing/revenue management solutions. He specializes in Revenue Management, Segmentation, Price Optimization, Forecasting, Implementations, and Technical Sales.
Wenshen Song has worked at PROS for two years in the Science and Research department, applying machine learning and deep learning methods in building models and solutions for price guidance, opportunity detection, and sales prediction. Song received his Ph.D. in Physics from Washington University in St. Louis where is was also a Research Assistant in Computational Laboratory of Quantum Materials (CLoQM), and a Teaching Assistant for undergraduate Physics.
Yan Xu is currently working as a Data Science Manager and Senior Scientist on a machine learning team at PROS, working on next-generation products of intelligent sales and pricing tools for modern commerce. She is also the founder and leading speaker for the machine learning meetup in Houston area. Xu received her Ph.D. form the University of Houston with research focus in machine learning, computer vision, and bioinformatics.
Michael Wu: Hello, everyone. I'm Dr. Michael Wu and I'm the Chief AI Strategist at PROS. Welcome to the homestretch as we are almost near the end of Outperform 2021. Today, I want to talk to you about a rare and interesting phenomenon called innovation convergence. Now this is a rare event because under most circumstances, technological innovation will typically diverge and become more different rather than converge and become more similar. So when something unusual like this happen is always an opportunity for us to learn something from it. Now with this audience I don't think I need to reiterate that pricing is actually a core competency of PROS. We actually have pricing solutions for the BBB side, and this essentially boils down to some kind of willingness to pay estimation algorithm. And we also have pricing for the airline side, and these are our flagship revenue management system that we've been supplying to the airline industry for many years....
So if you look at these two sets of solutions, they actually look nothing like each other. So an interesting question is, how did this pricing practice evolve and become something so different? So despite their differences, they do actually came from the same basic laws of economics. So if you recall from your econ 101, you might remember that the optimal price that maximizes your revenue. Is this concept of equilibrium price and this equilibrium price is achieved when your supply curve crosses your demand curve. Now, most supply and demand curve aren't just simple straight line that's drawn here, and these are merely a, I would say, a local approximation to the more realistic supply and demand curve that you can see in the real world. Now how this simple idea of matching supply and demand, turn into willingness to pay and turn into Revenue Management. So let's take a look at the airline case first. With airline what we're trying to do is that we're trying to price a seat on the flight. Now, if there's anything you know about seats on the flight is that their supply is actually fairly inelastic. And this is actually illustrated by this near vertical supply curve here. And what that means is that as demand increases, the airlines don't really have an easy way to increase the supply of seats to match that demand to maintain this optimal equilibrium. Now, airlines sometimes can actually swap planes, and they can make minor adjustments to the number of seats. But largely speaking, when the capacity of a particular market is determined at the network planning stage, the total number of seats for that market is largely fixed. Now, since airline cannot change the supply of seats very easily. What they do instead is they change the price of those seats to control the demand for their seats, and they will change the price to shift the demand for those seats to match the supply of seats that they have. And now they can achieve this equilibrium again. Now what allow airlines to do this effectively is really the predictability of people's travel behavior. Turns out, people's behavior in general is actually very predictable, and that includes travel behavior. In fact, there are books published saying that we are so predictable that we are basically predictably irrational. The bottom line is that if you can know the future demand for these seats, then you can price these seats optimally to shift the demand. So that it matches the remaining seats that the airlines have. And that's actually the basic idea behind Revenue Management. For B2B this situation is actually very similar to retail pricing in this age of globalization we have infinite production capacity so that we almost never run out of supplies. So we are basically operating in this range of infinite supply. And in this range, the supply curve is essentially flat and this is due to the economy of scale. When you take this economy of scale to the global level, the marginal cost of production is essentially zero. Basically, if you're able to produce millions and millions of units of goods to supply at the global level to the global market, producing one extra unit is not going to cost you anything more. And that's why the supply curve is so flat in this range. Moreover, the competitiveness of the market is going to keep the supply cost very constant and very stable. So what that means is for B2B and retail pricing we can treat the supply cost as a fixed cost that never changes as the demand fluctuate, so we can simply ignore the supply curve altogether and just focus on modeling that demand elasticity. Now what enable retailer and B2B to take advantage of this is really the efficiency of our supply chain, and that supply chain had been so efficient that most B2B and retails is able to find equivalent replacement product at the same price somewhere else. Finally, I just want to emphasize the main difference between B2B and retail pricing is that B2B pricing is negotiated based on business relationship and contacts, what that means is that they don't even need to be priced at equilibrium, they can be priced around equilibrium based on a particular customer's willingness to pay. And that's why B2B pricing essentially boils down to estimation and willingness to pay. Now, these solutions actually work extremely well until 2020 when the pandemic disrupted everything. The pandemic disrupted Revenue Management because they disrupted people's travel behavior. Now, the demand for these seats are much more volatile and they become harder to forecast. Airline used to be able to leverage many, many years of history and forecasts far out into the future. Up to about a year. And that's why people can book their flight up to a year in advance. But with the pandemic, the airline can no longer leverage this long history anymore because they're no longer relevant in forecasting future travel. Instead, they're stuck with a much shorter, relevant history. So what that means is that they can also only forecast a much shorter window into the future. Since everything just compress in the time domain, sometimes this is called the temporal squeeze problem. Now, interestingly, this is a problem that the retailer and B2B have dealt with for many years, and because the market is so competitive that buyers can always find an equivalent replacement easily. So demand volatility is not just common, but the norm for B2B and retails. So B2B and retail always have to deal with this temporal squeeze problem, and typically they don't even try to forecast their demand more than a few months out. But one thing that the retailer finds that’s extremely useful in dealing with this problem of temporary squeeze is data variety, since they can't use longer history, which essentially is this more of the same kind of data. They use more variety data, which essentially is a different kinds of data to help them do the forecast. So it's very common for retailers to leverage everything from purchase history to a clickstream to social media and literally anything they could get their hands on, to help them forecast the demand for their product. And this is something that the airline can really learn from the retailers. On the other hand, the pandemic has also disrupted B2B pricing because they have disrupted our supply chain. Today, we no longer have infinite supply, but instead we're constantly out of stock. And because it's a reason to supply cost and even the shipping costs have been steadily increasing. And become more volatile. So B2B and retailers can no longer treat their supply costs as a constant fixed cost anymore. Instead, they need to factor this cost into their pricing decision. Now, surprisingly, this is something that the airline had been doing for many decades with their practice of Revenue Management because the number of seats in the airline is always constrained. They always have to price their remaining seats differentially based on some estimate of opportunity costs based on market demand. And in Revenue Management. They will forecast this demand and then optimize the price based on that forecast of demand. And this is an approach that B2B can really learn from airlines. So as you can see, the challenge for B2B is actually an expertise for the airlines and vice versa. So what that means is the airlines can also learn a lot from the B2B and vice versa too. And PROS is actually at a very unique position to facilitate this knowledge transfer because not only do we have customers on both sides, we also have the science and the technology and the working experience on both sides. So even though we have experienced a crisis of a century last year, it is also an opportunity of a lifetime. As Winston Churchill once said, “Never let a good crisis go to waste.” We certainly didn't let this crisis go to waste because we innovated so much during this time. For airlines, we are learning from the retailers by incorporating more variety of third party data sources into our forecasting pipeline. Specifically, we're going to talk about our research on leveraging shopping data, as well as the events data to improve booking forecasts. On the B2B side we're learning from the airlines too. Inspired by RMs forecast and optimize approach with build wROXANNe as our next generation B2B pricing innovation. Now, to make the complex discipline of RM more accessible to business who have never done RM before. We also build DiANNe. Now what better way is there to introduce these innovation than to hear directly from the scientists who built them? So for this next portion, I'm going to invite the scientists who work on this innovation to join me and discuss some of their respective work. So for the research that involve the use of shopping data to improve booking forecasts, I want to invite Jiabing Li and Guangrui Xie to join me to work on building the base model that takes in the shopping data to predict booking. And then Guangrui is taking the booking to augment our existing patient forecaster.
Michael Wu: So Jiabing there are so many third party data out there, everything ranging from weather to a competitor fair. Why did you choose to use shopping data to predict booking?
Jiabing Li: Obviously, people often search before they book. Shopping data contain the information about people intend for future travels since shopping activities happen several weeks or even months ago before booking. Shopping data could provide a strong predictive signal to predict future booking trends.
Michael Wu: Now, this sounds so obvious, but the fact that this hasn't already been done suggests that this probably is something pretty difficult. So what are some of the machine learning challenges that you're encountering using shopping data?
Jiabing Li: Well, the shopping data is also very noisy and hard to work with in machine learning. The huge number and the variety of channels has different consumer coverage, and many of them are overlapping. The data sizes are huge because people can shop for many different itineraries for a long time before the mix of final booking. Finally, there is also huge variation in shopping activity across individuals. Some people, when they shop for few days before booking. Where others, can shop for several months or even set up scripts to automatically search and monitor the price.
Michael Wu: So how are you overcoming these challenges? This sounds like a pretty hard problem.
Jiabing Li: We use the multi-stage modeling method. We essentially try to predict the booking data in many stages using increasingly complex models, and each stage is a repeating of error of the previous stage. So the error would be decreased continuously from one stage to the next. The method is a glass box approach. It's a way to offer more interpretability to complex models that use black box models.
Michael Wu: So can you tell me more about this glass box approach? It sounds like a pretty interesting approach.
Jiabing Li: It's a fairly new approach, which includes a series of interpretable machine learning models. Our method uses the interpretable model at the early stages as the error decreases, we will be using more complex models at the later stages, even if we have to use the black box model at the last stage. The entire chain of models is mostly interpretable until the last stage. But since the last stage will be feeding two very tiny errors. Even if the black box stage is uninterpretable, it will only contribute very little to the booking prediction.
Michael Wu: So how well can we predict booking using shopping data?
Jiabing Li: We tested this using data from transportation company and airlines. For transportation we are seeing average correlation coefficient of 0.5, but it can go as high as 0.96. For airlines the average correlation coefficient value is 0.4, but it also can go as high as 0.9 for some of these.
Michael Wu: So now we have a predictive model that uses shopping data. But this really is just the first step because even though the shopping data may be highly predictive of booking, but maybe our Bayesian forecasts are so good that the shopping model is not bringing any additional benefit. So next, we must really quantify the error reduction when we actually using shopping data. So Guangrui, how are you using Jiabing’s model to augment existing booking forecasts?
Guangrui Xie: One of the approaches that we decided to try first is the ensemble idea. Here the ensemble model means that we combine the predictions from the shopping model with the predictions from the basic forecaster. This is definitely not the only approach that we can use to achieve the goal, but this is the one that we decided to try first and have got some promising results.
Michael Wu: So the Bayesian forecaster actually has very granular dimensions such as departure time, fair class, point of sale. Now, when people actually are searching for flights, they don't typically specify this granular dimensions, so these dimensions just simply don't exist in the shopping data. So obviously, the model will not have these dimensions. So the question is, how can you merge and combine these predictions when they don't even have the same dimension?
Guangrui Xie: Well, we managed to unify the focus dimension by redistributing the predictive bookings from the shopping model across all the missing dimensions based on the proportions of the predictive bookings from the basic forecaster. And this method proved to work well for our tests.
Michael Wu: So how much improvement can we get from using shopping data?
Guangrui Xie: I do have some exciting results to share. If you use a test period that includes a COVID timeframe, the assembly model gives a 43% of error reduction on average as compared to the basic forecaster.
Michael Wu: That is a pretty impressive result. Now you tested this during the COVID pandemic, right? So could this huge kind of performance live be a result of the fact that the Bayesian forecaster may not be doing as well as it could have during the pandemic?
Guangrui Xie: Well, if I use a test period that doesn't include COVID time frame, then there are some other cases 26% of error reduction on average as compared to the basic forecaster. So you can say that the shopping data is useful not only in the disruptions but also in normal times. But we expect when shopping data brings us more benefits in disruption than COVID.
Michael Wu: That is pretty amazing. Now besides augmenting booking forecasts, how else can we use shopping data?
Guangrui Xie: Well, in the Bayesian forecaster, there is a special model called the holiday and specialty model, and we use this model to adjust the booking predictions. We may have a booking peak due to a particular holiday or special event. With the shopping data, we can potentially replace this model entirely because when there is a spike in the shopping activity for a particular departure date, then we are going to anticipate there is an increased demand on that particular departure date. And this information itself is enough to tell the forecaster to adjust the booking prediction level. And it really doesn't matter what holiday or special event causes that spike. So as you can see, in the shopping data it a more natural way to deal with booking peaks that are due to your holidays or special events.
Michael Wu: Now, for the research that leverages events data, I would like to invite Wenshen Song to answer a few questions for this very early stage research. So Wenshen why do you think events data can improve booking forecasts?
Wenshen Song: Because events can help improve booking forecasts because events drive people's shopping behavior and their shopping behavior, where ultimately lead to changes in booking. Some events, such as holidays, or conferences and sports. They will increase the shopping for tickets, while other events, such as extreme weather, will have negative impacts on shopping. In fact, COVID is just like another global scale event. This event will lead to a series of other events, and those events will greatly influence people's shopping behavior.
Michael Wu: So what precisely are you doing with these events data? We're using the events to predict the shopping activities. The main reason is that most events are in early on, are known far in advance. For example, we have repetitive events such as holidays. We have events that were announced early on, such as sports or conferences or concerts, even extreme weather or even lockdown for covid, where we predicted or announced days or weeks before the event actually happened. So what we're doing here is, for announced events in the future. We predict its impact on shopping by looking at how similar events in the past. Driving shopping activities.
Michael Wu: Wait, so you're trying to use events to improve booking forecasts, right? So why are you using events to predict shopping activity? You know, why not just predict booking directly?
Wenshen Song: Actually, we can do that. But based on the chain of causality event data. Event data will have a more direct impact on shopping and then the shopping, we're determining the booking. Since Jiabing and Guangrui has already built a mode of using shopping data to predict the booking. Events data probably will not contribute too much to improving the booking forecast. This is because the impact of the events has already been reflected in people's shopping activities. If we don't have a shopping model, then the events data can definitely improve bookings. But as things we have already got the shopping data. We should predict the shopping using events and then we're predictive booking using the shopping data.
Michael Wu: That's very interesting. So it sounds like the real advantage of using events is that you can actually predict shopping activity before shopping actually happens. And then you can use the shopping activity to forecast demand at a much, much earlier time because most events are planned so far in advance. So it seems like this could even have benefits beyond revenue management, even in network planning, for example. So Wenshen clearly, these events data are extremely useful. So are there any challenges in using event data in machine learning?
Wenshen Song: Well, that's a very good question. Leveraging event data is actually very, very challenging, right? You have huge temporal and spatial overlapping and numerous other features which are not easy to deal with. So, for example, at any given time in a location, we could have thousands of events happening in different impacts, attendance by geographical space, et cetera events data can also change very dynamically. Right? For example, announced events could get canceled or postponed or updated at any given time. As a result, predicting shopping using event data, it's not an easy task.
Michael Wu: Now, despite these challenges, Wenshen is able to get some pretty promising results, we tested his model with about a year of data from a US based transportation company in the 48 states that it served, and we were able to achieve a predictive correlation coefficient about 0.6, and it can actually go as high as 0.72, which actually is fairly good for some states. So that's our innovation on the airline side by leveraging third party data. So now let's switch gears and take a look at the innovations on the B2B side. And for this, I'd like to invite Yan Xu and Jonas Rauch to talk about wROXANNe now, wROXANNe, because it's actually inspired by revenue management's forecast and optimize approach. It actually has two parts. So, Yan. Work on the first part, and that's the price prediction part where we are trying to predict the willingness to pay price and then Jonas actually works on subsequently the optimization part that generates the recommended price. So first of all, Yan, can you tell us what wROXANNe stands for?
Yan Xu: Well, the it consists of three parts and the “wro” stands for win rate optimized and “xa” stands for explainable AI and “ne” stands for the neural network.
Michael Wu: So wROXANNe takes a pretty novel and unconventional approach to predicting willingness to pay. Can you tell us what is the most significant change in this willingness to pay estimation process?
Yan Xu: Well, the biggest change is removing segmentation and goal segmenting. In may sound a bit surprising, but our research shows that removing segmentation can actually help improve the accuracy of customer when to pay estimation. Segments may also change dramatically due to, say, environment or market change, such as the COVID. It may become difficult to manage and without segmentation, we can really aim for the target to make the price prediction more accurately.
Michael Wu: So, you know, segment free is interesting. So, but what is the advantage?
Yan Xu: From a machine learning perspective, it utilizes that data pattern for more accurate price prediction. Previously, we segment the data into micro segments first and then estimate the customer willingness to pay or segment willingness to pay within each segment. This means that we can only use the data within each micro-segments to do the estimates. What we are doing now is we are directly predicting the prices. And this allows us to use the entire data set.
Michael Wu: Now this sounds like a much more challenging problem. So when this require more data, so what is the data requirement for wROXANNe?
Yan Xu: The data requirement is still the same. We still need the transaction table, customer table and product table and also the minimum data requirement with just the SKU I.D., say customer ID, date and price. And furthermore, since we don't have any sparcity issues and we can take in more attributes, there's really no limitation on how many attributes we can use or what attributes we can use. So even though this is a much more challenging problem and we can actually leverage more variety of adjectives to solve it.
Michael Wu: Now from the name wROXANNe, we know there must be a neural network in there somewhere, and you're using it to predict this willingness to pay price, right? So can you tell us why do you choose neural network?
Yan Xu: And actually, we have tested a variety of methods with data sets from different industries. It turns out the neural network gives us the best prediction accuracy, and on average, it can improve the accuracy by more than 10% and it also provides a very flexible framework to incorporate different types of attributes. For example, previously, it's very challenging to utilize the attribute with lots of values, such as zip code. We need to pre-group them first before putting them into the model well with the neural network. You can directly take that look.
Michael Wu: Now, that seems like a pretty interesting way to deal with sparseness, but deep neural network is essentially a black box, right? So how can we interpret and understand what this model is actually doing?
Yan Xu: We totally understand that the interpretation is the key to adoption. And we always have it in mind when we design the algorithm. And actually, there are many different ways of how we can investigate the model, and we can do that at different levels. For example, at the global level, we can generate feature importance to understand which feature or attribute contributes most to the prediction. At individual level we can explain the output from neural network with pricing model for like starting from the average price. How it adds up to the predicted price with positive or negative impact from each attribute. So we always keep model interpretation as our highest priority exposure.
Michael Wu: Great. So now let's move on to the optimization part of wROXANNe. So Jonas, what's the major improvement in the way that wROXANNe generates its recommended prices.
Jonas Rauch: But another main difference is going to be how we come up with the recommended price. Previous versions of guidance use a rule based approach, and now we can do true price optimization. And to be able to do that, we estimate a statistical model that predicts the purchase probability as a function of the offered price, and we can then use it predicted win rate curve to find the price that maximizes revenue or profit or some other defined objective function. And especially in the case of profit optimization, that means that the recommended price will depend on the input cost.
Michael Wu: This forecast and optimized approach is not new, and it's been using RM for many decades. So why haven't we done this before? And I guess the real question is, what changed recently have enabled us to do this model based approach?
Jonas Rauch: We know that estimating price elasticity is extremely challenging, especially given sparse data and two recent developments are making our lives a little bit easier now. One is that we just have more data available, and that's driven by the shift to digital sales channels, where it's a lot easier to capture the kind of data we need. This shift has been accelerated by the pandemic, and secondly, recent advances in machine learning methods really allow us to use the data. We do have a lot more efficiently than we used to in the past.
Michael Wu: So now we are using this win rate curve to come up with the optimal price, right? But this win rate curve itself, a model that's estimated from data that's potentially noisy, right? So when an error in the win rate estimation compound and make the price recommendation less optimal or less accurate?
Jonas Rauch: Yes, obviously no more or less perfect and a bad prediction for the window, but it could lead to a suboptimal recommended price. Now to account for that and to work around that, we decided to use a Bayesian model, which means that our model estimates its own uncertainty. And that means that the model is able to predict not just the mean win rate cost, but also a confidence interval around it. And if that confidence interval is very large, then we know that the model is very uncertain about its prediction, and we can use that as a trigger for a more thorough manual analysis or a user intervention. So that means where we cannot expect every recommendation to be perfect. We have an alerting mechanism in place that tells the user in which case more care is necessary to ensure that we don't make a bad decision.
Michael Wu: I know that estimating this win rate curve, it's a pretty challenging problem and probably require a lot of data. Will company have enough data to estimate this accurately.
Jonas Rauch: To accurately estimate a win probability model it's critical to not only have data about successful sales transactions, but also to have data about offers that were made and that did not end up resulting in a sale. This so-called loss information is critical because it allows our model to observe how a change in price will affect the expected win probability.
Michael Wu: But most of our customers are not tracking lost data yet, so we will wROXANNe still work for these customers?
Jonas Rauch: We know that many of our customers do not have lost data yet or might just be starting to collect it. So we made sure to also have a version of the model that is able to work with only the kind of data required by previous versions of Guidance and specifically this product information, customer information and then sales transactions. And this model obviously has to make additional assumptions. And it's going to be slightly less accurate, but it retains a lot of its advantages and that it's still able to predict confidence intervals. And in the optimization process, we are still able to react to a cost.
Michael Wu: Besides this price optimization, are there any other use cases for this predicted win rate since we work so hard modeling it? You know I'm curious what other use cases there are?
Jonas Rauch: Yes, and I just made two of them. The first one is estimating ROI. And the second one is balancing multiple objectives. And by that, I mean that a company might sometimes want to slightly deviate from a revenue or profit objective to achieve some secondary goal. For example, increasing market share or increasing conversion rate. And given the predicted win rate curve with our model, we can visualize the trade off between these different objectives and answer questions such as how much revenue would we have to give up in order to improve our conversion rate by let's say 1%.
Michael Wu: So there's wROXANNe, from willingness to pay price prediction to win rate optimize price recommendation. Now, if you're interested to learn more about wROXANNe, there is actually a deep dive session later on today and it will have a demo in it. So please check it out if you're interested. Last but not least, I'd like to invite Ezgi Eren and Kyle Schoener to join me to talk about DiANNe. Ezgi, work on developing this very innovative approach to do revenue management, and Kyle is actually applying this innovation to some of our customers in the travel industry. So Ezgi, first thing's first. Please tell us what DiANNe is and why do you call it DiANNe?
Ezgi Eren: DiANNe actually stands for direct adaptive neural network based revenue management. And what do we mean by all that? DiANNe is direct as it skips demand forecasting, and it generates prescriptions directly from historical data. It is also adaptive as it has that inherent ability to adjust its recommendations to shocks and shifts in demand patterns. And last but not least, then utilizes a neural network model as it is a predictive outwards, which makes it really flexible about the data sources and features used in prediction. So the name pretty much captures the main components and the strengths of this new normal approach to revenue management.
Michael Wu: So you work on DiANNe since the beginning of the pandemic, so how does that actually address some of the challenge created by the pandemic? If you think about it, demand forecasting has always been a challenge even prior to the pandemic for certain industries. We can talk about air cargo is a good example of that. Demand faced by the cargo industry is extremely late on lumpy. The bookings are really close to the departure date. They typically also come in as really large chunks of shipments. So this and other types of high levels of demand have become more pronounced since the start of the pandemic for many more industries. As a result, no one would be opposed to the fact that you really need more adaptive approaches. You also need less rigidity in data requirements and their strength to address both challenges. DiANNe is designed to work flexibly with data. It utilizes more data when available, but it can also work with less when needed. In terms of adaptability, it is quite responsive to the changes in the data due to the reactive nature. We set up simulations where we tested Diane's response to significant demand shocks and significant changes in demand patterns, and it has proven to be quite robust under various scenarios. Even the extreme ones where we simply cut the expected demand in half.
Michael Wu: DiANNe is a very novel approach to data revenue management. How is it different from the traditional RM?
Ezgi Eren: There are many aspects of DiANNe that distinguish it from a more conventional revenue management approach. It does not rely on a demand forecast. It does not require an optimization server, as is typically a large implementation with a quicker time to value. That makes it also possible by the industries that are less mature in their use of revenue management. It makes it useful for the cases where data quality and availability are questionable.
Michael Wu: So Ezgi, correct me if I'm wrong, I know it feels like that DiANNe is almost a light version of this traditional RM. So my question is, what are you giving up? So is the benefit and value you provide also lighter? Can you comment on this?
Ezgi Eren: DiANNe comes with great flexibility. Yet it is powerful. Using a neural network model for its predictions enables it to slice as many data sources as available. A few examples would be outside economic indicators, competitive information, or weather forecasts. Pretty much whatever is applicable to make predictions more powerful, and Diane is really adaptive, which is really critical under current circumstances.
Michael Wu: That is pretty amazing. It feels like we're getting something for nothing. So how can company use DiANNe and what's the data requirement for this technology?
Ezgi Eren: In addition to lighter implementation and lighter assumptions, DiANNe also has lighter requirements on the data. It is a data driven approach. It will require data for sure, and in particular, it requires historical transaction or booking data with pricing information included in it. But that's pretty much it. It will not require a vast amount of data, such as historical controls and availability, as is typically the case for a more conventional revenue management approach. And I must add that conventional revenue management works great when all that is available and usable, but sometimes that's a big challenge as it is right now. That's why we innovated DiANNe and making it a priority to keep the data requirements at a minimum.
Michael Wu: OK, so let's move on to Kyle. So, Kyle, you've been working on introducing DiANNe to the transport and logistics industry, and one of the important characteristic of DiANNe is its directness. So can you tell us why this is so important for the T&L industry right now?
Kyle Schoener: So it's very common practice with them in trying to solve a revenue problem. To do some prediction and send that to our optimization, that forecasting approach is usually to forecast demand for a given product. And once we have that demand forecast, we feed that into our optimization, determine our controls. Traditional forecasting, like predictability, they like patterns that are consistent because then it can forecast those into the future. So like a consistent sinusoidal pattern or a consistent pattern, traditional forecast methods work very well for. But let's say we have a pattern going on. It suddenly drops off. Like in the case of covid, then our traditional forecasting methods can, they don't know what to do. So we wanted to take an approach to be able to get optimal controls that were forecast free, independent of forecasts, and be able to generate them.
Michael Wu: Great, so direct really means just skipping this forecasting step and coming up with this optimal control and directly. So now how does this cargo industry think about this, this direct approach?
Kyle Schoener: Generally, they're really excited about it. They really like the idea of forecast free revenue management system. In the cargo industry that demand doesn't follow that consistent pattern. In fact, demand comes rather close to departure and really ramps up prior to departure. PROS is written even some papers, and deemed this the lumpy and late problem. And so because their demand is lumpy and late and doesn't follow that consistent pattern, they like the idea of being able to generate controls through an optimization that is independent of forecasts. And so this allows the classic revenue management problem to break away that type of additional forecasting and for us to generate opportunity curves, opportunity cost curves by looking directly at the pricing data as opposed to forecasting demand. And so they're excited about it, and they really want to get their hands on it and really liked the idea of that approach.
Michael Wu: I'm so glad to hear that. So how can the output of DiANNe be used by the T&L industry?
Kyle Schoener: So we call from DiANNe, we get out an opportunity cost benefit that both varies by time and by capacity. Within the cargo industry, this can be used as what we call an inter-condition. Inter-condition is a minimum price threshold that must be met for booking to be accepted. So as we're moving through our booking curve, DiANNe can provide this opportunity cost curve that can be fed directly into the reservation system given the current time to departure and how much capacity is available. So it becomes very dynamic and we can change. The second way within the industry, the T&L industry, that this opportunity cost factor out of DiANNe can be used is in terms of the margin optimization or profit optimization. We call for profit we typically want to find the marginal optimal location that provides the best price from a margin perspective. And within a margin perspective, we need a cost. So where can we get that cost? We can get that cost from DiANNe.
Michael Wu: I see so very interesting. So it seems like, you know, just profit or margin optimization use case in the T&L industry is a natural place for us to kind of combine this pricing and RM. So how do you think we can merge these two things?
Kyle Schoener: I think this is a great question. I think this question helps be driven by particular business case that PROS trying to solve. And that business case is I have products that I'm trying to price, and those products have a capacity constraint associated with it, and that availability should probably have an impact on my pricing. And so I think with DiANNe in our Guidance application, we can look now to emerging revenue management and pricing together. So our DiANNe provides our opportunity, call our guidance measures, our willingness to pay. It takes when more state into consideration. And so now I'm taking some pricing aspect and the cost aspect I can bring them together into my profit optimization within my profit optimization. Now I can maximize margin by incorporating DiANNe’s cost and utilizing pricing from my guidance application to help determine what that dynamic price ultimately should be.
Michael Wu: So that's DiANNe, from research to application in the T&L industry. What an amazing set of innovation. I am so privileged to be able to work with such a talented team of brilliant scientists. Now, in case you haven't noticed, we are seeing a rare convergence of these innovations. Our B2B solution is starting to look more like our RM with forecasting and optimization, and our RM is also starting to look more like our B2B solutions too. Leveraging more variety of data sources and taking into account of consumers' willingness to pay. And now we know why. And the reason is because business will evolve their practice to take advantage, what is certain, what is predictable in their respective industry. So B2B pricing is taking advantage of the certainty of efficient supply chain and evolve and become willingness to pay estimation. While airlines, on the other hand, is taking advantage of people's predictable travel behavior and essentially become revenue management today. Now exploiting these certainties is actually a good thing in the smart thing to do. But the thing is that change and disruption will always come. And in fact, change is the only constancy in life, and we have known this for thousands of years since the time of the Greek philosophers. So when these disruptions happen, business can no longer leverage what they know for certain anymore. Yet they still need to innovate in order to stay competitive, and the innovation does develop under these conditions will behave strangely in the sense that they would tend to converge rather than diverge. And we can see this in innovation that we develop, too. Now, whether it is an airline or B2B, our scientists is simply trying to innovate our pricing solution to help our customer address the challenge that they're facing today. So this rare convergence may be completely accidental. But the result is that now we have a whole spectrum of pricing, technology and innovations. And what that means is that the remember what change may come in the future or how much your business is being disrupted, or even how little your business is being disrupted. We will have the pricing solution that is perfect for you.