Two Biggest Challenges to Implementing Dynamic Pricing

Many people think dynamic pricing is crazy–witness this 50x surge Uber once saw. At today's exchange rate, that 425 KR is $50 USD–per minute. But is customer reaction to this the biggest challenge? 

We're not going to sugar coat it: implementing dynamic pricing is hard. And, you really can screw it up. Anything with tremendous potential to improve your business has the same potential to ruin it: just look at StubHub's fee-bundling experiment, that lost them 30% of their revenue–and, worse, created competitors in the form of SeatGeek, VividSeats and more. They went from around 80% marketshare to around 50%, all because of a badly managed pricing experiment. So yes, you can screw it up. Big time. 

But that doesn't make dynamic pricing–or any pricing strategy change–a bad idea. It just means you have to be careful, and thoughtful about implementation.

In our experience the two biggest challenges people run into are, in order, managing organizational change, and building a successful algorithm (or series of algorithms). If you can master these, you'll have a higher liklihood of success.

Managing organizational change




"The common theme across all our executive searches is change management. Very few people can show they can successfully lead change in billion-dollar organizations, and those people are in tremendous demand." 

- Consultant, Russell Reynolds Associates

In today's dynamic world, managing change is in extreme demand, especially at the executive level. While manager-level people may find the inspiration, or a strategy group may do the homework, only with executive committment will dynamic pricing–or any pricing change–become reality. Because it reaches through out the organization.

Here are 5 common areas that experience significant organizational change with dynamic pricing.

  1. Marketing. Learning how to message when you can't quote a simple number that's always available and easy to put on a billboard can be tough. But somehow, every airline and hotel has figured this out. Uber marketed its way to a $50 billion valuation. Can your marketing team figure it out?
  2. Sales. Sales managers are used to trusting their gut. But making gut decisions takes time, requires meetings, and is impossible to track or assess with data. Trusting a new, core tool after many years of doing things a certain way is far from an easy discussion. To counter reluctance, it helps that the model can be trained on and proven with backtesting on their historical gut-driven pricing decisions. And, many studies (and our direct experience) has shown that putting more, predictive pricing power in the hands of front line sales makes for faster decisinos, more closed deals, more profitable deals and happier customers. In sum, bigger bonuses. Does your VP of Sales want a bigger bonus for herself and her team? 
  3. Product. For digital businesses, the power to change price creates tremendous product opportunities–that can seem like overwhelming challenges if they aren't carefully prioritized and roadmapped. Many comapnies want to get pricing perfect right out of the gates: instead, make (large) incremental improvements. Can you start with one category, market, or segment, put points on the board, then expand? Why do you have to go all in on day one?
  4. Finance. Usually, we see finance driving dynamic pricing projects because they feel the pain of having to raise dilutive capital, they see the lost revenue in the numbers, they see the missed forecasts, and they know there's money to be found. However, they aren't (typically) operators, so a great CFO or VP Finance needs to help the rest of the organization realize how core pricing can be to growth, success, and overall performance–especially juxtaposed against other, lesser ROI projects, like the latest social networking or ad targeting fads. Can your finance organization win champions over in their counterparts across your organization?
  5. Customers. When ‘t Vlaams Broodhuys, a bakery chain in Sweden, eliminated cash money, 90% of its customers didn't care. But a few threw their coins at the staff, angry they couldn't use them. Reactions to dynamic pricing are similar: complaints about "Surge pricing" haven't stopped Uber from doing an average of one million rides per day, up from 0 just 5 years ago.  Or Amazon from having 304 million customers (the equivalent of the entire United States population), buying over $100 billion in merchandise from it in 2015–up from 0 just 17 years ago. Can you get the messaging right, so customers–who already expect to pay the "right price"–take the change in stride? 

Organizational change is difficult. Most companies will not be able to handle it, and will never implement dynamic pricing. Fine. Just remember - 88% of the companies in the Fortune 500 in 1955 had dropped off by 2014. Just 59 years had wiped out nearly 90% (87.8% to be precise). The ones that were left–Boeing, Proctor & Gamble, IBM–are famous for their ability to adapt and evolve. 

Getting the algorithm right

Modeling demand, clustering users, clustering products, detecting surges, (and, especially challenging, distinguishing a real surge from something meaningless) and building the outputs into product are all really hard problems. They are, however, all problems software can solve. Let's break them down–bearing in mind that every company is different, and the details matter.

Modeling demand

For simple businesses with a few hundred products, a handful of price points, and one or two sales channels (say, a website and a small wholesale business), perhaps a linear regression in excel could proivde a useful forecast of demand. But such a company would have little use for dynamic pricing. 

Modeling demand in complex environments–100 hotel properties around the world each with 200 rooms, available 365 days per year–either requires extremely detailed, complex human-made models, or machine learning models. The advantage of a human-made model is that you can look at each coefficient and think about it really hard: "When it's sunny, it looks like demand goes up by 2%, so we should raise prices then". The advantage of a machine learning model is that it can handle millions of factors, far more than what a human could consider, and be far more accurate. Comparing this against intuition (why are prices flat when it's sunny?) can lead to great results. 

The model provides the estimate of demand for the time period, which is critical for deciding what price to charge. If you want the basic mathematical formula, check out the book by our advisor Robert Phillips, Price and Revenue Optimization

Clustering users

All successful pricing strategies–even dynamic pricing–need to involve an intrinsic user segmentation. This may be done implicitly or explicitly. Implicitly, the price segments the market without advance knowledge of why. For example, Uber drives the price up if there is too much demand, leaving only the price insensitive and business people not paying themselves, in the market, without respect to why. The model didn't say "we'll hit business users with a 3x higher price because they book from the airport in the rain. Instead it increased price to match very high demand at the airport in the rain; this happened to hit business users. 

For some, slightly less dynamic industries, clustering users explicitly is more important. Explicit clustering takes into account the user characteristics and uses this to influence pricing. The classic example is airline fares for business travelers. Airlines realized business travelers rarely stayed over a Saturday night, so by raising the fare for trips that did not include a Saturday night stay, they charge business travelers more.  

Segmentation can either be explicitly used or implicitly taken advantage of, depending on the approach. Because even user segments can change, we prefer models that discriminate implicitly. 

Clustering products

Product clustering is critical for two reasons. First, many products will lack enough individual traffic or data to make accurate decisions in a timely manner, so aggregating them into clusters will enable much, much more accurate and stable models at next to no loss in clarity. Second, users think in clusters of products when they buy, so adapting that perspective in modeling demand closely aligns with consumer choice. 

Picture a hotel. 200 rooms. 160 of which are standard rooms, 30 of which are superior rooms, 8 of which are suites, and 2 of which are named suites ("Presidential Suite"–though, having worked for a US President once, I can say with authority that when the president stays at a hotel he takes over entire floors, not just one room–but I digress). The 160 standard rooms, while similar, will have different demand characteristics: some have views, some are on high floors with less noise, some are next to the elevator, etc. And each room has, at most, 1 booking per day, so grouping similar rooms together is critical. This grouping may change with time: in certain seasons, for example, views may be more or less linked to demand. But when considering suites–especially the named suites–clustering is critical. Here, with such low volume of bookings on any given time period, borrowing data is critical. In the case of the named suites, even borrowing cross-property. After all, if it's Ski Week, and rich people are taking their kids out of school on vacation, won't that impact be seen across many hotel properties? 

Clustering products is critical in building an accurate demand model. Sure, you can do other things with clusters of products, however, because of the importance and nature of this specific use, it's best practice to have dedicated, separate clustering algorithms just for price optimization and dynamic pricing. 

Detecting surges

Things don't always go as planned. If demand isn't matching to forecast, it can happen fast–and in many industries, you can sell out, losing revenue and–especially–profit. Think of tickets. Hamilton, the popular Broadway play, sells out nearly instantly. To scalpers, according to the New York Times. The same tickets are listed for 10x face value on StubHub moments after they go on sale. Presumably, Hamilton has a decent profit margin on its tickets. Let's say they net $50 on a $100 ticket. When a scalper sells that ticket for $1,000... that's $900 in lost profit for Hamilton. Ouch. 

Of course, everyone knows Hamilton is in demand, and they could fix this if they cared about the $65 million the New York Times estimates they lose to scalpers. I assume they don't care. But you might, if it were your business, and detecting a surge in demand is critical in setting prices to an equilibrium level. 

When Uber launched I had coffee with Ryan Graves, their now VP Operations (then CEO). Surge pricing was designed explicitly to always, always have a car available. If you open the app, there will be a car. It might cost $20/mile, but it's there. In San Francisco, in 2010, this was revolutionary: when my wife and I went out to dinner, frequently we'd walk 10 blocks, or drive, because there was never an avaialble taxi. On a rainy night, $20/mile was better than $30 in parking. 

But when Luxe Valet launched, they did not do surge pricing. Frequently, they would run out of valets, parking spaces, or both. As a customer trying to rely on their service, it was infuriating. I needed to park at my meeting–I was already in my car. And now, I couldn't. 

Most importantly, detecting surges avoids much more complicated planning. For example, consider selling nationwide, and needing to increase prices when it rains. You would need weather data, in 5 to 15 minute increments, for every US location, to build a meaningful model of demand considering rain. Or, you could just detect if demand is surging (for whatever reason–maybe rain) and increase prices. Far simpler, far more effective. Like Uber, it may take you months (or years) to figure out if that surge was because of rain, or a San Francisco Giants game, or the San Francisco Giants game being rained out–but in the moment, your prices reflected demand, and all was well. 

Integrating it all into the product

The "human computer interface"–that is, your product–is where your customers encounter and interact with your pricing. To be successful, an integration needs to present the right price, with applicable context, and in some cases applicable interactivity. And, an integration needs to learn from every interaction. 

Our philosophy hinges on how important this product integration part is: it is by far the most critical, so to have the focus and care to develop it, test it, and get it right, a company needs to know the models will work and be right. So Perfect Price focuses on the model side of it, freeing up company resources to decide and implement the right customer interface.

There are some great examples. Airbnb has a red-to-green bar that is designed to prevent hosts from overcharging. Lyft calls its surge pricing "Prime Time" which makes sense because it always seems to hit right during rush hour, intuitively a prime time for taxis.  Airlines and online travel sites like Expedia now suggest other days with lower prices, creating a context and putting the consumer in charge: instead of being "gouged" they are choosing to pay more for the convenience of travelling on the desired day.  

The first step to overcoming challenges

Small steps are always easier than huge leaps. Good places to start include doing a study to determine, based on survey or other data, how much money you are losing by not having dynamic pricing (or some other advanced pricing strategy) in place. Maybe it isn't significant. 

If it is significant, your next step is a pilot on a limited–but statistically significant and representative–portion of your product or customers (segmentation depends heavily on your business). While the up front work involved in doing this will probably be near or equal to a full deployment, the organizational change will be much easier to manage.  

If you get the algorithm right, the prices will make sense–whether high or low–and, most importantly, even if they complain a little bit, people will buy. But if, like some online marketplaces, the algorithm is wrong or runs wild, you'll at best get bad data (thinking dynamic pricing isn't working when, in fact, its the algorithm) and at worst, like StubHub, hurt your business permanently.

And, because I cannot say it enough, in our experience organizational change is always the hardest part. The modeling–we can handle that, and maybe you can too. But without buy-in in all corners of an organization, failure is all but guaranteed. 


Related Posts