Determining what price to charge for your product or service can at times be deceptively easy: figure out what your competition is doing and either match or beat that price. This approach works fine for commodity markets, where price transparency and comparison is easy. But what happens when there aren’t readily available comparisons? Or when you have “similar” product characteristics, but no other information that you can access that may have influenced pricing decisions such as other services included, seasonality, location, etc. In this case, you often go with a gut feeling or accumulated experience to price products and hope there is adequate demand at your chosen price level.
For most organizations, even if they did have data on all the product characteristics and external factors they would still struggle to process this information in a way informs day-to-day pricing decisions. The task of combing through and analyzing large data sets, determining correlations and assigning weighting factors to various product characteristics and other variables, is still beyond the capabilities of most organization’s technology stack designed for managing supply chains and customers, not high-powered analytical analysis.
Machine learning technology is starting to fill this gap, and traditional companies and startups are changing how pricing is done using smart analytics, processing power, and human intuition to optimize pricing. Let’s take a look a couple of real world applications in the insurance and hospitality industies.
Insurance Industry Application
As one of the largest insurers in the world, AXA has massive amounts of data on customer claim histories, and they are putting this to good use to help prevent large loss claims. Every year, 7%-10% of the company’s customers cause an accident, with most involving small claims of hundreds or thousands of dollars.
However, approximately 1% of customers involved “large-loss” cases of over $10,000 and Axa needed a better way to predict and hence prevent the number and size of the large-loss cases. They had been using a more traditional machine learning technique called Random Forest but were only getting prediction accuracy rates of less than 40%. In the hopes of getting better results, they started using Google’s TensorFlow deep learning solution and saw their prediction accuracy climb to 78%. They were able to do this by tapping into the advanced neural network model that Google had been refining over the years, and combining this with the scale of their cloud offerings to deliver the computing power necessary to handle the processing load. Axa is now in a position to accurately price risk based on better understanding of the attributes of policyholders and other factors that lead to large-loss cases.
AirBnB’s Pricing Algorithm
Airbnb’s pricing challenge is a bit more complicated than most, as the users, not the company are responsible for setting prices. To enable hosts with pricing decisions, the company needed to provide the tools and data to help them optimize the price received while maintaining occupancy levels. While conducting user research, Airbnb observed that during the initial sign up process, when hosts came to the pricing page they immediately began to search for other similar properties. The problem was that not only was this a laborious and time-consuming process, but they often had trouble locating similar properties. They discovered what most people learn when trying to sell their home, that it’s tough to find exact comparable properties. They also had to contend with pricing comps across an entire city, spanning multiple neighborhoods. They needed a way to automate this analysis and provide meaning price guidance to hosts.
So the technical team at Airbnb set their sights on solving two problems: 1. Automate the property comparison process, and 2. Understand supply and demand dynamics to make timely price adjustments.
Unlike eBay, where there aren’t any location or time dependencies — you can buy and sell anything from anywhere at any time — lodging is very location and date-specific. And in the Airbnb model, can be as varied and idiosyncratic as the people that own the properties. To solve for this, Airbnb developed a list of the prime characteristics of properties, applied weightings to each one based on their importance to potential renters, and then ran these assumptions against years of transaction data to model against actual outcomes (i.e. what was the final price).
They were looking for how each variable correlated with price to understand the key drivers of value and to inform their pricing engine to make better pricing recommendations. They discovered things like the number of ratings correlated with higher demand, and that the use of certain types of photos translated to higher prices. Surprisingly, the professional photos of living rooms didn’t fare as well as the nice cozy bedroom shots taken by the owner.
With these new insights, Airbnb was able to provide a more useful pricing tool for hosts that not only allowed them to price their properties based on more comprehensive comparative analysis but also provide dynamic pricing recommendations in response to changing demand. Similar to how airlines handle pricing, hosts get ongoing guidance based on market conditions so they can make adjustments the will drive higher occupancy.
Machine learning augments human decisions by narrowing a set of choices. But just running a “black box” in the background that produces the miraculous answer is not sufficient. In the above example, insurance agents need to be able to explain the rationale behind auto premium price differences and rental hosts need to understand why and how price recommendations were determined to maintain trust and confidence in the information. It’s important to keep in mind that while the machine learns and provides answers, humans still need to explain what it means and why the results should be trusted. The product lead for Airbnb put it best, “We wanted to build an easy-to-use tool to feed hosts information that is helpful as they decide what to charge for their spaces while making the reasons for its pricing tips clear.”