Machine Learning-Based Dynamic Pricing at Scale

Dynamic pricing is a pricing strategy where the price of a product or service changes flexibly in response to current market demands. It’s long been the norm in the hospitality and airline industries, but more and more industries have turned to these algorithmic approaches in order to increase their profits. It’s a complex problem that requires that a business be able to anticipate and adjust to all the factors that might affect the price of a product, over short time frames, making Machine Learning and Artificial Intelligence-based approaches ideal. What’s more, because there may be many products and/or many locations where the customer purchase behavior might change, the optimal approach might involve a large set of models customized by product or location.

In this article, we take a look at some of the challenges of dynamic pricing from an ML deployment perspective.

Dynamic Pricing as a Multi-Model Problem

In order to best set prices, pricing teams develop models to estimate price sensitivity: how changes in price change the rate at which prospective customers purchase a product or subscribe to a service. In general, higher prices return a higher per-unit profit, but lower overall sales volume. The goal is to find a price that balances per-unit profit and sales volume to achieve optimal revenue.

Customer price sensitivity can be related to a number of factors, such as:

• customer segment
• product type
• scarcity/availability of the product or service
• the competitive landscape

and more.

Sales volume is also driven by how many people see the product; this can be affected by:

• seasonal demand
• environmental factors that help drive or decrease demand
• promotions and marketing campaigns

and other factors.

In dynamic pricing, the goal is to estimate all the factors that affect supply and demand and optimize the price accordingly — in relatively short time frames, sometimes even at the pre-purchase level. Estimating every one of those factors can be complex, and in order to determine pricing strategy even for a single product, a pricing team may have to develop multiple models:

• elasticity, or price sensitivity models
• supply and demand forecast models
• customer segmentation models
• website traffic models

and more.

In a dynamic pricing scenario, all these models, working in tandem, algorithmically determine the prices that consumers will see.

Each type of model may be the responsibility of different members of the team; the models may have diverse retraining schedules. Because a model may depend on other models in the pricing stack, a change to one model can potentially affect others.

For example, a change to a customer segmentation model, even if it’s an improvement, can change the prices that a repeat customer sees. If that price is higher, there is a risk of alienating or driving away that customer. So the effect of the new customer segmentation model has to be evaluated in the context of the entire pricing process. This may result in changes to other models as well; for example, the price sensitivity model may have to be retrained using the new customer segmentation model.

All of these changes may need to be A/B tested, as well.

To make things more complicated, customer behavior and price sensitivity — and therefore pricing strategy — can vary not only across products, but across sales regions, or from store to store in the case of physical commerce. This means that, depending on the domain, pricing may require models customized to different levels of granularity: regional, seasonal, locational, or even at the product level.

All of this leads to a plethora of models, working jointly to set prices. When a business has to consider dozens of regions, hundreds of stores, and hundreds or even thousands of products, this can quickly lead to thousands of models. The resulting complexity can overwhelm the data scientists and the pricing team who have to coordinate the evaluation and deployment of these models and monitor their performance in production.

For example, a nationwide real-estate company came to Wallaroo looking for a way to improve its pricing processes. They wanted to optimize their pricing based on region, location, product type, and sales channel. This required several models per region and channel, across hundreds of regions. These models need to be refreshed on a weekly basis. The complexity of developing, managing, evaluating and eventually deploying these pricing models threatened to overwhelm their relatively small data science team.

Using Wallaroo, the team was able to reduce the time to deploy their prototype models to less than a quarter of the time it took before, with substantially fewer compute resources. With the ability to quickly refresh and redeploy their pricing models, as well as monitor their performance on an ongoing basis, the data science team now has more time to focus on improving their models, and the overall pricing process.

Wallaroo for Managing Complex ML Processes

The Wallaroo platform can alleviate the pain of managing thousands of models in production. Our self-service deployment toolkit provides an easy-to-use SDK, graphical UI, and API to facilitate collaboration amongst all the members of the pricing team.

Streamline team collaboration: With the SDK and GUI, team members can check the status of their own deployed models, as well as those of their teammates. This facilitates coordination when multiple models must work together in tandem to put pricing processes into production.

Organize complex ML processes: Data scientists can chain together multiple interacting models and associated processing steps in a pipeline via the SDK. Pipelines can then be deployed as a unit, again via the SDK, simplifying the transition into production. If models or other components in a pipeline need to be updated: no problem! Pipelines painlessly support switching compatible components in and out.

Monitor Models in Production: Wallaroo’s basic data and output validation checks that the data going into a model and the predictions that come out look valid and in range. In addition, our Model Insights feature provides advanced data drift detection, providing insight into the environment that the models are operating in, and how it might be changing. With Model Insights, data scientists can potentially catch changes in the environment that necessitate model retraining, or other interventions, in a timely manner.

With dynamic pricing, businesses can implement pricing strategies that optimize revenue, while responding quickly to market fluctuations and consumer demand. Wallaroo’s last-mile platform reduces the effort involved in deploying and managing complex pricing processes in production, leaving the pricing team more time to focus on improving and enhancing their pricing strategies.

Want to Hear More?

Interested in knowing more about how Wallaroo can help you streamline your dynamic pricing processes? Email us at deployML@wallaroo.ai