Pricing Team has some awesomely smart data scientists. We put one of their machine learning models behind a RESTful service with a minimalistic UI -- now anyone can play around with their models!
What it does
The Predictr service exposes two different pricing models: Elasta and Demand Kinks, both model demand for a product as a function of the product's price. Now anyone can upload their own pricing strategy and the service will leverage the models to predict the effect of your pricing strategy on VCD and Revenue.
How we built it
We got all the necessary formulas from our data science team, we collaborated with each other to organize the service endpoints, sql code, and business logic into an architecture that made some sense, and then we coded it up!
Challenges we ran into
These models were really complicated: long formulas with complicated inputs. We spent a lot of time debugging weird results.
Accomplishments that we're proud of
We made the hard work of data scientists available to everyone at Wayfair! That's awesome!
What we learned
How to stand up services, how pricing strategies work, and a lot about pricing sql tables.
Who is the intended audience for this project?
Pricing and business analysts as well as data scientists conducting pricing experiments who wish to investigate the effect of a proposed pricing strategy on VCD and Revenue.
What problem is this trying to solve?
Currently, Pricing Analysts and Catalog Management leverage a model known as Constant Elasticity to predict the effect of proposed pricing strategies on business metrics. The Elasta team has recently produced a new model, Demand Kinks, which introduces a number of corrections and improvements to Constant Elasticity at the cost of being far more complex. As a result, it has become difficult for analysts to leverage Demand Kinks when analyzing the effect of pricing strategies on business metrics.
What is the proposed solution?
We stood up a Flask service with a user-friendly UI that applies the Constant Elasticity and Demand Kinks models to a proposed pricing strategy. The service accepts a set of new prices for a set of products and computes the expected changes in demand, revenue, and VCD. The service also exposes a RESTful API endpoint for data scientists and developers to query programatically.
How does this solution benefit the intended audience (and anyone else)? What is the impact?
This solution provides a simple and intuitive way for pricing analysts to iterate through a variety of pricing strategies and get valuable quantitative metrics of the expected effect of those strategies.
What's next for Predictr
Dockerization and horizontal scalability. We would love to containerize our application so that we can deploy it to Kubernetes and harness their awesome compute power.