A Peek into Machine Learning at Square

Heads up, we’ve moved! If you’d like to continue keeping up with the latest technical content from Square please visit us at our new home https://developer.squareup.com/blog

We’ll be posting more about specific projects, methods, and how we use machine learning in our day-to-day work — stay tuned! For now, here’s a glimpse into machine learning at Square.

Square has a large, unique data set that covers businesses across the U.S., Canada, Australia, Japan, and the UK. Machine learning is a key way we are taking a data-driven approach to automating both our internal tools and customer-facing experiences.

Square takes a new approach to supporting small business owners by starting from a position of trust. This is possible because of the strength of our machine learning models, which are at the core of our business and enable us to onboard individuals who might otherwise have been excluded from the economy. More than 90% of sellers self-onboard and are accepted — more than double what’s typical in the industry.

How do we decide when to use machine learning?

Our teams typically go through a development workflow similar to the following:

1. Define the customer problem. What are we trying to solve? Has someone already done it? Is it worth solving? What techniques are applicable? Is machine learning appropriate? Or necessary? Answering these questions ensures the analysis has actionable outcomes.
2. Collect and understand the data. When data doesn’t already exist for consumption, we work with engineering to implement the necessary instrumentation and collection. Once we have collected an initial sample, we validate its quality through data investigation. For existing data, we process and clean a sample, or the entire dataset, to use in building our model.
3. Build a model. Having the data and a clear problem statement, a model is chosen to fit the requirements. The model is evaluated on run time, interpretability, accuracy, and generalizability.
4. Iterate. There are always improvements that can be made to optimize our models. Whether through refactoring code, feature engineering, or implementing a different machine learning model, the team here at Square is constantly working on making the process cleaner, faster, more accurate, and more interpretable to our end consumers.

We have implemented machine learning in most areas of our business including risk management, marketing, sales, analytics, and customer support; and we are continuing to find more ways to integrate it. Below are some examples of where we are using machine learning to scale our business and improve customer experience.

Risk management and fraud detection

Back when Square was founded, as a brand-new company, we had very little data to work with so we started with simple heuristics, similar to a basic SQL query, plus manual review. Over time, with more data and experience, we were able to build out machine learning models that accurately detect risk and fraud automatically. This has allowed Square to be more precise, efficient, and scalable as we push for best-in-class risk management without negatively impacting customer experience. We’ve invested heavily in using machine learning to minimize the amount of manual work that is required to assess risk. This is something we track very closely and over the past six months, we have been able to reduce this manual work by 40%.

Square Capital business lending

Our real-time data-driven understanding of our sellers helps us reduce the risks associated with business lending and credit lines. Because we can look into a seller’s payment history, inventory movements, and hiring practices, we can quickly and easily determine who qualifies for a small business loan based on their ability to repay. We’ve been able to grow Square Capital loan volume 68% year-over-year, with an average loan size of $6,000. We’re not competing with financial institutions, but rather creating a new market where previously the only other option for a seller was asking for a loan from friends or family.

Predicting best product fit for sellers’ needs

We can also utilize Square data and machine learning to help our sellers find the best Square products for their business. We have a data science team devoted to using domain knowledge and feature engineering to predict a seller’s likelihood to use any given suite of Square product offerings, such as Invoices, Appointments, Gift Cards, Instant Deposit, Customer Engagement, Point of Sale, Employee Management, Loyalty, and Payroll. This prediction allows us to surface Square products to the right sellers at the right time — streamlining a seller’s search and potentially bringing new products to their attention that could help them grow their business.

Predictive Customer support

Square’s customer support team is constantly innovating on support tactics to provide fast, efficient answers to our sellers’ questions. Using machine learning models, we are able to draw connections and make recommendations dynamically and in a personalized way. We predict likely issues a customer is having, allowing us to proactively resolve their issues. This method intelligently balances both technical and human solutions to customer success — and gives them time back so they can focus on running their businesses.

Work on these machine learning challenges and more with us! square.com/careers/data