Coinmama uses Pecan to foresee customer behavior

Use Case

Customer behavior predictions

To reduce chargebacks and identify fraudulent transactions, Coinmama needs reliable, accurate methods of analyzing its data. Using data to dig into customer behavior can be a complex, inefficient project — until you find the right tools.
coinmama-img
coinmama-logo

Industry: Cryptocurrency exchange

Company Size: Over 3.5 million customers in 200 countries with annual sales over $130 million

Solution: Fraud detection, payments review

Platform Use Case: Customer behavior predictions

Data Stack: Snowflake, Tableau

140 hours

Saving an estimated 140 hours per month of analysts’ time in the review process

1,000s

Generating accurate predictions for thousands of transactions per week

2/3

Reducing the number of transactions requiring manual review by two-thirds

What Coinmama Says About Pecan

We’ve reduced our rate of false positives by about two-thirds and are now manually reviewing a much smaller list of transactions. Pecan gives us the ability to prioritize that list as well, so we can tend to the riskiest customers first.
Tammy Rotem
Data Analyst

Challenge

Evolving from static, manual analytic methods

Coinmama initially used a custom-written, manually run SQL script to apply static business rules to new transactions. However, this method was limited to a set of variables and rules that were established based on past analyses.

In addition to making this process more efficient, Coinmama also wanted to be able to look for emerging patterns of risk within its data. Recognizing those new patterns would make it possible to update predictive models so that they would constantly reflect current trends in customer behavior.

Solution

Using Pecan automates and democratizes data access

Today, Coinmama’s data analysts create predictive models using Pecan’s platform. Every transaction passes through their Pecan models. Using the transaction and customer features known to be currently most relevant, the models score each transaction for how likely it is to be fraudulent.

Pecan passes those scores back into Coinmama’s Snowflake data warehouse. Finally, the transactions, predictions, and explanations for those predictions are displayed on a Tableau dashboard that’s used by the payments team and refreshed daily. Additionally, feature importance details reveal which customer and transaction characteristics are most meaningful in predicting customer behavior.

In addition to the predictions for the individual transactions, the data team shares the feature importance information from Pecan with the payments team via a Tableau dashboard. Because these analyses are available to everyone on both the data and payments teams, they can have cross-team discussions of notable data points and emerging trends, together building a more thorough understanding of customers and behavior patterns.

Results

Highly precise predictions speed reviews and dramatically reduce time and effort

Armed with the dashboard, the Coinmama payments team can make fast, data-driven decisions in their review of customer transactions. The most concerning transactions are prioritized with Pecan’s scores.

Only about a third of the transactions that previously would have been reviewed now require scrutiny, saving analysts on Coinmama’s payments team an average of 8 minutes per transaction. That saves the team 35 hours of work each week. Additionally, using predictive modeling, Coinmama can uncover 15% more fraudulent transactions than with hand-coded scripts, leading to better early detection techniques and mitigation strategies.

In more technical terms, the Coinmama data analysts refined their Pecan model to achieve an AUC of 0.92. The Pecan model’s precision is +0.42 higher at the 95% confidence level than the review team’s previous, rules-driven baseline model. Ultimately, the Pecan model achieved precision of 0.87 at the 99% confidence level. The Coinmama team continues to refine and update the model as conditions change.

Pecan’s predictive analytics platform, combined with the savvy of Coinmama’s analytics teams, has been a powerful force for saving time, reducing costs, democratizing internal data access and discussions about data, and gaining up-to-date, dynamic insights into customer behavior. The same kinds of predictive methods used by Coinmama can also be used by companies to determine which customers are likely to have the highest lifetime value, when a customer is likely to churn, or which customers will be more likely to respond to an upsell offer and what that offer should be.

Contents

Download the Full Study

Get the full details plus more data and results by downloading our PDF.

It’s time to refine your outreach with customer foresight