Maveryx Success Stories

Learn how Alteryx customers transform their organizations using data and analytics.

Showcase your achievements in the Maveryx Community by submitting a Success Story now!


Predicting Defectors with Alteryx Designer

Name: Rossella Melchiotti & Viktor Kazinec
Title: Data Scientist; Head of Analytics Delivery
Company: Close Brothers
Overview of Use Case

Close Brothers is a leading UK merchant banking group providing lending, deposit taking, wealth management services, and securities trading. In 2016, Close Brothers created their first Alteryx workflow. In 2018, they had 121 Alteryx users and unlimited licenses. Their marketing team uses Alteryx to predict defectors by running targeted marketing campaigns. The work is done on Alteryx from beginning to end, including data access, exploration and cleaning, feature generation, predictive modeling, and model comparison. The final model has been put into production as a day-to-day tool, and the accuracy of the predictive model is 80-90%.

Describe the business challenge or problem you needed to solve

The Marketing team wanted to predict defectors by running targeted marketing campaigns. First, they defined a defector as “a customer that doesn’t pick up a new deal within 24 months of current deal maturity.” 


They had 12 years’ worth of data and 4 different business units, and each one had between 4,000 to 10,000 agreements. Each deal had 30 or more variables, such as whether the customer is an early or regular payer, information about the agreement, term, amount, interest rate, and what assets they had financed.


Describe your working solution

Close Brothers tackled this project in 3 months. “We operate as an internal consultancy within Close Brothers. This project has been trusted to the BI team, and we connect our Alteryx workflows directly to our local database.  This is a classification model, so we can get a probability of customers defecting in the next 2 years. It outputs in a CSV in Excel and we send to the marketing agency to run the campaigns,” said Viktor Kazinec.


To build the model, they used 2/3 of training samples and 1/3 of validation samples. They achieved high accuracy rates for predicting defectors (84%-90%) and predicting loyalist (46% - 86%). “Another challenge was how this model would perform on current data. You are building a model in the end of 2017, but you are only using data from 2005 to 2015. So we performed an out of time testing. We used the training samples from 2005 to 2014 and validation from 2014 to 2015,” he said.


  1. Data cleaning: remove dirty data, duplicates and generate new variables. All the dataset is stored in a SQL server database





  1. Generate outcome: defectors were not a variable in the database





  1. Predicting future defection: train one model per branch because their customers are different – Decision Tree, Random Forest, Logistic Regression







Describe the benefits you have achieved

Campaign Results

  • Triple conversion rate
  • 20% higher average deal size


Model Maintenance

  • Easy to update every 6-12 months
  • Easy to adjust for changes in data, customer type or behavior

Related Resources