Kensington Mortgages slashes mortgage performance modeling HPC runtimes from 20 hours to 25 minutes with Azure Batch service

The company serves borrowers such as self-employed or small sole proprietors and those who make a high percentage of their income in the gig economy. Using Vector, its in-house mortgage performance engine, Kensington applies its own algorithms to analyze its rich dataset to identify opportunities while supporting the design of products and policies along with measuring the attributions of risks associated with its lending and servicing its portfolio of loans. This helps the company better understand buyer behavior, identify possible risks, assess potential losses, and take appropriate actions while also being able to forecast expected cash flow. 

“To create better products for the people who we’re trying to serve as well as improving our proposition as an originator and servicer of mortgages, we need a level of understanding that takes a lot of information, history, and knowledge—and that’s where all our data comes in handy,” says Amir Mehr, Portfolio Strategy, Credit and Analytics Director at Kensington Mortgages.

Data-driven decision making

Kensington created Vector more than a decade ago in a spreadsheet using logistic regression to process a limited number of variables for moderately sized portfolios before moving to more complex versions using .NET. However, the move to state-of-the-art machine learning algorithms boosted model performance and accuracy, and this led to exponentially increased computing requirements that Kensington’s on-premises infrastructure couldn’t meet.

With an average of 1,000 forecasts a year, the company’s infrastructure would require more than 20,000 hours to process its data. Kensington needed a solution that could process data much faster and reduce its average runtime to minutes instead of hours.

“We use Vector everywhere throughout the business,” says Eduardo Marcano, Senior Vice President, Portfolio Strategy, Credit and Analytics at Kensington Mortgages. “The tool provides real numbers, and we rely on it to help us make real decisions every day.”

Expanded computing capability

To achieve greater computing power, improve its models, and support a larger digital transformation, Kensington decided to migrate Vector to the Microsoft Azure cloud platform. “We chose to work with Microsoft because its solution aligned with our roadmap for the future—Azure has it all,” says Sughasini Ravikumar, Senior Analytics Solutions Architect at Kensington Mortgages. “Microsoft is more than just a vendor. The team came right through the journey with us.”

Kensington recognized Azure Batch as key to its solution. The company used the service to migrate Vector directly to Azure and maintain usability without having to rewrite its models and algorithms. Kensington also relies on Batch to schedule compute-intensive tasks and for the flexibility to distribute and scale tasks in the cloud to meet the company’s needs, which Kensington could not do with its static on-premises infrastructure. To handle that scaling, Batch creates and manages a pool of Azure Fsv2 virtual machines (VMs), installing the applications and scheduling the jobs that run on the nodes. The company doesn’t have to install and manage cluster or job scheduler software. Batch has a high-scale job scheduling engine at its core, which works behind the scenes to run jobs, note failures, requeue work, and scale back the pool based on parameters Kensington sets.

The company adopted Azure high-performance computing (HPC) to support its underlying infrastructure and run large-scale parallel processing, which offers significantly faster computations for more nodes with little overhead. “Using Azure HPC is a good strategy for us because we have the type of workflow that benefits from that massive parallel computation. The nature of our algorithms necessitates a lot of computation because of complex branching and our need for accuracy,” says Marcano. “We do billions of calculations for a given portfolio. We model each individual loan through its whole life, which is typically over 30 years.”

Parallel processing

Kensington runs forecasts on its multi-billion-pound mortgage portfolio up to 100 times a month. Prior to the Batch deployment, loans were randomly distributed during processing. In some cases, nodes finished processing in 10 minutes, but they took 45 minutes in others. Batch offered the performance optimizations, flexibility, and configurability Kensington needed to customize the way it distributes loan calculations. Through Batch, the company now orchestrates and distributes all the calculations in a parallel manner. “When you want to split your tasks individually, Azure Batch is completely the best fit,” says Ravikumar. “It serves that purpose very well.”

With a better understanding of its new computing complexity and supported by Azure HPC, Kensington developed a custom algorithm and designed a way to run it on Batch. This gave the company a platform to equally distribute the calculations and bring runtime down to around 10 minutes per node.

Now, Kensington can take advantage of Azure resources like machine learning operations to help widen the scope of the machine learning algorithms that feed into Vector. Over the next 18 months, Kensington plans to use Azure tooling and computing power to round out and improve its whole software development cycle with machine learning operations. “We really want to make the most of our Azure environment to develop new tools rather than maintaining existing tools manually,” says Marcano.

Improvements across the company

Kensington now has the additional computing power it needs to optimize Vector, and it has reduced timesteps and increased accuracy in its results, including its regulatory risk management calculations. The company slashed Vector’s total runtime from 20 hours to just 25 minutes—for 1.9 billion total calculations with every run. Kensington wants to reduce time even further moving forward. Its goal is just five minutes, and the company’s engineers are confident they will get there. 

“Everything is in one place,” says Ravikumar. “Performance optimization is so easy with Azure Batch and supporting Azure platform services.”

The company continues to evolve Vector and make greater use of Azure Batch and high-performance computing, enhancing processes companywide. With the stronger predictions and insights it gains from its optimized models and results, Kensington can refine its customer experiences and serve even more customers across the United Kingdom.

“We’ve used Azure Batch to build new capabilities that will help us identify niches and create better tailored products,” concludes Vicki Harris, Chief Commercial Officer at Kensington Mortgages. “We can make sure that our customers’ mortgages are better suited to their particular circumstances. Ultimately, that means better outcomes for them.”

Find out more about Kensington Mortgages on Twitter, Facebook, and LinkedIn.