October 12, 2020 6 min read

Geospatial AI for Distressed Loan Analytics (Part 2)

Written By CAPE Analytics

In our previous post, How Distressed Loan Traders Are Leveraging Geospatial Imagery, we explored various imagery sources and how they are being used in the market today. With distress growing in the market, this post will explore how computer vision can scale the use of geospatial imagery and, in doing so, unlock insights into individual properties, as well as unique ties between property conditions, value, and loan analytics at an aggregate level. 

Capturing images manually, such as those in a Broker Price Opinion, is expensive, slow, and results in inconsistency. Reviewing images manually is time-consuming and leads to irregular outputs. These challenges reduce the utility of the information in large scale applications such as modeling or even for the efficient analysis of large pools. As volumes increase, these challenges only compound.  

Computer vision can help. Computer vision offers the ability to analyze properties in seconds instead of days, so evaluating images on large pools of mortgages becomes straightforward.  Computer vision offers consistent treatment of these images, so more powerful and predictive models are available. Computer vision offers fast, reliable analysis, encouraging the broader use of images — and allowing participants to catch outliers they would otherwise have missed and thus make more profitable decisions.  

Process for turning imagery into loan analytics

Transforming imagery into structured data is an involved process. At Cape Analytics, we begin by stitching together a coast-to-coast view of the United States from imagery provided by our imagery partners. Then, computer vision algorithms identify key elements in the imagery relevant to risk and valuation. Next additional data, ranging from parcel vectors to area hazards, are layered in to create a comprehensive view of the collateral/property. Finally, the outputs are returned as structured data linked to an address.   

Impact of geospatial property data

These property characteristics can greatly impact risk. As an example, let’s use the Roof Condition Rating (RCR) created by Cape Analytics. Cape’s RCR rates every roof on a five-point scale ranging from severe to excellent, as shown in the figure below.   

Above imagery provided by Nearmap


RCR has proven useful to insurance providers, where policies on properties with a bad roof had loss ratios three times higher than those on properties with good and excellent roofs.  RCR has also proven useful in loan analytics for predicting mortgage risk. In one study of 443,274 properties, mortgages on properties with roofs in severe condition were twice as likely to default as those on homes with roofs in excellent condition.  

Other variables prove useful as well. For example, using computer vision to identify trash littering a parcel (Cape’s “Yard Debris”) from overhead imagery has helped analysts flag risk on properties that drive-by inspections failed to identify. The detection of negative condition elements early in the process allows analysts to make more informed bids and focus diligence efforts where they are most impactful. Clearly, catching these issues before close improves returns.

Further, these risks change over time. Transforming images to structured data allows portfolio managers to more easily monitor those changes. Understanding this change retrospectively allows analysts to trace the path of millions of properties, useful in model fitting and market analysis. Understanding this change prospectively enables easy-to-manage change detection, flagging portfolio risks early.

Deploying property data into workflows

This conversion of images into structured data allows risk analysis at a massive scale.  With scalable cloud infrastructure, properties are analyzed quickly and results can be integrated into customer workflows. Data delivery is possible via API, batch files, or a customer web portal.  Often, combinations of delivery methods are used, where large loan portfolios are analyzed via API and a web portal is used to home in on an individual property and more fully comprehend the risk of previously unknown characteristics that could hurt investment returns. 

By changing the scale of what is possible, computer vision is increasing the capability of real estate and loan analytics across the sector.  This proves particularly useful in these uncertain times.  

If you would like to see computer vision in action to more fully understand the condition of your portfolio, please contact us about running a sample.


3 Ways Instant Property Condition Assessments Will Transform Real Estate Investment in 2021