2020 has been a year of massive uncertainty. As of this writing, there are more than 6.8 million COVID cases in the United States. U.S. unemployment remains elevated at 8.4%, and 3.8 million borrowers are in forbearance. A large number of forbearance plans are set to expire in late September or early October, and the market is watching closely to see which path borrowers follow after expiration: will they extend forbearance, move to current or fall into some stage of delinquency? The answers will significantly affect the magnitude and timing of the non-performance we face in the mortgage market.
Regardless, the market is preparing for a substantial uptick in activity, and traders are looking for ways to improve their ability to make buy and sell decisions quickly while choosing the right disposition path for each loan. In general, the greater the credit risk of a mortgage, the more critical the collateral value behind a loan. In looking at non-performing assets specifically, property condition is a primary consideration traders should take into account. And to better understand property condition, we’ve seen a growing reliance on geospatial imagery.
For example, as analysts try to understand property value before bidding on loan portfolios, they will often review satellite images from Google Maps and run a “Zestimate” from Zillow with its accompanying photos. The basic assumption of most automated valuation models, including Zestimate, is that the property is in average condition and analysts spend a significant amount of time looking at imagery in an attempt to discover the outliers. This continues after having won the bid, investors’ diligence teams may rely on valuation providers to deliver images from various street-level angles, which are often incorporated into an opinion of value. Additionally, diligence analysts scour the images these reports contain to glean new information to help further assess a property’s condition.
To better understand the evolving uses of imagery by analysts, let’s start by examining the types of imagery used today. It turns out, not all imagery sources are equal, and different image capture methods come with distinct advantages and disadvantages. In addition, the manual review of imagery by analysts comes with its own set of problems — problems solved by artificial intelligence.
Geospatial Imagery Sources For Distressed Mortgage Trading
It is useful to categorize exterior imagery by capture method. On one end of the continuum, there are high coverage, lower-resolution imagery sources such as satellites, and on the other end are lower coverage or low latency, high-resolution imagery sources like those offered in drive-by inspections.
Satellite imagery offers great coverage, offering overhead views of the entire United States and, in fact, the globe.
In the distressed mortgage market, we often hear people using tools such as Google or Bing or Zillow to get eyes on a property. These sources utilize multiple imagery sources, including satellite imagery. Analysts who access commercially available satellite imagery are able to understand the layout of a neighborhood and parcel from these images. An analyst may also be able to key in on other critical variables — for example, is the property next to a landfill? However, a fundamental trade-off is image recency, as the images can be out of date and the update schedule can be inconsistent. In the images shown below, the image on the left is over four years old, and the image on the right is 18 months old. The difference is striking: in the first image, none of the 16 homes in the new development have been built yet.
Move the slider to see the difference:
Commercial satellite providers are reasonably priced and offer a top resolution of one square foot per pixel (with 30 cm per side per pixel, this is known as 30 cm Ground Sampling Distance, or GSD, but is colloquially referred to as 30 cm resolution). Typical resolution ranges from 50 to 80 cm resolution.
Aerial imagery is another important source of overhead images. Multiple commercial aerial imagery providers rotate their fleets of airplanes around the United States throughout the year. Aerial imagery providers visit highly populated areas every few months, so the images offer good data currency at a similarly reasonable price.
Move the slider to see the difference:
A key advantage of aerial imagery is resolution. While the leading satellite imagery providers offer 30 cm resolution, aerial providers step this resolution up to 7 cm or better. The images shown above are examples of the difference in resolution between satellite and aerial. An analyst can identify debris in both images, but the aerial photo allows for the identification of critical details, such as debris type, driveway cracks, landscaping, and roof condition.
Finally, we end with the type of imagery most traders will be familiar with: drive-by imagery. Broker Price Opinions or Property Condition Reports typically include imagery captured in drive-by inspections. Both analysts and brokers use these images to determine the curb appeal of a given property. Often, they are used to influence the broker’s opinion of value. The resolution of these images is extremely high, typically below 1 cm square per pixel.
A key distinction from previous methods is that drive-by inspections require sending a person onsite. Inspection providers have extensive networks, so most properties are in coverage, but while other sources of imagery are stored and instantly available, inspections have turn-times that vary by provider and geography. Generally, results are returned within a week, although less populated areas can have extreme outliers. The requirement that a qualified individual visit the property increases the price significantly.
A critical challenge of drive-by inspections is the perspective that the name “drive-by” implies — these images typically reflect only the view from the street. As such, drive-by images do not capture the entire parcel and include more human variability in capture quality than other imagery collection sources.
Address Increasing Distress With Computer Vision
Regardless of the collection method, imagery has grown in importance to traders of distressed whole loans. However, relying on humans for these evaluations can present challenges, namely speed and scalability. A human analyst takes a couple of minutes to evaluate each set of property images, while computers take a couple of seconds to provide equally accurate and detailed information. This means that a pool of 1,000 loans would take a single analyst over 30 hours to review. In contrast, it takes a computer under a minute, given the massive power of parallel computing available.
Additionally, each evaluation completed by a human analyst would have some degree of uniqueness. While beauty is in the eye of the beholder, so, too is condition. This will naturally lead to evaluations varying between analysts and even across the same analyst as externalities change, like how much time is left to evaluate the pool, what time of day it is, and whether the kids are schooling remotely today. The use of computer vision and machine learning helps minimize these variables, which we will address in detail in the second post in this series.
With so many borrowers in delinquency and forbearance plans expiring, distress is likely to increase. Although critical information is contained in imagery, analysts do not have the capacity to review images of each property manually. For example, Fannie Mae recently announced a reperforming loan sale transaction containing 19,800 loans. An analyst would need over 300 hours to check each image, even if it took her only a minute per property. Computer vision allows investors to evaluate each of these properties in a fraction of the time. In the next post, we’ll dive into how computer vision can help address these and other challenges. To learn more, read Part 2 of this post series.
Interested in learning more about the use of geospatial imagery in the distressed mortgage market? Sign up for our on-demand webinar or for regular updates by subscribing to our newsletter, below.