Freddie Mac is harnessing big data and artificial intelligence (AI) to drive significant value for our customers, their borrowers and the industry. There’s no shortage of data at Freddie Mac and we’re putting it all to work to build solutions that address customer pain points or real business opportunities, including the current appraisal process.
The Dawn of Artificial Intelligence
Artificial intelligence, specifically, machine learning, allows us to gain insights from huge quantities of data; cloud computing allows us to process this data in very little time.
Machine learning methods typically have an advantage when it comes to making predictions and accounting for complex interactions among different factors. Estimates from machine learning methods can provide a confidence metric, which is useful in many risk management optimization routines, but absent in an appraisal. In particular, a prediction model provides a metric that informs the user how much confidence they can have in the estimate of value. For example, a track suburban home that is easy to value might be assigned high confidence, while a rural property with few nearby comparables, low confidence.
Because of its sheer volume—and several enabling factors such as storage cost reductions, machine learning and cloud computing—data have become an important consideration for companies. Leveraging all of this data, new AI technologies make prediction cheap! As a result, not only are we going to start using a lot more prediction, we’ll see it emerge in surprising new places.
At the same time, the cost of prediction will impact value of other things; increasing the value of big data and cloud computing and diminishing the value of skills such as human prediction. The human/machine tradeoff comes down to comparative advantage: As machine learning replaces the predictions humans make, the value of human prediction will decline, but the value of human judgment will increase.
Leveraging Big Data
Our lenders approached us with a pressing pain point. There is a concern in certain markets that there is a shortage of appraisers and few new entrants into the profession; completion of an appraisal report has extended turn times; and fee increases in those markets
So, we looked for ways to improve timeliness and customer satisfaction while maintaining credit quality and prudent risk management. We did this by:
- Reviewing the functions performed by the current appraisal process, including inspecting the property to determine marketability and condition, developing an opinion of value, protecting against fraud via appraiser independence, and helping to mitigate repurchase risk.
- Leveraging and enhancing our automated valuation model, Home Value Explorer® (HVE®), which uses machine learning to estimate the market value of over 90 million single-family, condo, and two-unit properties. HVE® consistently performs well in internal and external tests and is driving interest in our credit risk bonds.
- Leveraging our prior experience. Over the past 15 years, Freddie Mac has offered several alternatives to a full appraisal. These alternatives ranged among the following: no inspection, exterior inspection, interior and exterior inspection, and exterior appraisals.
- Using big data, including public records data from multiple providers, MLS data from multiple providers, repeat sales data, and appraisals in the Uniform Collateral Data Portal® (UCDP®).
Mixing all these ingredients (HVE®), prior experience, machine learning and big data—allowed us to develop logic and algorithms to estimate a property’s value and mitigate property condition risk. As a result, we sometimes waive the need for an appraisal via our offering, Automated Collateral EvaluationSM (ACE), in Loan Product Advisor.
The Future of Appraising
ACE represents an important step in improving valuation practices. To date, ACE is applied to a relatively modest portion of loans that Freddie Mac funds. This is to be expected as we need to be prudent and allow the solution to be proven over time.
A major benefit of machine-learning methods is that they can scale in a way that humans cannot. However, this comes with the downside that they struggle to make predictions in unusual cases, instances where data are scarce or the property is atypical. For rare events, machine learning methods have limited use.
Perhaps the biggest weakness of machine learning methods is that they sometimes provide wrong answers that they are confident are right. This is why human judgment provided by the appraiser is so important. Appraisers will continue to be needed to ensure that the sale is an “arms-length” transaction, that there are no excessive seller contributions, and to identify a property’s condition, features and marketability.
Data are clearly important and will be central to innovation going forward—from pure automation to traditional appraisal. So too will two major valuation-related initiatives spearheaded by our regulator, the Federal Housing Finance Agency: UAD Redesign (i.e., forms redesign) and the Appraisal Modernization initiative.
Many of the advances in appraising will come by finding solutions for the ‘gray’ space that lies
between the very simple cases handled by ACE and the very complex cases that require a full Uniform Residential Appraisal Report. Appraisers will be an integral part of formulating these advances by focusing on areas of comparative advantage.
The combination of humans and machines often generates the best predictions. As we attack the gray space, it might be beneficial to provide the appraiser with tools he/she can leverage in developing their report. This will also help capture standardized feedback for areas of the report where Freddie Mac information differs.
Finally, the stated purpose of most appraisals is to estimate the most probable price a property would bring in a competitive and open market. But, from a risk management perspective, sales price alone is insufficient to understanding the risk of a loan supported by a property serving as loan collateral. Appraisers need to identify instances when the sales comparison approach is not appropriate and bring back the other approaches as a check on what the market is doing. At the end of the day, borrowers and investors care about the sustainable value of a property, which is only suitably approximated by the sales comparison method when the market is at or near equilibrium.
Have any comments or would you like to submit an article of your own? Email firstname.lastname@example.org.