Location matters: enhancing credit risk models with ML/AI socioeconomic insights

Product

Locality Health Score (LHS)

At SiteData.io we leverage the power of machine learning and socioeconomic data to provide unique, actionable insights into the credit health of localities across the U.S. Our Locality Health Score (LHS) feature empowers credit professionals to make data-driven decisions, optimizing loan portfolios and maximizing returns.

We market historical and current LHS scores for over 60,000 U.S. cities and neighborhoods. These scores were generated by our proprietary ML/AI models fine tuned to optimally mirror the expected principal recovery percentage of loans issued in localities with widely varying socioeconomic conditions.

LHS can be utilized in credit risk models to optimize the interest rates of consumer loans in order to regularize expected return. Principal loss (unpaid principal) on high scoring LHS loans can be 30% less than on low LHS loans. On a $1B loan portfolio reducing loss by 1% yields an extra $10M. That is the magnitude of difference our predictors can provide, and we have the data analysis to back that up. Optionally, LHS can help in selecting which loans to include in an optimized portfolio.

Outside of the credit space, here are some other areas where LHS could be of use:

• Insurance: Insurers could use LHS to help assess the risk of insuring individuals, businesses and properties located in different cities, potentially influencing premiums.
• Property Investment: Real estate investors could leverage LHS to identify promising locations for property acquisition and disposition, using both current scores and, for trend analysis, historical scores.
• Targeted Marketing: Marketers could tailor their strategies based on the health scores of different localities.
• Asset Management: Asset managers could use LHS to identify localities with economic growth potential for investment opportunities.
• Mortgage Lending: Lenders could incorporate LHS into their risk assessment models for mortgage rates and approvals.

Note: Our models do not include any featurization for: geographic location (like "city" or "state"), race, color, religion, national origin, sex, marital status, or age. 1

Roadmap

What we offer, and what's ahead

Let's begin with what we have to offer... We have a database containing over 60,000 cities, neighborhoods and counties with socioeconomic data harvested from many sources. We've used this data to build machine learning models to forecast expected percent of loan principal to be repaid derived from the (featurized) socioeconomic data, rescaled into a 0 to 1 Locality Health Score (LHS) feature, for ease-of-use and model dimensionality reduction.

• We market a complete by-city (or neighborhood) dataset of LHS scores, both current and historical, in 3 flavors, which may also be purchased point by point.
• Also, we have a complete by-3 digit zip codes (like 941xx) dataset of LHS scores, both current and historical, in 3 flavors.
• LHS is a function of socioeconomic ratios. We can compute it for any geography where those inputs exist — city, neighborhood, census tract, or 3-digit ZIP.

Here are other data products under consideration. Please let us know what interests you!

• ML/AI models featurized down to the zip code level.
• ML/AI models to forecast future by-city unemployment, income levels, home prices, etc. given exogenously forecasted US values. We don't know next year's national unemployment figure, but if we did, we could forecast expected unemployment of localities.
• Given a loan dataset including location of issuance and outcomes, we can report back on to what extent it was already featurized using socioeconomic data, and also expected additional base points loss due to it being not included in the credit risk model.
• We would love to partner with you on your proprietary data, loan data, or whatever, either on a consulting basis (e.g. to build custom ML/AI models for you) or as a mutual joint venture.

Blog

From our research

The Economic Health of a Borrower's City Affects Loan Performance

We have created a locality database containing geographic, economic, and demographic information for the U.S. as a whole, for all 50 states (plus Washington D.C.), for all of the over 3,000 U.S. counties, for over 38,000 cities. (circa 2014.)

Using ML/AI to Create a Powerful Location Based Feature for Credit Risk Models

With an interest in improving credit risk models created for consumer loans, we focus here on analyzing how the socioeconomic environment in which a borrower resides may impact expected loan performance, with the goal of adding a meaningful new feature. (revisited in 2024.)

The Tale of Two Cities: Miami and Corpus Christi

A city's LHS over time indicates if the condition of the city is stable, improving or deteriorating. So in addition to its current LHS score, the LHS trendline can be a useful indicator for decision making.

About Us

The story behind LHS

Hello! I'm Harlan Seymour, founder of SiteData.io.

In 2014, I happened upon the peer-to-peer lending space, which was publishing rich data on all of their consumer loans on offer. I love data! So, I went to work studying it to see if I could gain an edge over the baseline performance.

And I did! I found that featurizing the socioeconomic factors of borrowers' cities to create a Locality Health Score (LHS) allowed me to select loans predicted to outperform the mean by at least 1%.

From 2014-2016, I invested almost $1M in consumer loans at Lending Club, and I did indeed achieve about a 1% lift over the baseline return.

Since then, I became fascinated by machine learning and AI, winning a Kaggle AI contest, and, as a founding engineer at Afresh.com, building a large quantile regression ML model to forecast demand for fresh food at thousands of stores. See:

• 1st Place Solution (in DonorsChoose.org Kaggle competition)
• Predicting Demand at Hundreds of Stores with Multi-Task Learning and Good Features

Now in 2026, I have systematized LHS for over 60,000 cities and neighborhoods, using the more powerful modeling techniques available today, creating annual scores for these localities from 2014 to 2026, so now trends over time can be discovered. I believe that using our LHS in credit risk models will significantly improve loan portfolio performance!