Milliman developed several integrated machine learning modules to leverage hundreds of local data series to evaluate the current state of the housing markets and produce estimates of future trends. The modules are components of M-PIRe (Mortgage Platform for Investments and Reinsurance); M-PIRe is a cloud-based platform designed specifically to evaluate mortgage market trends and Credit Risk Transfer deals.
The modules estimate:
- Future mortgage demand by geography, product, credit quality, and originator
- Home price conditions (opportunities and risk)
- Credit quality of mortgage originations with consideration of borrower attributes, underwriting and product features, and economics
This report provides the results of the home price conditions model as of October 2018.
A tale of two metros
While the general housing market appears healthy, with national average home prices expected to rise over the foreseeable forecast period, pockets of risk are developing in local housing markets. The table in Figure 1 provides a list of the top five and bottom five metropolitan statistical areas (MSAs) ranked by expected home price appreciation (HPA) over the next year.
Figure 1: Top 5 and bottom 5 MSAS: Expected HPA and probability of home price declines
| Geographic Area | Expected one-year HPA | Probability of negative HPA |
| Madera, CA | 4.80% | 2% |
| Homosassa Springs, FL | 4.78% | 3% |
| Sierra Vista-Douglas, AZ | 4.39% | 4% |
| East Stroudsburg, PA | 4.11% | 1% |
| Fresno, CA | 3.96% | 3% |
| Boulder, CO | -1.06% | >50% |
| Austin-Round Rock, TX | -0.89% | >50% |
| Bismark, ND | -0.68% | >50% |
| Denver-Aurora, CO | -0.62% | >50% |
| Fort Colins, CO | 0.60% | >50% |
| San Francisco, CA | 2.67% | 4% |
The model ingests many data series on local (i.e., MSA) housing trends, demographics, and economic conditions (e.g., number of housing starts, median home prices, home price appreciation over past several quarters, migration trends, new household formations, housing affordability, unemployment rates, per capita income, and various ratios and changes in these data series). These data elements are combined with an advanced machine learning algorithm to estimate the probability that home prices in a given market will decrease or increase within the forecast quarter. For example, the model estimates the probability of home prices decreasing by more than 25%, by between 25% and 15%, by between 15% and 10%, and by similar increments up to the probability home prices will increase by more than 25%. These probabilities are used to create an expected value for the change in home prices over the forecast period.
This article will look at Denver and San Francisco to demonstrate how the model works and what factors are considered.
Why Denver and San Francisco?
These two metro areas both experienced above-average home price appreciation over the past five years and have seemingly healthy housing markets. However, one is identified as a high-risk area, and the other as a high-cost area. San Francisco is not ranked as one of the top five MSAs, but it provides a good contrast to Denver for the purposes of this article. The side-by-side comparison on the next page will examine the drivers behind these two different paths.
|
Denver, Colorado |
San Francisco-Redwood City, California |
|
Home prices: The median home price in Denver increased by approximately 65% between 2012 and 2017. While appreciating home prices typically have a positive impact on future home prices, prices increasing too fast too quickly can result in unsustainable prices and affordability issues. |
Home prices: The median home price in San Francisco increased by approximately 60% between 2012 and 2017. Similar to Denver, appreciating home prices typically have a positive impact on future home prices, but if prices increase too fast too quickly it can result in unsustainable prices and affordability issues. |
|
Housing affordability: As home prices have increased, housing affordability has decreased. Housing affordability is measured as an index1 and represents the ability of a typical household to purchase a typical house in a given market. The housing affordability index in Denver decreased from 169 (indicating the average family has more than enough income to purchase a home) in 2012 to 105 in 2018 (indicating the average family has just enough income to purchase a home); this represents a 40% decrease in affordability. |
Housing affordability: Housing affordability has been on the low end of the spectrum in San Francisco since the early 2000s. The housing affordability index in San Francisco has hovered around 50 between 2000 and 2018, reaching a low of 42 in 2007 and a high of 77 in 2012. The housing affordability index is currently at 60, representing a 20% decrease in affordability from 2012. |
|
Housing sales to housing starts: As home prices heated up in Denver, many builders began constructing new homes. In 2012, for every housing start, there were 10 sales (in other words, somewhere around 10% of sales were new construction). In 2017, the ratio of housing sales to housing starts decreased from 10 to 1 down to 3 to 1. In other words, in 2017 around 33% of housing sales were new construction. This greatly increases the supply of housing and puts downward pressure on future home price increases. |
Housing sales to housing starts: San Francisco does not have a lot of room to expand geographically, so housing starts are fairly limited (for example, between 2012 and 2017, there were a total of just over 11,000 housing starts in San Francisco—this compares to 200,000 housing starts in Denver over the same time period). The majority of housing sales in San Francisco are existing homes, and the ratio of housing sales to starts in San Francisco is about 15 to 1. |
|
Housing starts per household formation: Between 2012 and 2017, Denver has grown significantly, with the number of households increasing by 8.3% (equating to 95,000 new households). During this same time period, there were 200,000 housing starts, or two new homes for every household formation. When there is a sustained period of housing starts outpacing housing formations (i.e. potentially more new homes being built that there are new buyers), negative pressures to home prices can arise. |
Housing starts per household formation: Between 2012 and 2017, the number of new household formations in San Francisco increased by 4.3% (equating to 32,000 households). During this same time period, there were only 11,000 housing starts, or one new home for every nine formations. Given the cost of homes in San Francisco and the scarcity of home starts, the demand for housing far exceeds the supply. |
|
Median house price to median household income: The median house price in Denver was $418,000 in 2017 and the median household income was $75,000. This represents a ratio of approximately 5.6 to 1. Median household income has increased by 22% from 2012 through 2018 (this compares to a 65% increase for home prices). Similar to the housing affordability index, this ratio indicates housing is becoming less affordable for the typical Denver family. |
Median house price to median household income: The median house price in San Francisco is $1,110,000 and the median household income is $110,000. This represents a ratio of approximately 10.0 to 1. Median household income has increased by 40% from 2012 through 2018 (this compares to a 60% increase for home prices). While home prices certainly are expensive in San Francisco, the supply of new homes is limited and the growth in home prices has not far outpaced growth in median household income. |
Figures 2 and 3 provide a visual of the model results for Denver and San Francisco, respectively, from 2002 through 2018. The orange and red bars represent probabilities of home price declines, and the yellow and green bars represent probabilities of home price increases. From these charts, we can see the model's estimated significant appreciation between 2012 and 2018 for both cities. However, it estimates home price declines in 2018 for Denver while projecting only a slowdown in growth for San Francisco over the next year.
Figure 2: Historical and Forecast Home Price Index, Denver
Figure 3: Historical and Forecast Home Price Index, San Francisco
Closing remarks
While the change in home prices from 2012 through 2017 is similar for Denver and San Francisco, underlying trends in housing availability, demographics, and income result in different forecasts for future home prices. The above analysis takes the results of Milliman’s proprietary house price forecast model and selects two markets for a deep dive drill-down into the root of the forecasts. The model uses advanced machine learning techniques to evaluate nonlinear relationships and correlations between hundreds of data series and home price movements. This allows Milliman to efficiently capture nuanced and granular differences between markets in a way that was not possible several years ago.
Milliman produces forecasts of home prices for all 50 states and over 400 MSAs. These forecasts are used internally to evaluate market conditions for estimating future mortgage origination volume, loan performance, and evaluating credit-risk transfer bonds and reinsurance treaties.
While the above charts are shown through 2018, model forecasts are generated over five-year forecast horizons.
Documentation and details of the models can be provided upon request.
Disclaimer
In performing this analysis, we relied on data and other information provided by third parties. We have not audited or verified this data and other information. If the underlying data or information is inaccurate or incomplete, the results of our analysis may likewise be inaccurate or incomplete. In that event, the results of our analysis may not be suitable for the intended purpose.
We performed a limited review of the data used directly in our analysis for reasonableness and consistency and have not found material defects in the data. If there are material defects in the data, it is possible that they would be uncovered by a detailed, systematic review and comparison of the data to search for data values that are questionable or for relationships that are materially inconsistent. Such a review was beyond the scope of our assignment.
Differences between our projections and actual results depend on the extent to which future experience conforms to the assumptions made for this analysis. It is certain that actual experience will not conform exactly to the assumptions used in this analysis. Actual amounts will differ from projected amounts to the extent that actual experience is better or worse than expected.
1National Association of Realtors. Methodology: About the Index. Retrieved November 21, 2018, from https://www.nar.realtor/research-and-statistics/housing-statistics/housing-affordability-index/methodology.