Mortgage — Performance models


Mortgage performance models

Milliman is proficient in working with and analyzing complex data and with building econometric models that are transparent, intuitive, and informative for our clients. We have developed models for default scoring and assessing repurchase risk. Our risk models are built using industry data but can also be customized to the client, to ensure that they reflect the experience of your institution and your data.

We have used our expertise to assist multiple clients in developing econometric models for evaluating mortgage risks—both at the point of sale and for seasoned mortgages.

Milliman’s mortgage performance model is a loan-level model that estimates monthly probabilities of prepayment, delinquency, foreclosure/REO, and severity rates. The mortgage performance model is developed from a set of dynamic econometric models and is used by management to evaluate mortgage performance for analyzing mortgage credit risk transfer opportunities, mortgage servicing rights, whole loans, and others.

The model has the ability to estimate prepayment, default, and severity rates for a variety of loan types including fixed rate mortgages, adjustable rate mortgages, and FHA/VA mortgages. For actual loss severity exposures, the model takes into account expected foreclosure expenses at the local level and credit enhancement provided through mortgage insurance.

Milliman’s mortgage performance model is being used by many banks, (re)insurers, investors, and others for evaluating their exposure to the mortgage market. The models have also been relied upon by government agencies and policy makers in evaluating various mortgage policy changes and proposals.

Actual Versus Expected

Prepay Model
Foreclosure Model
Severity model



More information on the model, including data used to develop the models, fit statistics, and independent variables, can be provided upon request.



Milliman’s Mortgage Default Score is a loan-level algorithm that assesses the default risk of a mortgage either at origination or for a seasoned loan, where default is defined as the probability of foreclosure. Milliman’s Mortgage Default Score is a composite default rate calculation that combines three attributes of mortgage credit risk to estimate the frequency of borrower defaults:

  • Creditworthiness of the borrower
  • Underwriting properties of the loan
  • Macroeconomic influences

The Mortgage Default Score is developed from a dynamic econometric model and can be used by management to identify, evaluate, and manage mortgage opportunities.

Actual versus Expected

By FICO score
By combined loan-to-value
By home price appreciation
By origination year




Graph 1
Graph 2



Identify and manage the loans that are driving the risk of a portfolio.

Graph 1 provides an example in which a portfolio of loans is sorted by the Milliman Default Score and the cumulative loss is estimated for each loan. The blue bars on the chart show the cumulative percentage of the portfolio by Milliman Default Score, and the red bars represent the cumulative percentage of estimated risk. In this exhibit, 10% of the portfolio has a Milliman Default Score of 25% or greater; this 10% of the portfolio also accounts for 60% of the risk.

Identify the source of risk for a portfolio.

Milliman can segment the composite default score calculation into its three components to identify which component(s) contribute the most to default risk within a portfolio. This helps mortgage lenders understand what drives the risk in their portfolio and improve their origination process to mitigate default risk in the future.

Graph 2 provides a portfolio component analysis by Milliman Default Score cohort. For example, of the loans with a default score of 30%, the majority of the risk from this portfolio is from risky underwriting products (the red line on the chart). Therefore, to reduce its risk in this cohort, the lender could add underwriting guidelines to limit the amount of “risk layering” in order to reduce the overall risk of the portfolio.

Lenders with large correspondent channels use the Mortgage Default Score to compare their default risk for borrower creditworthiness and underwriting adjustments. They can then identify specific correspondent lenders that are delivering a more risky product while controlling for economic influences.

The graph below demonstrates a sample of the Mortgage Default Score being used to monitor new production for a lender’s correspondent channel. The chart ranks correspondents by the estimated default rate on their production. In the figure, we identify Correspondent #2 as presenting the most risk to the lender in that the volume of production is high and the correspondent is delivering more risky loans to the lender. This information would be used by a client to take corrective action through lowering purchases, altering underwriting guidelines, or increasing pricing to the correspondent if the trend continues.

Monitor credit quality of loans from correspondent lenders of other channels.


Milliman developed a loan-level model to estimate repurchase risk at origination based on the characteristics of the mortgage. The model is used by clients for a variety of purposes including repurchase reserving, quality control discretionary review, and benchmarking.

The graphs below show a comparison of Milliman’s model against historical experience using an out-of-sample dataset.

Graph 1
Graph 2



Using the above models, sellers can proactively reduce their repurchase risk by incorporating a pre-closing sampling process to identify loans with the highest risk of defects and take corrective actions before loans are sold.

The effectiveness of a pre-closing review process depends on the ability of the lender to identify two critical outcomes:

  • The characteristics and groups of mortgages that are more likely to have a defect
  • The characteristics and groups of mortgages that are more likely to be reviewed by the government-sponsored enterprises (GSEs)

Using data collected from internal quality control reviews and historical repurchase requests from the GSEs, lenders can use predictive modeling techniques to “score” applications by their repurchase risk prior to closing. Applications above a certain threshold receive a second review to check for errors or omissions in the application pre-funding. A study performed by a quality control company on QC reviews performed in 2012 and 2013 found that approximately 40% of all post-funding defects could have been identified and potentially corrected through a pre-funding audit. Using targeted sampling to identify high-risk loans could only increase this hit rate.


Assume lender “XYZ Mortgage Company” developed a scoring algorithm that segments its production in three risk levels of defect risk: low, medium, and high. The table below demonstrates how the process described above can reduce XYZ’s repurchase risk on 1,000 loans delivered to the GSEs. We assume that 40% of potential defects are cured through a pre-closing quality control review.

Annual Production 10% Random Sample
Pre-Funding Review
10% Targeted Sampling
Pre-Funding Review
Defect Risk Level Number of Loans Actual Defect Rate Number of Files Reviewed (10% of all buckets) Total Actual Number of Defects Number of Potential Defects Cured Number of Files Reviewed* Total Actual Number of Defects Number of Potential Defects Cured
A B C=A * 0.1 D=B * C E=D * 0.4 F G=B * F H=G *0.4
Low 750 20% 75 15 6 10 2 1
Medium 200 50% 20 10 4 40 20 8
High 50 80% 5 4 2 50 40 16
Total 1,000 29% 100 29 12 100 50 25
*Assume 100% high risk, 20% medium risk, and the remaining low risk loans are reviewed.

In the above hypothetical example, XYZ would be able to significantly reduce its repurchase exposure by targeting high-risk loans pre-closing. Specifically, a random pre-funding review would correct 12 defects and a targeted approach would correct 25 defects while reviewing the same level of 10% of the loans. Assuming an average loan balance of $200,000 and a severity of 30% for a repurchase, this would result in a reduced repurchase exposure of over $750,000 for XYZ.

10% Random Sample
Pre-funding Review
Targeted Sampling
Number of Defects Cured A 12 25
Average Loan Amount B $200,000 $200,000
Average Severity of Purchase C 30% 30%
Savings from Pre-Funding Review D=A * B * C $720,000 $1,500,000


Related services

Mortgage consulting

Mortgage compliance

Mortgage valuations and reserving

Mortgage feasibility studies

Mortgage government services

Mortgage credit risk sharing analysis

Practice Leaders

Connect with Milliman experts

Glowacki Jonathan

Jonathan Glowacki

Principal and Consulting Actuary

Jonathan is a principal and an experienced consulting actuary with the Milwaukee office of Milliman. He is a fellow and chartered enterprise risk analyst through the Society of Actuaries. He is proficient in working with and analyzing ...

We’re here to help

Ask the tough questions. We’re ready for them.