The MIAC CORE™ is the name of MIAC’s new class of behavioral models for all asset classes including residential non-agency whole loans/MSRs, residential agency whole loans/MSRs, commercial whole loans, auto loans, credit cards, unsecured consumer loans, etc.

Each of these models has two components – a Voluntary Prepayment Model and a Loss Model. Each of these model components has been built to measure their COnditional REsponse to at least five macro factors required for bank stress testing.  Each of these model components is the CORE™ of MIAC’s cash flow behavior for asset valuation, balance sheet, capital requirements and net interest income simulations.

Developed to address the growing needs of market participants

The MIAC CORE™ – Non-agency Loss Model (“NALM”) was developed to address the growing needs of market participants to understand the dynamics of the credit behavioral response to changes in important macro economic factors.  With the arrival of CCAR and DFAST, not only will financial institutions need asset and enterprise-specific behavioral models, that have been calibrated to respond to requisite macro factors, but asset pricing methods adequately robust to incorporate these macro factors in their underlying pricing model and, by extension, into their forecasted cash flows.

The very significant contribution of “NALM” is the granularity of the cash flow simulation process within MIAC’s software suite, Vision™.  This unparalleled level of specification is possible because the dataset, being so large, allows for granular transition roll rates to be measured with statistical confidence.

Thus granular delinquency, foreclosure, curing, and REO simulation processes will improve the modeling of delinquent payment advances costs for mortgage servicers and will also improve the near-term forecasts for expected REO liquidations. Additionally, allowance for Loan and Lease Loss (ALLL) and SOP 03-3 accounting challenges can be addressed with a new level of accuracy.

Two very large datasets

The MIAC Non-agency Loss Model was built with two very large datasets:

  • The first dataset
    is the entire population of loans that were used as collateral for residential MBS private-label (non-GSE) bonds.  This population included over 23 million loans, nearly one billion loan monthly transactions, and starts in the middle of the 1990s.
  • The second dataset
    is culled from five of the largest Puerto Rican banks.  This dataset starts in the early 2000s and includes nearly 1 million unique loans.  Puerto Rican borrowers display unique payment behaviors and the MIAC NALM is able to measure this behavior accurately.

Utilizing this large dataset, MIAC can measure the asset specific, geographic specific, and institution specific behavioral response to the requisite macro factors – HPI, CPI, GDP, unemployment, and interest rates.  In addition, the borrower behavior while in foreclosure, including the cure rates, REO entry rates, the timing of REO entry and the timing for Cure from FCL, are all precisely measured.

New Approach to Modeling Mortgage Losses

The primary goal of a mortgage loss model is to forecast, as precisely and consistently, the frequency, timing and severity of mortgage loss behavior.

Historically, when modelers approached this problem, they utilized historical regression behavior thereby needing large datasets to build and calibrate their regression models. However, the availability and numerical size of actual REO data was difficult to obtain and much smaller than foreclosure entry data.

So rather than regress loan attributes to REO completion rates, the modeler regressed loan attributes to FCL Entry rates, which was a larger data set than REO completion rates.

A separate model was built to measure the time from foreclose entry to REO entry and ultimately REO completion or liquidation.  The sample size of foreclosure entry was larger and thus enabled the models to have better statistical significance than if the process where REO completion was directly regressed.

Wall Street dealers, rating agencies, bank research departments, and independent vendors shared this approach.  And once a loan entered foreclosure, there was a high probability that the loan would ultimately enter and complete REO.  Typically, only a modest percent would cure from foreclosure, so measuring and regressing FCL Entry was seen as a more defensible approach.

“MIAC choose a newer method that is more robust”

The MIAC Approach

However, when MIAC approached this problem with a very large dataset and re-examines the trade-offs, MIAC choose a newer method that is more robust. D30 Entry events are far more frequent than FCL Entry events.  FCL Entry events are far more frequent than REO Completion events.

If one regresses the key loan attributes against foreclosure entry, the results of the statistical correlations are unimpressive (R-squares in the 0.2s). The statistical correlations improve dramatically, however, when the D30 Entry status of a loan is included in the independent parameters. The statistical correlations improve approximately four times (high 0.8’s) when delinquency status is included.

Moreover, the delinquency status dominates the regression outcome – to the tune of six times. This has dramatic consequences for mortgage loss modeling.

Therefore, in building a regression model to forecast the frequency and timing of delinquency 30 days (D30 entry), using the key loan attributes and the macro factors, the correlations are statistically significant and dramatically better than foreclosure entry.   And very importantly, the transition roll rates have a surprising level of stability provided that key loan level attributes are included. The transition roll rates do have a dependency on some of the loan attributes, but these are identified.

This new approach to modeling D30 entry and then modeling “loan sector”-specific delinquency transition roll rates can improve the loss model frequency forecasting accuracy by more than four times over the traditional methods.  This is a game changer for mortgage market participants.

What about the Roll Rate Transitions?

With very large datasets, modelers can measure the components of the roll rate transitions with more precision and with statistical significance.

Because MIAC Analytics™ software is used to model more residential whole loans than any other analytical solution in the industry, we choose to increase the granularity of the cash flow simulation process to include every measurable transition roll rate that could be measured with statistical significance.  A challenging modeling and model validation quest that we have achieved.

Below is a chart of all the transition roll rates that can be measured with statistical significance and are included in the MIAC Analytics™ cash flow process.

MIAC CORE Non-Agency Loss Model - image page 3

Figure 1:     MIAC CORE™ Non-Agency Loss Model      Source:    Mortgage Industry Advisory Corporation

What about the timing?

The model contains several timing curves. The first curve shows the timing of when a loan will enter D30. Specifically, MIAC segregates the population of current and performing loans into two categories – Clean Current and Dirty Current. Clean Current loans are loans that have a “current” delinquency status (e.g. not delinquent) and have never been delinquent.  Dirty Current are loans that presently have a “current” delinquency status but have, at one point in their history, been delinquent.

The Dirty Current loans have a distinctly different D30 entry frequency model and a distinctly different timing curve from the Clean Current loans. The propensity for a loan to enter D30 is dramatically higher for a loan that has previously been delinquent. Not surprisingly, the timing of their D30 is distinct from the Clean Current loans.

For those populations of loans that MIAC has collected and normalized the loan performance data, MIAC segregates discrete databases of loans that have been ‘EverDelinquent’, ‘EverBankrupt’, or ‘EverFCL’.  Because of the statistical significance of having this data, MIAC refers to this segregation as our ‘Data Enrichment Process’.

The other timing curves are loan sector, geographic and servicer specific.  These three curves are “Time_To_Enter_REO”, “Time_To_Cure_Exit_FCL”, and “Time_IN_REO”.

The “Time_To_Enter_REO” curves are of course unique for loan sector, state and servicer. After a loan enters foreclosure, some percentage will ultimately exit foreclosure and enter REO.  Likewise, another percentage will ultimately exit foreclosure and cure.   There are many possible delinquency states that the cured loans will enter, including bankruptcy.  And those remaining loans will continue in foreclosure. The timing of when the loans leave FCL to enter REO is governed by the “Time_To_Enter_REO” curves.

The timing of when the loans leave foreclosure to cure or enter bankruptcy is governed by the “Time_To_Cure_Exit_FCL” curves.  For each sector, state and servicer there is a Time_To_Enter_REO. The software uses the appropriate curve. Because foreclosure timelines are distinct by sector, state and servicer, MIAC constructed these “Time_To_Cure_Exit_FCL” curves to also be unique by each loan sector, state and servicer combination.

By extension, once a loan enters REO, the model includes state and servicer specific timing curves, “Time_IN_REO” curves, which are distributions of the time it takes for loans to liquidate.

Model Validation

How is a MIAC client assured that the specifics of model functionality and the MIAC Analytics™ software is operating in a manner consistent with the underlying statistical model?  MIAC has built model validation tools to enable both software licensing clients as well as asset valuation clients to engage MIAC analysts in the details and defense of the model behavior.

Over the past several years, the users of financial models have been required to demonstrate that they understand not only the implications of the models but also the mechanics and methods of their financial models.

MIAC embraces this approach and has a long-standing practice of working with clients to open up the underlying black-box pricing models and work collaboratively toward comprehensive model comprehension.


The MIAC CORE™ – Non-agency Loss Model has several innovations and improvements from prior mortgage industry tools.

In this article we have focused only on the high-level summary of the key innovations and improvements. The major innovations are: migrating the frequency simulation of D30 to dramatically improve the predictive ability of the model; increasing the granularity of the transition roll rates; and increasing the granularity and specification of timing in the foreclosure process.

The details of the model behavior will be shared collaboratively with our clients as we mutually address mortgage valuation and balance sheet management challenges.

Request a demo

Find out more about MIAC CORE™ – request a demo

Joseph Furlong, Managing Director, Due Diligence and Borrower Analytics Group

MIAC Perspectives – Summer 2015
MIAC CORE ™ : Non-Agency Loss Model