Details Behind Behavioral Modeling

In Episode 7 of The AiCR Exchange, Joe Furlong sits down with Dick Kazarian, Managing Director of MIAC’s Borrower Analytics Group. Dick has been building and applying behavioral models across mortgage markets through multiple credit cycles, including the peso crisis, the great financial crisis, and COVID. He walks through how models are built, where they break down, and why subject matter expertise cannot be replaced by machine learning alone. 

What does MIAC’s Borrower Analytics Group do? 

MIAC’s Borrower Analytics Group builds and maintains the behavioral models that support MIAC’s valuation, hedging, brokerage, and advisory work. The group produces three types of analytical output: valuations that calculate market price given a spread or spread given a market price, hedging analytics that measure how much a price will move if rates change, and stress analysis that runs macro factor scenarios to project how assets will perform under adverse conditions. The models support both MIAC’s internal production teams and external software clients who license MIAC’s analytical platform. 

How are behavioral models built? 

Model development follows a defined lifecycle. It starts with understanding the business requirements, then moves to data acquisition, preparation, and normalization. The team identifies what datasets will best answer the question the model is trying to address, specifies the model structure, and runs the estimation process. Extensive back testing and diagnostic work follow. The model then has to be implemented into software, which Dick describes as where most of the real work begins. Implementation testing surfaces issues that estimation alone will not catch. From there the model goes through documentation and validation, and then into ongoing performance assessment because models get out of date and have to be monitored continuously. 

Why does data quality matter so much for behavioral modeling? 

Data quality shapes every output a model produces. If the data going in is incomplete, inconsistent, or reflects anomalous periods that were not accurately characterized, the model will produce outputs that cannot be trusted. Dick is direct: everyone is in the data business. The only question is whether they are going to be good at it. Models built on incomplete or distorted data will fit the past but fail to forecast well. The great financial crisis is the clearest example. Loan level LTVs reached 130 during the crisis, a scenario no model built in 2006 had historical data to account for because no originator was creating loans at those LTVs. The data for that scenario simply did not exist before it happened. 

Why can AI and machine learning not replace subject matter expertise in behavioral modeling? 

Machine learning models have a role in mortgage analytics but are not a replacement for subject matter expertise. Dick identifies the conditions where machine learning works well: many variables, limited economic intuition about how those variables relate to each other, large amounts of training data, and stable environments without policy interventions. Mortgage markets fail on several of those conditions. The mortgage is the largest liability most homeowners carry, which means regulators, Congress, and state agencies intervene constantly. HARP changed prepayment speeds overnight. Dodd-Frank changed appraisal quality structurally. COVID broke the historical relationship between unemployment and credit losses entirely because government relief programs meant people were financially better off unemployed than employed. A machine learning model trained on historical data would have produced wrong outputs for all of those scenarios. Subject matter expertise is what allows a team to recognize when the data is misleading and adjust accordingly. 

How do prepayment and default assumptions get formed? 

Prepayment and default behavior is driven by a combination of loan attributes and macroeconomic factors. Loan balance, LTV, credit quality, loan age, and product type all matter. So do interest rates, home price appreciation, unemployment, and inflation at both the national and MSA level. VA loans are a useful example of how product-specific factors interact. VA loans prepay at roughly 50% CPR, significantly faster than other product types, because VA borrowers tend to have strong credit, and the VA streamlined refinance process is efficient enough that borrowers can refinance quickly once rates drop. LTV matters for both credit performance and prepayments across all product types, but VA borrowers show better credit performance than FHA borrowers at comparable LTVs, which is an important distinction for anyone building or interpreting models across government loan products. 

What is the role of through-the-cycle data in behavioral modeling? 

Building models that hold up across different economic environments requires data that spans multiple credit cycles. A model estimated only on data from a stable period will not know how borrowers behave when home prices fall sharply, unemployment spikes, or policy interventions alter the normal relationships between variables. The bubble vintages from 2003 to 2005 produced credit losses that models of that era could not have anticipated because the LTV distributions that emerged during the crisis were outside the range of any prior origination experience. Post-crisis originations showed lower credit losses than pre-crisis originations even after adjusting for credit quality differences, largely because appraisal quality improved significantly after Dodd-Frank reforms eliminated broker-ordered appraisals. Having data through multiple cycles is what allows a modeling team to understand which relationships are structural and which are period-specific. 

What is ahead for the residential mortgage market? 

Dick’s view as of early 2026 centers on two themes. The first is expanding data availability. Credit bureau data that shows a borrower’s total indebtedness, not just mortgage debt, is becoming more accessible and will enable more precise behavioral modeling over time. The second is policy risk. The mortgage market is subject to constant regulatory and legislative intervention, and those interventions can produce seismic changes to prepayment and credit behavior. LLPA structures are currently a focus of political debate from both sides, and changes to them could have significant downstream effects on prepayment speeds and credit performance. Dick’s closing point: policy risk is very difficult to forecast and very impactful when it arrives. 

Frequently Asked Questions About Behavioral Modeling in Mortgage Markets 

What is behavioral modeling in mortgage finance? 

Behavioral modeling in mortgage finance is the process of building quantitative models that forecast how borrowers will behave over time, specifically their likelihood of prepaying, defaulting, or becoming delinquent. These models take loan-level attributes and macroeconomic inputs and produce outputs used for valuation, hedging, stress testing, and capital planning. 

What inputs do prepayment and default models use? 

Prepayment and default models use a combination of loan-level attributes including LTV, loan balance, credit score, loan age, and product type, as well as macroeconomic factors including interest rates, home price appreciation, unemployment, and inflation. The interaction between loan attributes and macro factors is what drives borrower behavior, and models need data that spans multiple economic cycles to capture how those relationships change across different environments. 

Why did behavioral models fail during the great financial crisis? 

Models built prior to the great financial crisis lacked historical data on borrower behavior at the LTV levels that emerged during the crisis. No originator created loans at 130 LTV under normal conditions, so no training data existed for that scenario. Additionally, appraisal quality was poor before Dodd-Frank reforms, meaning the collateral values underlying loan LTVs were inflated in ways the models did not account for. The crisis exposed how dependent model accuracy is on the quality and completeness of the data used to build them. 

What is MIAC’s Borrower Analytics Group? 

MIAC’s Borrower Analytics Group builds and maintains the behavioral models that support MIAC’s valuation, hedging, and advisory services. The group develops prepayment, default, and severity models across residential, commercial, and consumer loan types, and supports both MIAC’s internal production units and external clients who use MIAC’s analytical software platform.