Friday, October 17, 2014

How to Develop a REO and Foreclosure AVM


The traditional AVMs are appropriate for origination and portfolio management – they are however totally inappropriate for the REO and Foreclosure (“foreclosure”) market. Subsequent to the bursting of the recent housing bubble, the foreclosure market has been growing nationwide, creating the need for specialized AVMs that could appropriately address that market segment. 

As such, the development of the foreclosure AVM is fundamentally and technically different from the traditional AVMs. Obviously, if the traditional AVMs are applied onto the foreclosure portfolio, the portfolio values would be unjustifiably inflated, creating false expectations from the market. Even when a statistical haircut is applied to the traditional AVM values to make them usable in the foreclosure environment, the application does not generally work considering foreclosure markets are inherently pocket-oriented. Of course, specialized foreclosure AVMs would be cost ineffective where foreclosures command a very small percent of the overall market.

 While developing a specialized foreclosure AVM, the following modeling issues must be paid special attention to:
 

1. Sales Sample  

Unlike the traditional AVMs which depend strictly on recent arm’s length sales, the foreclosure AVMs must be developed off of the universe of foreclosure and short sales. Therefore, it would be easier to generate sales samples in those Metropolitan Statistical Areas (MSA’s) where foreclosures are common or are elevated.  

Due to the bulk purchases (REO to Rental) by major institutions in Sun states in 2012-13, the incidence of foreclosures had significantly subsided causing that glut of inventory to take a nosedive and prices to jump. As the institutions exited the market in late 2013 paving the way for the homebuyers and small investors to return, the foreclosure has started to return to normalcy there.  

So, only the foreclosure and short sales since their exit (Q3-2013: AZ, CA, FL, GA, NV and TX) must be considered for the sale sample. In other states where institutional involvement was significantly lesser, two years worth of sales (since Q3-2012) could be represented. 

In testing the representativeness of the sample, each of the three variable groups – continuous, categorical and fixed neighborhood – must be separately tested. At least, two of the four common continuous variables (Land Area, Bldg/Living Area, Baths/Bath Fixtures and Year Built/Age) must pass the 10th to 90th percentile test (not just the median), while a group of categorical (Style, Story, Exterior Wall, Heating and Cooling, Grade, Condition, etc.) and fixed neighborhood variables (Towns, Villages, School Districts, etc.) must also be successfully tested. Since these are non-continuous variables, the frequency percent test would be appropriate. The most liquid categories must successfully pass the test (please read our book entitled ‘How to Build a Better AVM’ for more developmental details).
 

2. Time Adjustment

Time adjustment must take place in the form of quarters (instead of monthly) as quarterly adjustments are inherently statistically smoother than their monthly counterparts. Additionally, the broader markets (25th vs. 50th vs. 75th percentile) must be separately evaluated to find out if they have moved in tandem. If they have moved in tandem, the same set of time coefficients could be applied; or else, separate time coefficients must be derived from each value strata of the broader market.  

If the two most recent quarters show faster decay in value, an additional negative adjustment could be factored in, allowing adequate marketing gestation between the development and the shelving of values. Since time is a surface correction, it would be inappropriate to drill further down to the geographic areas, etc. (please read the chapter on Time Adjustments in our book entitled ‘How to Build a Better AVM’ for more details).
 

3. Multiple Regression Analysis (MRA) 

While in traditional AVMs the categorical variables are generally assigned starting linear values auto-regressively, the foreclosure AVMs may take a combination of heuristic values for hierarchical variables and auto-regressive values for the other variables, considering the smaller sample sizes and the resulting illiquidity in many categories. In other words, hierarchical variables like Grade and Condition may assume heuristic values while non-hierarchical variables like Style, Story, Exterior Wall, Town and School Districts may take auto-regressive values.  

As the MRA models are run, the starting values would be adjusted. Due to the expected paucity of data points and variables in foreclosure AVMs, the significance level of the independent variables should be set at a lower level than the traditional AVMs. For example, if a minimum t-value of 2 is used to control the significance level in the traditional MRA model, it should be set at, say 1, in the foreclosure MRA model.  

Likewise, while running the multicollinearity matrix, the threshold must also be higher than the traditional model. If the threshold in the traditional model is set at +.3 or -.3, it should be significantly higher, say +.5 or -.5, in foreclosure model so that the variables are not easily removed from the equation.
 

4. GIS/Spatial Analysis

Traditional AVMs require significant amount of GIS and spatial analysis to define and quantify non-fixed planes and pocket areas. Those variables are then added to the MRA equation, along with the fixed neighborhood variables. Since foreclosures are more clustered and/or pocket-oriented, the concept of non-fixed planes does not take form. 
  
Therefore, instead of the traditional three-stage MRA model – additive, multiplicative and GIS – the foreclosure model would require only a two-stage model. In fact, if the foreclosure pockets within the model could be clearly identified and defined (using State Plane X-Coordinate & Y-Coordinate or Latitude & Longitude), they should then be introduced in the MRA as linearized clusters or pockets. Any more GIS/Spatial analysis would be overkill.
 

5. Residual Analysis  

Traditional MRA models are generally optimized by way of residual analysis. The most common form of residual analysis is the optimization of sales ratios (MRA Estimate / Adjusted Sale Price) of the categorical variables, meaning that the categorical variables are iteratively adjusted in a way that the individual sales ratios approach a median value of 1 (or 100 if percentile).  

On the other hand, since foreclosure sales samples are made up of widely diverging elements (level of consumer distress, urgency on the part of the lending institutions to liquidate REO portfolios, etc.), extensive residual and outlier analyses would be a waste, further shrinking the modeling sample (please read the chapter on Residual Analysis in our book entitled ‘How To Build A Better AVM’ for more details).  

Therefore, the generally accepted quality metrics like Coefficient of Dispersion (COD), Price-related Differentials (PRD) would be unnecessary. Instead, a valuation review group consisting of internal foreclosure experts, selling brokers and AMCs (when involved) should be encouraged to study the foreclosure AVM values to determine their effectiveness.

In 2010, Freddie Mac announced the use of AVM values for review appraisals to help lenders “more easily identify potentially inflated appraisal values that may need additional review early in the origination process.” On the heels of this announcement, the lenders holding large inventories of REO and foreclosures (late stage) would also be better served with the proposed specialized foreclosure AVMs, not only to price their portfolios more accurately, but to prevent appraisal frauds as well. Those AVM values should be used as control values to trigger the Supervisory QC.  

Simply put, the incoming REO appraisals must be compared against the AVM values and those that deviate from the internally acceptable range (say, +/- 20%) must be flagged for the Supervisory QC. Without such internal controls, those portfolios would be susceptible to appraisal frauds. AMCs should also use a similar approach to score and rank their appraisers.


Thanks,

Sid Som MBA, MIM
President, Homequant, Inc.
contact@homequant.com 

Note - Reprinted from my book 'Alternate Applications of AVM.'  Now available on Kindle on Amazon (search 'Sid Som').


Follow us on Twitter
https://twitter.com/SidSom1 
 

No comments:

Post a Comment