Sunday, April 22, 2018

How to Develop an Automated Valuation Model (AVM)

(Click on the image to enlarge)

A Market AVM is developed using a group of recent arms-length sales, generating an econometric solution (blue graph). 

1. If the sales sample (must be representative of the population it is derived from) comprises an extended timeframe, say 12 to 18 months, sales are generally time-adjusted. 

2. Time-adjusted sales, rather than the Pure Sale (as originated), therefore becomes the dependent variable, while Size (Land and Bldg), Bldg Age, Land & Bldg Type/Quality (Site Elevation, Type of Lot, Bldg Grade, Condition, Exterior, Style, etc.) and Location (Value influencing Fixed Neighborhoods like School Districts, Assessment Districts, etc.) serve as the independent variables in the Multiple Regression Analysis (MRA) equation that usually defines the model structure. 

3. Of course, the most efficient models are generally built around multiple stages (Additive=Quantitative /Continuous Vars; Multiplicative-1=Qualitative a.k.a. Descriptive Vars; Multiplicative2=Fixed Location Vars) AND multiple cycles (Cycle-1=Outliers Defined/Removed; Cycle-2=Residuals Analyzed and Corrected; Cycle-3=Final Model Developed post Hold-out Testing). 
 

4. As the model gets further fine-tuned (requiring the aforesaid iterative cycles involving removal of outliers and correction of residuals, which are not shown here), the model R-squared tends to approach .90 and the COD/COV falls below 10, thus convincing technicians to apply the model on to the assessment population in order to generate the assessment roll (or on to the mortgage portfolio to have it re-priced in line with the current market).

5. Of course, it's always wise to test the draft model on to a hold-out dataset (cycle-3 as mentioned above, with the hold-out having very similar attributes as the population dataset) before applying on to the actual population. This interim hold-out test helps identify and correct any model errors and inconsistencies, instead of having to discover and correct them post population application which is very time-consuming and cumbersome. Since this test must poduce very similar results as the model's as well as the population's, this is an excellent QA/SQC step as well.



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

Link to our Trend-setting AVM Book on Amazon (Kindle/Paperback)

No comments:

Post a Comment