Sales can come variously from MLS organizations, County Clerk Record filings (in DISCLOSURE States - where filing dates, loan amounts and sale prices are Recorded as opposed to NON-DISCLOSURE states where at most the loan amounts and filing dates are given in addition to grantor/grantee names, addresses), private data providers, and various other sales providers.

County Assessor records are purchased/licensed by county as are County Clerk (recorder) data sets.

All three data sources are then placed into a database(s) and cross-indexed usually on APN number between the three data sources.

(The above usually holds true for the DISCLOSURE states. There usually are NO COMMON INDICES for the NON-DISCLOSURE states. This means that matching these involves using addresses, legal descriptions, names etc., and the classical free-text UNIQUE matching problem ensues. This is why the NON-DISCLOSURE states are under-represented by most AVM's).

^TOP 

At the AVM design level the MINIMUM number of fields necessary are:

Situs address | situs city | land_sqft | zip | Gross living Area (GLA) | current year assessed values land & GLA | and total Taxes by land & GLA| sale price | sale date |

Most all providers of data for AVM's generally provide the above.

The best AVM's databases would also have, in addition to the above:

APN | total rooms | bedrooms | baths | half_baths | style | stories | census_tracts |census_blocks| condition/grade | subdivision name/code | latitude | longitude | two years of assessed taxes / assessed values by land value and improvement value as well as TOTAL assessed value and taxes | county clerk file number | type of instrument | filing date

Problem:

Even with a minimum data set, as given above, not all records will contain the significant fields such as street number, street name, sale price, sale date.

Therefore, these records will be unavailable to the AVM unless strong theoretical statistical methods regarding missing data elements are utilized and whether or not these can be used is data-dependent. These require long development times, testing times, model generation times, etc. that would usually exceed the time line for annual updates from the three sources. (Such methods are problematical since most lenders that use AVM's only want actual data as provided by Assessor, Sales and County Clerks but interesting enough, lenders do accept multiple linear regression as a GENERAL AVM Method since that method is also used by Human Appraisers and is sanctioned by the Appraisal Institute).

At this point the AVM providers now have a data base, either minimum or maximum from the three sources, and receive updates at some frequency form the data sources ranging from one year on master assessor records to generaly one month on sales and County Clerk records and a methodology for updating same.

Now the AVM providers begin to develop their models across all data sets and in some rare cases a model for each county data set because the underlying model is fragile. (This would generally be a poor AVM).

How they do this depends upon their underlying model of choice. If a Neural Net is involved they begin training runs against the known sales prices, dates, county, address, zip codes, and land, GLA. This produces, in essence, a statistical model involving Neural Net Algorithms (essentially specialized statistical methodologies). Both model building and model generation/reliability testing is achieved during the training/testing phase.

If a replacement cost method AVM is used, base tables for each property style, construction/grade, exterior walls, etc. are used. Marshall & Swift is the industry standard and such data can be licensed from them. Then it is merely a matter of taking the square footage GLA, M&S tables, and generating a replacement cost. This generally yields a PREDICTED SALES price in the +/- 4-10% accuracy range, where "accuracy" means deviation from the ACTUAL SALES PRICE, but it is highly property data specific and requires a much larger database.

^TOP 

A second step involving Multiple Linear Regression to fine-tune the Replacement Cost AVM is often used. Here such issues as ranges of home GLA's, land square footage, total rooms, bedrooms, baths, half_baths, subdivision sales performance, etc. are further refined to create an "adjustment factor" of the base Predicted Sales price by property characteristic. This results in an Adjusted Predicted Sales Price within a +/ 3-5% range. (Lenders only require a +/- 7% variance from Human Appraisers of the actual sales price).

A Third Method combines Replacement Cost, Comparable Sales, Statistical Methods, and Artificial Intelligence using an EXPERT SYSTEM as opposed to a the Neural Net AI approach. This approach is the MOST ACCURATE METHOD. (To our knowledge ONLY RDI USES THIS APPROACH).

In this approach, all possible property and land characteristics, sales, subdivision, location, and other factors that a HUMAN Appraiser would use in EACH of the Classical Appraisal Methods has been KNOWLEDGE ENGINEERED by having, say, 10 MAI's value the same 25 properties and capturing their DECISION RULES into the Expert System as well as conflict resolving heuristic rules between methods, or involving MISSING DATA in one comparable to another as well as rules to SELECT COMPARABLES.

Another level involving statistical methods for such items as SUBDIVISION performance where historical sales values are calculated yearly by subdivision, section (from the legal descriptions) are used to fine tune valuation.

The result is an AVM that adjusts to missing data, can select comparables or more than a mere 1/2 mile radius from the subject based only on date and GLA (as do most other AVM methods since they lack DECISION RULES for selecting comparables and do not MAXIMIZE the subject property characteristics against comparables AND use DECISION RULES to reduce variance).

The result is an AVM that can run anywhere in the U.S. that produces a PREDICTED SALES PRICE WITHIN +/- 2% OF THE ACTUAL.

Not only that, it can also POSTDICT VALUES, meaning that by selecting a date range say five years earlier, it can produce the SALES PRICE ON THAT DATE.

Obviously in a short discussion such as this details regarding adjustment factors by the Neural Net, Replacement Cost Method, and Comparables Method or combined with an Expert System cannot be addressed in detail. Suffice to say that FEW AVM's can use more than one method. Look to the results. RDI's AVM OUTPERFORMS ALL OTHER AVM's as to ACCURACY.

All AVM's have to match addresses where TaxID's are not known. This poses an additional layer of difficulty regarding standardization since most AVM's take batch input via the Internet or individual keyed input and then must "find" the equivalent addresses. (Assume both the assessor address and the input address differ from USPS data). To the degree that this occurs it accounts for the "address not found" or "no valuation" return.

A final note, any AVM that gives a High range - Low range of value and something such as a "confidence level" should be suspect. What should appear is the PREDICTED VALUE, not a RANGE. The range is a statistical "fudge" based upon the holes in the AVM method attempting to adjust to missing data. What this means is that the AVM provider's DATABASE is INCOMPLETE and the AVM Method CANNOT VALUE so it distrusts itself. Wonderful!!!

At RDI we have the knowledge and expertise in Artificial Intelligence, Address parsing, Expert Systems, Database design and build, AVM model building as well as proprietary systems and algorithms to custom build an AVM. Call us for Consulting Rates and Turn-Key pricing.

The AVM provider industry is a money maker on the Internet with few competitors. RDI can deliver a Turn-Key AVM that has unparalled accuracy and speed. The AVM cash flow is impressive and ranges from $5.00 to $20 per valued property with some lenders running in excess of 20,000 valuations per month.