Neural networks can be used to predict the sale price of a home. The information provided by the neural network helps appraisers make assessments, helps sellers determine appropriate asking prices, and helps homeowners decide if improvements would be cost-effective. As the neural network designer, your primary responsibilities are to clearly define the problem and present the data in such a way that the network can find patterns. Once this is accomplished, training the network is mostly a fine-tuning process.
The Problem
Traditional methods for determining the value of real estate include appraisal by a certified expert, computer-assisted appraisal and, of course, the actual sale price. The problems inherent with these valuation methods are the inconsistency between appraisers, the inability of machines to consider more than rules and mathematical formulas, and the effects of changing market conditions.
Neural networks do not fall victim to these problems. When applied to real estate appraisal, neural networks are able to predict the actual sale price of properties with 90% accuracy. Neural networks perform better than multi-variate analysis, since networks are inherently nonlinear. They can also evaluate subjective information, such as a neighborhood rating, which is difficult to incorporate into traditional mathematical approaches. Richard Borst, a Senior Vice President at Day & Zimmerman, Inc., the nation's leading provider of mass appraisal services to state and local governments, has successfully trained a neural network to appraise real estate in the New York area. His network incorporates eighteen data items which include the number of dwelling units, fireplaces, plumbing fixtures, square feet of living area, age, months since last sale, and air conditioning. He uses 217 sales records from 1988 and 1989 with prices ranging from $103,000 to $282,000. His network was trained on a 386 using BrainMaker Professional v2.5 (California Scientific: Nevada City, CA).
The Data
The data used in Mr. Borst's network, collected by the mass appraisal firm, Cole-Layer-Trumble, represent sales from a single area. The data chosen are similar to what an appraiser would examine to make an assessment. The table below lists all variables used in the original network design and the range of possible values for each. All values are continuous except two, heating type and neighborhood group. These two inputs represent categories, but since each has only two possible values, they don't need to be divided into separate inputs.
| Name | Description | Range |
|---|---|---|
| SALEPRIC | Actual sales price of home | $103,000-250,000 |
| DWLUN | Number of dwelling units | 1-3 |
| RDOS | Reverse (months since sale) | 0-23 |
| YRBLT | Year built | 1850-1986 |
| TOTFIXT | Number of plumbing fixtures | 5-17 |
| HEATING | Heating system type | coded as A or B |
| WBFPSTKS | Wood burning fireplace stacks | 0-1 |
| BMNTGAR | Basement garage | 0-2 |
| ATTFRGAR | Attached frame garage area | 0-228 |
| TOTLIVAR | Total living area | 714-4185 |
| DECK/OFP | Deck / open porch area | 0-738 |
| ENCLPOR | Enclosed porch area | 0-452 |
| NBHDGRP | Neighborhood group | coded as A or B |
| RECROOM | Recreation room area | 0-672 |
| FINBSMT | Finished basement area | 0-810 |
| GRADE% | Grade factors | 0.85-1.08 |
| CDU | Condition / desirability / useful | 3-5 |
| TOTOBY | Total other value (bldg & yard) | 0-16400 |
Reference
R. A. Borst, Artificial Neural Networks: The Next Modeling / Calibration Technology for the Assessment Community?, Property Tax Journal (International Association of Assessing Officers), 10(1):69-94, 1991.