Real estate appraisal using predictive modeling

Real estate abstract

An automated real estate appraisal system (100) and method generates estimates of real estate value using a predictive model such as a neural network (908). The predictive model (908) generates these estimates based on learned relationships among variables describing individual property characteristics (905) as well as general neighborhood characteristics at various levels of geographic specificity (906). The system (100) may also output reason codes indicating relative contributions (1009) of various variables to a particular result, and may generate reports (701) describing property valuations, market trend analyses, property conformity information, and recommendations regarding loans based on risk related to a property.

Real estate claims

What is claimed is:

1. A computer-implemented process for appraising a real estate property, comprising the steps of:

collecting training data;

developing a predictive model from the training data;

storing the predictive model;

obtaining individual property data for the real estate property;

generating a signal indicative of an appraised value for the real estate property responsive to application of the obtained individual property data to the stored predictive model;

developing an error model from the training data;

storing the error model; and

generating a signal indicative of an error range for the appraised value responsive to application of the individual property data to the stored error model.

2. The computer-implemented process of claim 1, wherein the error model comprises a regression model.

3. A computer-implemented process for appraising a real estate property, comprising the steps of:

collecting training data;

developing a predictive model from the training data;

storing the predictive model;

obtaining individual property data for the real estate property;

generating a signal indicative of an appraised value for the real estate property responsive to application of the obtained individual property data to the stored predictive model;

developing a lower percentile error model from the training data;

developing an upper percentile error model from the training data;

storing the lower percentile error model;

storing the upper percentile error model;

generating a signal indicative of a lower bound value for the real estate property responsive to application of the obtained individual property data to the stored lower percentile error model; and

generating a signal indicative of an upper bound value for the real estate property responsive to application of the obtained individual property data to the stored upper percentile error model.

4. The computer-implemented process of claim 3, wherein:

the lower percentile error model is a computer-implemented neural network; and

the upper percentile error model is a computer-implemented neural network.

5. A computer-implemented process for appraising a real estate property, comprising the steps of:

obtaining individual property training data describing past real estate sales;

aggregating the obtained property training data into area training data sets, each area training data set describing a plurality of sales within a geographic area;

developing a predictive model from the training data;

storing the predictive model;

obtaining individual property data for the real estate property; and

generating a signal indicative of an appraised value for the real estate property responsive to application of the obtained individual property data to the stored predictive model.

6. The computer-implemented process of claim 5, wherein the step of aggregating is repeated using successively larger geographic areas until the number of sales within the geographic area over a predetermined time period exceeds a predetermined number.

7. A computer-implemented process for appraising a real estate property, comprising the steps of:

collecting training data;

performing the iterative substeps of:

applying input data to a predictive model;

ranking output data produced thereby responsive to a measure of quality; and

adjusting operation of the model responsive to the results of the ranking substep;

storing the predictive model;

obtaining individual property data for the real estate property; and

generating a signal indicative of an appraised value for the real estate property responsive to application of the obtained individual property data to the stored predictive model.

8. The computer-implemented process of claim 7, wherein the predictive model comprises a computer-implemented neural network having a plurality of interconnected processing elements, each processing element comprising:

a plurality of inputs;

a plurality of weights, each associated with a corresponding input to generate weighted inputs;

combining means, coupled to the weighted inputs, for combining the weighted inputs; and

a transfer function, coupled to the combining means, for processing the combined weighted inputs to produce an output.

9. A computer-implemented process for appraising a real estate property, comprises the steps of:

selecting a geographic area surrounding the real estate property;

obtaining area data for the geographic area;

collecting training data;

developing a predictive model from the training data;

storing the predictive model;

obtaining individual property data for the real estate property; and

generating a signal indicative of an appraised value for the real estate property responsive to application of the obtained individual property data and the obtained area data to the stored predictive model.

10. The computer-implemented process of claim 9, further comprising the steps of:

developing an error model from the training data; storing the error model; and

generating a signal indicative of an error range for the appraised value responsive to application of the individual property data to the stored error model.

11. The computer-implemented process of claim 10, wherein the error model comprises a regression model.

12. The computer-implemented process of claim 9, further comprising the steps of:

developing a lower percentile error model from the training data;

developing an upper percentile error model from the training data;

storing the lower percentile error model;

storing the upper percentile error model;

generating a signal indicative of a lower bound value for the real estate property responsive to application of the obtained individual property data to the stored lower percentile error model; and

generating a signal indicative of an upper bound value for the real estate property responsive to application of the obtained individual property data to the stored upper percentile error model.

13. The computer-implemented process of claim 12, wherein:

the lower percentile error model is a computer-implemented neural network; and

the upper percentile error model is a computer-implemented neural network.

14. A computer-implemented process for appraising a real estate property, comprising the steps of:

collecting training data;

developing a predictive model from the training data;

storing the predictive model;

obtaining individual property data for the real estate property, the individual property data comprising a plurality of elements;

generating a signal indicative of an appraised value for the real estate property responsive to application of the obtained individual property data to the stored predictive model; and

for each element of the individual property data:

determining a relative contribution of the element to the appraised value;

determining from each relative contribution a reason code value; and

generating a signal indicative of the reason code value.

15. A system for appraising a real estate property, comprising:

a predictive model for determining an appraised value for the real estate property;

training data input means, coupled to the predictive model, for obtaining training data;

training data aggregation means, coupled to the training data input means, for aggregating the training data into training data sets, each training data set describing a plurality of sales within a geographic area;

a model development component, coupled to the predictive model, for training the predictive model from the training data;

a storage device for storing the trained predictive model;

individual property data input means, coupled to the predictive model, for obtaining individual property data and sending the individual property data to the predictive model;

area data input means, coupled to the individual property data input means and to the predictive model, for selecting a geographic area surrounding the real estate property, obtaining area data, and sending the area data to the predictive model; and

an output device, coupled to the predictive model, for generating a signal indicative of the appraised value.

16. The system of claim 15, wherein the predictive model comprises a neural network.

17. The system of claim 15, further comprising:

an error model for determining an error range for the appraised value;

and wherein:

the training data input means is coupled to the error model;

the model development component trains the error model from the training data;

the storage device stores the trained error model;

the individual property data input means is coupled to the error model and sends the individual property data to the error model;

the area data input means is coupled to the error model and sends the area data to the error model; and

the output device generates a signal indicative of the error range.

18. The system of claim 17, wherein the error model comprises a regression model.

19. The system of claim 15, further comprising:

a lower percentile error model for determining an lower bound for the appraised value;

an upper percentile error model for determining an upper bound for the appraised value;

and wherein:

the training data input means is coupled to the error model;

the model development component trains the lower percentile error model and the upper percentile error model from the training data;

the storage device stores the trained lower percentile error model and the trained upper percentile error model;

the individual property data input means is coupled to the lower percentile error model and the upper percentile error model, and sends the individual property data to the lower percentile error model and the upper percentile error model;

the area data input means is coupled to the lower percentile error model and the upper percentile error model and sends the area data to the lower percentile error model and the upper percentile error model; and

the output device generates a signal indicative of the lower bound and the upper bound for the appraised value.

20. The system of claim 19, wherein:

the lower percentile error model comprises a neural network; and

the upper percentile error model comprises a neural network.

Real estate description

CROSS-REFERENCE TO RELATED APPLICATION

The subject matter of this application is related to the subject matter of pending U.S. application Ser. No. 07/814,179, for "Neural Network Having Expert System Functionality", by Curt A. Levey, filed Dec. 30, 1991, the disclosure of which is incorporated herein by reference.

The subject matter of this application is further related to the subject matter of pending U.S. application Ser. No. 07/941,971, for "Fraud Detection Using Predictive Modeling", by Krishna M. Gopinathan et al., filed Sep. 8, 1992, the disclosure of which is incorporated herein by reference.

37 C.F.R.1.71 AUTHORIZATION

A portion of the disclosure of this patent document contains material which is subject to copyright protection. The copyright owner has no objection to the facsimile reproduction by anyone of the patent document or the patent disclosure, as it appears in the Patent and Trademark Office records, but otherwise reserves all copyright rights whatsoever.

BACKGROUND OF THE INVENTION

1. Field Invention

This invention relates generally to real estate appraisals and sales price predictions. In particular, the invention relates to an automated real estate appraisal system and method that uses predictive modeling to perform pattern recognition and classification in order to provide accurate sales price predictions.

2. Description of Related Art

The "appraised value" of a real estate parcel, or property, comprises some estimate of the full market value of the property on a specified date. A property's appraised value is of great importance in many types of real estate transactions, including sales and loans.

Conventionally, appraised value is determined by a professional appraiser using both objective and subjective factors. One disadvantage of such a method is the difficulty in ensuring that the appraiser conducts a neutral, unbiased analysis in arriving at the appraised value. This difficulty is often compounded by the fact that the appraiser may be retained and paid by an interested party in the contemplated transaction, such as a lender, mortgage broker, buyer, or seller.

In order to reduce bias and provide more accurate appraisals, statistical techniques may be used to obtain an independent, consistent, mathematically derived estimate of a property's value to assist an appraiser in generating an appraised value. Traditional statistical techniques, such as multiple linear regression and logistic regression, have been tried, but such techniques typically suffer from a number of deficiencies. One deficiency is the inability of traditional regression models to capture complex behavior in predictor variables resulting from nonlinearities and interactions among predictor variables. In addition, traditional regression models do not adapt well to changing trends in the data, so that automated model redevelopment is difficult to implement.

One example of the difficulty of applying a regression model to appraisal problems is the uncertainty as to the optimal temporal and geographical sample size for model development. A model developed using all homes in one square city block might theoretically be an effective predictor for that particular neighborhood, but it may not be possible to develop such a model with sufficient stability and reliability, due to the relatively small sample size. On the other hand, a model developed using all homes sold in the United States in the past month might have a sufficiently large sample size, but might be unable to capture local, neighborhood characteristics to provide an accurate appraisal. Thus, a significant deficiency of traditional regression modeling techniques when applied to real estate appraisals is the inability to successfully model neighborhood characteristics while including a sufficiently large sample size to develop a robust, stable statistical model.

It is desirable, therefore, to have an automated system that uses available information regarding real estate properties to provide accurate estimates of value. Preferably, such a system should be flexible enough to allow model development in a relatively small geographic area, it should be able to handle nonlinearities and interactions among predictor variables without advance specification, it should have high predictive accuracy, and it should have capability for redevelopment of the underlying system model as new patterns of real estate pricing emerge.

SUMMARY OF THE INVENTION

In accordance with the present invention, there is provided an automated system (100) and method for real estate appraisals, which uses one or more predictive models such as neural networks (908) to generate estimates of real estate value. The predictive models (908) generate these estimates based on learned relationships among variables describing individual property characteristics (905). The models (908) also learn relationships between individual property characteristics (905) and area characteristics (906). Area characteristics (906) are stored and applied at a level of geographic specificity that varies according to the amount of data available at each of several successively larger geographic areas. In this way the models (908) are able to capture local neighborhood characteristics without unduly reducing sample sizes, which would reduce reliability and predictability.

The learned relationships among individual property characteristics (905) and area characteristics (906) enable the system (100) to estimate the value of the property being appraised. Error models (909) may also be provided to generate an estimated value range or error interval for the sales price. The appraised value and error estimate may then be provided as output (907) to a human decision-maker, along with other related information such as: reason codes that reveal the relative contributions of various factors to the appraised value; and various measures of market trends. Finally, the system (100) periodically monitors its performance, and redevelops the models (908,909) when performance drops below a predetermined level.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block diagram of an implementation of the present invention.

FIG. 2 is a sample data entry screen that forms part of a typical input/output interface for the present invention.

FIG. 3 is a sample quick data entry screen that forms part of a typical input/output interface for the present invention.

FIG. 4 is a sample record selection screen that forms part of a typical input/output interface for the present invention.

FIG. 5 is a sample sales price estimate screen that forms part of a typical input/output interface for the present invention.

FIG. 6 is a sample area averages screen that forms part of a typical input/output interface for the present invention.

FIG. 7 is a sample report produced by the present invention.

FIG. 8 is a flowchart illustrating the major functions and operation of the present invention.

FIG. 9 is a block diagram showing the overall functional architecture of the present invention.

FIG. 10 is a block diagram showing the property valuation process of the present invention.

FIG. 11 is a flowchart showing a method of determining area and obtaining area data according to the present invention.

FIG. 12 is a flowchart showing a method of generating reports according to the present invention.

FIG. 13 is a flowchart showing a method of performing market trend analysis according to the present invention.

FIG. 14 is a flowchart showing a method of determining property conformity according to the present invention.

FIG. 15 is a flowchart showing a method of comparing an estimated property value to user-specified values according to the present invention.

FIG. 16 is a flowchart showing a method of generating recommendations according to the present invention.

FIG. 17 is a diagram showing an example of geographic subdivision according to the present invention.

FIG. 18 is a flowchart showing a method of aggregating individual property data into successively larger geographical areas according to the present invention.

FIG. 19 is a diagram showing an example of the relationship between individual property characteristics and area characteristics.

FIG. 20 is a diagram of a single processing element within a neural network.

FIG. 21 is a diagram illustrating hidden processing elements in a neural network.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

The Figures depict preferred embodiments of the present invention for purposes of illustration only. One skilled in the art will readily recognize from the following discussion that alternative embodiments of the structures and methods illustrated herein may be employed without departing from the principles of the invention described herein.

Referring now to FIG. 1, there is shown a block diagram of a typical implementation of a system 100 in accordance with the present invention. The user supplies property data to system 100 via input device 105. Central processing unit (CPU) 101 runs software program instructions, stored in program storage 107, which direct CPU 101 to perform the various functions of system 100. In the embodiment illustrated herein, the software program is written in the Microsoft Excel macro language and the ANSI C language. Each of these languages may be run on a variety of conventional hardware platforms. Data storage 103 contains data describing real estate properties, as well as regional data. It also contains model parameters. In accordance with the software program instructions, CPU 101 accepts input from input device 105, accesses data storage 103, and uses RAM 102 in a conventional manner as a workspace. CPU 101, data storage 103, and program storage 107 operate together to provide predictive neural network models 908 for real estate appraisal, as well as error models 909 for generating error ranges for the appraised values. If desired, multiple models 908 and 909 may be used (for example, one for each geographic region), particularly when property pricing characteristics vary widely from region to region. After neural network models 908 and error models 909 process the information, as described below, to obtain estimates of property value and error range, a signal indicative of the estimate and error range is sent from CPU 101 to output device 104.

In the embodiment illustrated herein, CPU 101 can be a mainframe computer or a powerful personal computer; RAM 102 and data storage 103 are conventional RAM, ROM and disk storage devices for the CPU; and output device 104 is a conventional means for either printing results based on the signals generated by neural network models 908 and error models 909, displaying the results on a video screen using a window-based interface system, or sending the results to a database for later access.

Referring now also to FIGS. 2 through 7, there are shown sample screens from a conventional window-based interface system (not shown) that forms part of output device 104. FIG. 2 shows data entry form 201 that allows the user to enter data describing a property for appraisal. Form 201 is also known as a Uniform Residential Appraisal Report (URAR) form. It contains a number of data fields 202. Scroll bars 203 are provided to allow navigation throughout form 201.

FIG. 3 shows quick data entry form 301 that allows quick entry of property data without using a complete URAR form 201. This form is intended for use when a quick estimate of property value is required. A number of fields 302 are provided, which represent a subset of the fields 202 in URAR form 201. Data entered on URAR form 201 for a particular property is automatically transferred to quick form 301, and vice versa.

FIG. 4 shows record selection screen 401 that allows the user to select among previously-entered property records in order to view URAR form 201 for the selected record. Record selection screen 401 lists a plurality of records 402, showing the address 403, city 404, map reference 405, sale price 406, assessor parcel number (APN) group 407, ZIP code 408, and sale date 409 for each record. Scroll bars 410 are provided to allow navigation throughout the list of records, and selected records are indicated by highlighting 411.

FIG. 5 shows sales price estimate screen 501 that provides appraisal information. Estimated sales price 502 is shown, along with lowest typical sales price 503 and highest typical sales price 504. Also shown are positive contribution factors 505 that tend to drive the price of the property up, and negative contribution factors 506 that tend to drive the price of the property down.

FIG. 6 shows area averages screen 601 that shows average values 602 for several property criteria 603 for a selected geographic area, alongside comparative values for a selected property 605. Clicking arrow buttons 606 changes the level of geographic specificity, according to the following sequence: neighborhood, local, extended, region, and county. The example shows neighborhood values, representing the average values for all properties sold in the same neighborhood as the selected property, over a period of time prior to the selected sale.

Referring now to FIG. 7, there is shown statistical review report 701 summarizing property information and estimated value, and providing recommendations regarding loan processing with respect to the property. This type of report would typically be used when system 100 of the present invention is employed to appraise properties in connection with loan processing. Identification portion 702 identifies the loan, property, and appraisal to which report 701 pertains. Explanatory portion 703 gives general explanatory information concerning the report. Regional trend analysis portion 704 reports average sales prices for the county and ZIP code in the four preceding semiannual periods, indicating market stability and providing a broad foundation for valuation and risk analysis. Local trend analysis portion 705 reports average sales prices for smaller geographic areas, such as the census tract, map grid, and assessor parcel number (APN) group, in the four preceding semiannual periods, indicating local market stability and providing a further information useful for valuation and risk analysis. Subject conformity portion 706 compares sales price, square footage, and price per square foot for the property with the norms for the neighborhood. Subject valuation portion 707 provides a value range for the property based on the characteristics of the property and the region, and compares the value range with an appraisal value determined by an independent human appraiser. Subject valuation portion 707 also provides an indication of the loan-to-value (LTV) ratio of the loan, and a comparison with a user-supplied maximum LTV ratio. Summary and recommendations portion 708 summarizes the information given in the other portions and recommends one of "Proceed", "Caution", or "Suspend".

Referring now to FIG. 8, there is shown an overall flowchart illustrating the major functions and operation of system 100. First neural network models 908 are trained 801 using training data describing a number of individual real estate properties, characteristics, and prices, as well as area characteristics. If real estate pricing characteristics vary widely from region to region, it may be advantageous to use different models 908 for the different regions (counties, for example). Once neural network models 908 are trained, neural network model parameters are stored 802. Error models 909, which are typically regression models, are then trained 803 using additional training data and output from neural network models 908. Once error models 909 are trained, error model parameters are stored 804, and system 100 is able to estimate prices and pricing errors for a subject property. System 100 obtains 805 property data describing the subject property 905, as well as data describing the area in which the subject property is situated 906. System 100 then applies 806 property data 905 and area data 906 to the appropriate stored neural network model 908. It then applies 807 property data 905 and area data 906 to the appropriate stored error model 909. The models 908 and 909 estimate sales price, reason codes (described below), and estimated error, which are output 805 to the user, or to a database, or to another system via output device 104.

Referring now to FIG. 9, the overall functional architecture of system 100 is shown. System 100 is broken down into two major components: model development component 901 and property valuation component 902. Model development component 901 uses training data 904 describing a number of real estate properties, characteristics, and prices to build neural network models 908 containing information representing learned relationships among a number of variables. Together, the learned relationships form models 908 of the behavior of the variables. Although neural network models 908 are used in the embodiment illustrated herein, any type of predictive modeling technique may be used, such as regression modeling. For purposes of illustration, the invention is described here in terms of neural network statistical models 908. Model development component 901 also uses training data 904 to develop error models 909, which are typically regression models used to estimate error in predicted sales prices generated by neural network models 908.

Property valuation component 902 feeds input data describing the subject property 905 and its geographic area 906 to neural network models 908 and error models 909. It obtains results from models 908 and 909 and generates price estimates, error ranges, and reason codes. A report is prepared using this information, and the report is output 907 either to a screen display, printer, or stored in a database for future access.

Each of the two components 901 and 902 of system 100 will be described in turn.

Model Development Component 901

Neural networks employ a technique of "learning" relationships through repeated exposure to data and adjustment of internal weights. They allow rapid model development and automated data analysis. Essentially, such networks represent a statistical modeling technique that is capable of building models 908 from data containing both linear and non-linear relationships. While similar in concept to regression analysis, neural networks are able to capture nonlinearity and interactions among independent variables without pre-specification. In other words, while traditional regression analysis requires that nonlinearities and interactions be detected and specified manually, neural networks perform these tasks automatically. For a more detailed description of neural networks, see D. E. Rumelhart et al, "Learning Representations by Back-Propagating Errors", Nature v. 323, pp. 533-36 (1986), and R. Hecht-Nielsen, "Theory of the Backpropagation Neural Network", in Neural Networks for Perception, pp. 65-93 (1992), the teachings of which are incorporated herein by reference.

Neural networks comprise a number of interconnected neuron-like processing elements that send data to each other along connections. The strengths of the connections among the processing elements are represented by weights. Referring now to FIG. 20, there is shown a diagram of a single processing element 2001. The processing element receives inputs X.sub.1, X.sub.2, . . . X.sub.n, either from other processing elements or directly from inputs to the system. It multiplies each of its inputs by a corresponding weight w.sub.1, w.sub.2, . . . w.sub.n and adds the results together to form a weighted sum 2002. It then applies a transfer function 2003 (which is typically non-linear) to the weighted sum, to obtain a value Z known as the state of the element. The state Z is then either passed on to another element along a weighted connection, or provided as an output signal. Collectively, states are used to represent information in the short term, while weights represent long-term information or learning.

Processing elements in a neural network can be grouped into three categories: input processing elements (those which receive input data values); output processing elements (those which produce output values); and hidden processing elements (all others). The purpose of hidden processing elements is to allow the neural network to build intermediate representations that combine input data in ways that help the model learn the desired mapping with greater accuracy. Referring now to FIG. 21, there is shown a diagram illustrating the concept of hidden processing elements. Inputs 2101 are supplied to a layer of input processing elements 2102. The outputs of the input elements are passed to a layer of hidden elements 2103. Typically there are several such layers of hidden elements. Eventually, hidden elements pass outputs to a layer of output elements 2104, and the output elements produce output values 2105.

Neural networks learn from examples by modifying their weights. The "training" process, the general techniques of which are well known in the art, involves the following steps:

1) Repeatedly presenting examples of a particular input/output task to the neural network model;

2) Comparing the model output and desired output to measure error; and

3) Modifying model weights to reduce the error.

This set of steps is repeated until further iteration fails to decrease the error. Then, the network is said to be "trained." Once training is completed, the network can predict outcomes for new data inputs.

In the present invention, data used to train models 908 are drawn from various database files containing data on individual properties. These data are aggregated to obtain medians and variances across geographic areas. Thus, models 908 are able to capture relationships among individual property characteristics, as well as relationships between individual property characteristics and the characteristics of the surrounding geographic area.

Referring now to FIG. 19, there is shown an example of this technique. House 1901 has associated with it individual property characteristics 1902, namely 2500 square feet, 3 bedrooms, and a 6000 square foot lot. In order to provide effective predictive modeling of the estimated selling price of house 1901, neural network models 908 use area characteristics 1904 for geographic area 1903. The area characteristics 1904 are 2246.4 average square feet, 2.5 average bedrooms, 7267.2 average square foot lot size, and $267,000 average selling price. These represent averages for homes sold in area 1903 in the last x months, where x is a predetermined time period. By comparing individual property characteristics 1902 with area characteristics 1904, neural network models 908 are able to more effectively estimate the selling price of house 1901.

An important factor in the effectiveness of neural network models 908 is the sample size used to train models 908. Conventional regression models and model designs for real estate appraisal generally use small samples in an attempt to provide a homogeneous group of properties in the developmental sample. See J. Mark & M. Goldberg, "Multiple Regression Analysis and Mass Assessment: A Review of the Issues", The Appraisal Journal v. 56(1)., pp. 89-109 (1988), and H.-B. Kang & A. Reichert, "An Empirical Analysis of Hedonic Regression and Grid-Adjustment Techniques in Real Estate Appraisal", AREUEA Journal v. 19, no. 1, pp. 70-91 (1991), the teachings of which are incorporated herein by reference. For example, properties within a single city block would generally provide effective predictor models for capturing neighborhood characteristics within the block. A problem with this approach is that a large number of distinct models must be built. Since each model is created using a set of training data describing properties within the associated city block, an extremely large number of properties is required to effectively train all the models.

On the other hand, use of larger geographic areas such as ZIP codes results in diminished ability to capture local neighborhood characteristics.

The method of the present invention provides effective predictor variables that preserve information describing neighborhood characteristics without unduly increasing the number of models and predictor variables required for training. It accomplishes this by aggregating individual property data in the training data set into area characteristics in a flexible manner, using the smallest geographic areas containing sufficient data to produce reliable models 908. The models 908 are thus able to capture area characteristics for relatively small geographic areas where the data describing these characteristics are available.

Referring now to FIG. 17, there is shown an example of geographic subdivision according to the present invention. Each region 1701 is divided into successively smaller geographic areas. In the example shown, the geographic areas are ZIP codes 1702, census tracts 1703, map coordinates 1704, and assessor parcel number (APN) groups 1705. Other geographic areas, such as census blocks, or lot blocks, may also be used.

Referring now also to FIG. 18, there is shown a flowchart of the aggregation method. System 100 uses data describing real estate sales activity for each month of a user-specified training period, such as eighteen months. For each month within the training period, system 100 performs the steps shown in FIG. 18. System 100 initially defines 1804 the "neighborhood" as the smallest geographical area, such as the APN group 1705. If there have been any sales in the previous 12 months 1805, system 100 proceeds to step 1813. If not, it defines 1806 the neighborhood as the next larger geographic area, the map code 1704. If there have been at least 3 sales in the previous 12 months 1807, system 100 proceeds to step 1813. System 100 continues to enlarge the definition of the neighborhood until a predetermined minimum number of sales have occurred within a predetermined period of time. The minimum number of sales and the period of time in steps 1805, 1807, 1809, and 1811 may vary according to the optimal sample size and geographic specificity required. In addition, the number and size of the geographic areas may vary. Once the predetermined minimum number of sales for a particular geographic area has been met, system 100 determines 1813 medians, averages, and variances for various property characteristics such as sales price, square feet, number of bedrooms, etc.

Property characteristics used as predictor variables in the embodiment illustrated herein include, for example:

PREDICTOR VARIABLES IN AREAS MODEL

AIR.sub.-- COND: type of air conditioning

BA.sub.-- FLCND: condition of bathroom floor

BA.sub.-- FLMAT: bathroom floor material

BA.sub.-- NUM: number of bathrooms

BA.sub.-- WNCON: condition of bathroom wainscot

BA.sub.-- WNMAT: bathroom wainscot material

BEDRMS.sub.-- N: number of bedrooms

FRPL.sub.-- NUM: number of fireplaces

FRPL.sub.-- TYP: type of fireplace

FL.sub.-- ZONE: flood zone?

FLR.sub.-- MAT: main floor material

FND.sub.-- INF: foundation infestation?

FND.sub.-- SETL: foundation settlement?

IMP.sub.-- TYPE: improvement type (attached, townhouse, etc.)

LANDSCAP: adequate landscaping?

MAN.sub.-- HOME: manufactured house?

OWN.sub.-- TYPE: ownership type (condo, single family residence, etc.)

P.sub.-- COND: condition of parking structure

P.sub.-- SPACES: number of parking spaces

P.sub.-- STRAGE: type of parking (garage, carport, etc.)

P.sub.-- DOROPN: electric garage door opener?

ROOFTYPE: type of roofing material

R.sub.-- TOT.sub.-- N: number of rooms

LOT.sub.-- SHAP: lot shape

SITE.sub.-- INF: site influence (ocean, mountains, etc.)

SI.sub.-- STM: public or private street maintenance?

SI.sub.-- STT: street surface material

PARCLSIZ: size of parcel (typical, undersized or oversized)

SQ.sub.-- FT.sub.-- LA: square footage of living area

STRYSFDU: number of stories

STYLECOD: style of house (colonial, ranch, etc.)

TOPOCODE: topography of lot (level, hilly, etc.)

WAL.sub.-- EXTT: exterior wall material

POOLTYPE: pool, spa, both, or none

AGE: age of home

HOA: home owner's dues?

ECONLIFE: economic life (remaining years) of house

LN.sub.-- LOT: natural log of lot size

APN.sub.-- COMP: median comps price (local neighborhood)

ZIP.sub.-- COMP: median comps price (zip code or county wide)

ZIPCMPBE: median # of bedrooms in comps (zip code or county-wide)

ZIPCMPBA: median # of bathrooms in comps (zip code or county-wide)

ZIPCMPSQ: median square footage in comps (zip code or county-wide)

ZIPCMPAG: median age in comps (zip code or county-wide)

ZIPCMPRM: median # of rooms in comps (zip code or county-wide)

ZIPCMPGA: median # of parking spaces in comps (zip code or county-wide)

ZIPCMPFP: median # of fireplaces in comps (zip code or county-wide)

APNDIFAG: age differential (current property minus local comps)

APNDIFBA: # bathrooms differential (current property minus local comps)

APNDIFBE: # bedrooms differential (current property minus local comps)

APNDIFFP: # fireplaces differential (current property minus local comps)

APNDIFGA: # park spaces differential (current property minus local comps)

APNDIFRM: # rooms differential (current property minus local comps)

APNDIFSQ: sq. footage differential (current property minus local comps)

Once system 100 has obtained predictor variables as described above for each month in the training period, the predictor variables are fed to networks 908 and networks 908 are trained. The embodiment illustrated herein uses a modeling technique known as a backpropagation neural network 908. This type of network 908 estimates parameters which define relationships among variables using a training method. The preferred training method, well known to those skilled in the art, is called "backpropagation gradient descent optimization" and is described in Gopinathan et al., although other well-known neural network training techniques may also be used.

Once the neural networks have been trained using training data, the network model definitions are stored in data files in a conventional manner. These data files describe the neural network architecture, weights, the data configuration, data dictionary, and file format.

In addition to developing and storing neural network models 908 as described above, model development component 901 also develops error models 909. Typically, these error models 909 are implemented as conventional regression models, known to those skilled in the art, although other predictive modeling-techniques, such as neural networks, may be used. As with neural network models 908, different error models 909 may be provided for different regions.

To develop error models 909, system 100 determines the absolute percent error of the neural network model estimate for each record in the training data set. Based on a set of input parameters, error model 909 is trained to forecast the absolute percent error of the neural network model estimate. Training methods for regression models are well known in the art. An example of a set of input parameters used in the embodiment illustrated herein is given below:

PREDICTOR VARIABLES IN AREAS ERROR MODEL

PRED.sub.-- SP: predicted sales price

PRED.sub.-- SP2: square of PRED.sub.-- SP

PRED.sub.-- SP3: cube of PRED.sub.-- SP

APNPDIF: normalized difference between PRED.sub.-- SP and local median price

APNPDIF2: square of APNPDIF

APNPDIF3: cube of APNPDIF

ZIPPDIF: normalized difference between PRED.sub.-- SP and zip code median price

APNSRC: size of local neighborhood

AIR.sub.-- COND: type of air conditioning

BA.sub.-- FLMAT: bathroom floor material

BA.sub.-- WNCON: bathroom wainscot condition

FND.sub.-- INF: foundation infestation?

FRPL.sub.-- TYP: type of fireplace

IMP.sub.-- TYPE: improvement type (attached, townhouse, etc.)

MAN.sub.-- HOME: manufactured house?

OWN.sub.-- TYPE: ownership type (condo, single family residence, etc.)

PARCLSIZ: size of parcel (typical, undersized or oversized)

POOLTYPE: pool, spa, both, or none

P.sub.-- COND: condition of parking structure

P.sub.-- DOROPN: electric garage door opener?

P.sub.-- STRAGE: type of parking (garage, carport, etc.)

SI.sub.-- STM: public or private street maintenance?

TOPOCODE: topography of lot (level, hilly, etc.)

WAL.sub.-- EXTT: exterior wall material

AGE: age of home

AGE2: square of AGE

AGE3: cube of AGE

APNDIFAG: difference between age of home and local median age

APNDFAG2: square of APNDIFAG

APNDFAG3: cube of APNDIFAG

APNDFBE3: cube of difference between # bedrooms and local median

APNDFFP2: square of difference between # fireplaces and local median

APNDFFP3: cube of difference between # fireplaces and local median

APNDIFGA: difference between # parking places and local median

APNDFGA2: square of APNDIFGA

APNDFRM2: square of difference between # rooms and local median

APNDFSQ3: cube of difference between square footage and local median

BA.sub.-- NUM: number of bathrooms

ECONLIFE: economic life (remaining years) of house

ECONLIF3: cube of ECONLIFE

PSPACES3: cube of number of parking spaces

R.sub.-- TOT N: total number of rooms

R.sub.-- TOT N2: square of R.sub.-- TOT.sub.-- N

R.sub.-- TOT N3: cube of R.sub.-- TOT.sub.-- N

SQ.sub.-- FT.sub.-- LA: square footage

ZIPCMPAG: difference between age of home and zip code median age

ZIPCMAG3: cube of ZIPCMPAG

ZIPCMPBA: difference between # bathrooms and zip code median

ZIPCMBA2: square of ZIPCMPBA

ZIPCMPBE: difference between # bedrooms and zip code median

ZIPCMBE2: square of ZIPCMPBE

ZIPCMBE3: cube of ZIPCMPBE

ZIPCMFP2: square of difference between # fireplaces and zip code median

ZIPCMFP3: cube of difference between # fireplaces and zip code median

ZIPCMGA2: square of difference between # parking places and zip code median

ZIPCMSQ2: square of difference between square footage and zip code median

ZIPCMSQ3: cube of difference between square footage and zip code median

ZIP.sub.-- COMP: median sales price across zip code

ZIPCOMP2: square of ZIP.sub.-- COMP

ZIPCOMP3: cube of ZIP.sub.-- COMP

Property Valuation Component 902

As seen in FIG. 9, property valuation component 902 reads data 905 describing the property to be appraised (known as subject property) and data 906 describing the surrounding geographical area 906, and generates as output 907 a price estimate for the subject property. Furthermore, property valuation component 902 outputs a range of values based on the estimated maximum error of the price estimate, as well as a list of contributing variables, or reason codes, for the price estimate.

Property data 905 are generally entered by the user on a data entry form such as those shown in FIGS. 2 and 3. The data may be entered either interactively, or in batch mode using tape or disk storage devices. Property data 905 describe the particular property to be appraised, and they typically include the same types of predictor variables as listed above for training data 904.

Area data 906 are collected from databases describing properties in geographical areas surrounding the subject property. The method by which area data 906 are collected is described below. Typically, area data 906 include averages of the same types of predictor variables as listed above for training data 904.

Referring now to FIG. 10, there is shown a flowchart of the property valuation process of the present invention. System 100 obtains 1001 property data 905 from user input or batch input. Based on property data 905, it determines 1001 the appropriate region to be used for the analysis. As shown in FIG. 17, a region 1701 is a relatively large geographic area containing a number of smaller geographic areas. Each region 1701 may be associated with a separate neural network model 908, as well as a separate error model 909. System 100 then loads 1003 neural network model 908 and error model 909 for the region 1701 containing the subject property.

System 100 then determines 1005 which area to use in the analysis and obtains area data 906. In determining 1005 the optimal area for the analysis, system 100 uses a technique that captures local neighborhood characteristics while including a sufficiently large sample size to preserve predictability and reliability. Generally, system 100 accomplishes this by seeking the smallest geographic area containing both the subject property and at least one other property that was sold within the past x months, where x is a predetermined time period.

Referring now also to FIG. 11, there is shown the method of determining 1005 the optimal area to use. As shown in the flowchart, system 100 uses the smallest geographic area containing at least one property that was sold within the past x months, in addition to the subject property. The minimum number of properties, the time period, and the particular areas available for use, may vary depending on the level of geographic specificity and sample size desired.

System 100 applies 1006 the appropriate neural network 908 to subject property data 905 and area data 906. Neural network 908 generates estimated value 1007. System 100 then determines 1008 reason codes indicating which inputs to model 908 are most important in determining estimated value 1007. Any technique to generate such reason codes may be used. In the embodiment illustrated herein, the technique set forth in co-pending U.S. application Ser. No. 07/814,179, for "Neural Network Having Expert System Functionality", by Curt A. Levey, filed Dec. 30, 1991, the disclosure of which is hereby incorporated by reference, is used. System 100 uses the reason codes to generate 1009 a list of contribution factors to the estimated value, shown in FIG. 5 as positive contribution factors 505 and negative contribution factors 506.

System 100 also estimates 1010 the error range of its appraisal. In the embodiment illustrated herein, error estimation is performed by applying error model 909, typically a regression model, to subject property data 905 and area data 906. Error model 909 uses conventional regression techniques to generate an absolute percent error estimate E. System 100 generates a lower bound and an upper bound for the error range by applying the following formulas:

where P is the estimated property value and E is the absolute percent error estimate.

Alternatively, system 100 may estimate the error range using a technique known in the art as robust backpropagation, as described in H. White, "Supervised Learning as Stochastic Approximation", International Joint Conference on Neural Networks, San Diego, Calif. (1990), the teachings of which are incorporated herein by reference.

When robust backpropagation is used, system 100 does not require error model 909. Rather, two additional secondary neural network models 908 are used. Each of the two secondary models 908 is trained to estimate a specified percentile of the conditional distribution of sales prices. For example, the first secondary model 908 may be trained to estimate the 10th percentile of the conditional distribution, while the second secondary model 908 estimates the 90th percentile. These models 908 are trained and implemented in the same technique and using the same predictor values as described above for neural network model 908. When estimating a sales price for a subject property, the property data is sent to secondary models 908 in addition to primary neural network model 908. Secondary models 908 produce lower and upper bounds L and U for the error range.

Whichever technique is used to generate the lower and upper bounds L and U, system 100 then outputs 1011 the estimated property value, as well as the range defined by L and U designated in FIG. 5 as a lowest typical sales price 503 and a highest typical sales price 504.

Finally, system 100 generates 1012 reports as appropriate and as requested by the user. A typical report is statistical review report 701 shown in FIG. 7.

Referring now to FIG. 12, there is shown a method of generating 1012 reports according to the present invention. System 100 analyzes 1202 market trends as shown in FIG. 13. It first determines 1302 and 1303 county and ZIP code sales price trends over the past 24 months. Then it determines 1304 local sales price trends by census tract, map code, and APN group over the past 24 months. It classifies 1305 trends as stable, moderate upward or downward trend, or steep upward or downward trend. The trends and their classifications are used in generating regional trend analysis portion 704 and local trend analysis portion 705 of statistical review report 701 shown in FIG. 7. If alternative geographical subdivisions are used, the above-described method of market trend analysis is altered accordingly.

System 100 then determines 1203 the degree of conformity of the subject property with regard to the neighborhood. This is done according to the method shown in FIG. 14. System 100 determines the median and variance of neighborhood sales prices 1402, square footages 1403, and prices per square footage 1404. Medians and variances for other variables may be collected as well, if desired. The distribution within the neighborhood is used in generating subject conformity portion 706 of statistical review report 701. System 100 determines 1405 whether the property deviates by more than one standard deviation from the neighborhood norm. If not, system 100 classifies 1406 the property as conforming. If the property deviates by more than one standard deviation, system 100 determines 1407 if the property deviates by more than two standard deviations. If not, system 100 classifies 1408 the property as non-conforming. If the property deviates by more than two standard deviations, system 100 classifies 1409 the property as extremely non-conforming. Additional levels of conformity classification may be provided. System 100 uses the conformity classification in generating summary and recommendations portion 708 of statistical review report 701.

System 100 generates 1204 subject valuation portion 707 based on the estimated value determined by neural network 108.

System 100 generates 1205 summary and recommendations portion 708 using the method shown in FIGS. 15 and 16. Referring now to FIG. 15, there is shown the method of comparing an estimated property value to user-specified values. System 100 determines 1502 whether the appraised value as determined by a human appraiser falls within the valuation range generated by neural network model 908 and error model 909. If not, system 100 determines the percent outside the range and outputs 1503 this value in summary and recommendations portion 708. System 100 then determines 1504 if the loan-to-value (LTV) ratio, based on the estimated value of the property and the amount of the contemplated loan, is within a user-specified maximum LTV. If not, system 100 determines the percent above the maximum and outputs 1505 this value in summary and recommendations portion 708.

Referring now to FIG. 16, there is shown the method of generating recommendations. System 100 determines 1602 whether the LTV is within the maximum LTV. If not, it recommends suspension of the loan 1603. If the LTV is within the maximum LTV, system 100 determines 1604 whether the property is conforming. If the property is not conforming, system 100 determines 1605 whether the property is extremely non-conforming. If the property is not extremely non-conforming, system 100 recommends caution with regard to the contemplated loan 1606. If the property is extremely non-conforming, system 100 recommends suspension of the loan 1607. If the property is conforming, system 100 determines 1608 whether the market is declining. If so, it recommends caution 1609. If the market is not declining, system 100 determines 1610 whether the appraisal as performed by the human appraiser falls within the range generated by neural network model 908 and error model 909. If the appraisal does not fall within the range, system 100 recommends caution 1611. If the appraisal falls within the range, system 100 recommends that the loan proceed 1612. System 100 outputs its recommendation as part of summary and recommendation portion 708 of statistical review report 701.

As an additional disclosure, the source code for the embodiment illustrated herein of the invention is included below as an appendix. It should be noted that terminology in the source code may differ slightly from that in the remainder of the specification. Any differences in terminology, however, will be easily understood by one skilled in the art.

From the above description, it will be apparent that the invention disclosed herein provides a novel and advantageous method of real estate appraisal. The foregoing discussion discloses and describes merely exemplary methods and embodiments of the present invention. As will be understood by those familiar with the art, the invention may be embodied in many other specific forms without departing from the spirit or essential characteristics thereof. For example, other predictive modeling techniques besides neural networks might be used. In addition, other variables, geographic subdivisions, and report generation techniques might be used.

Accordingly, the disclosure of the present invention is intended to be illustrative of the preferred embodiments and is not meant to limit the scope of the invention. The scope of the invention is to be limited only by the following claims. ##SPC1##