Automated-valuation-model training-data optimization systems and methods

• US 9,582,819 B2
• Filed: 09/26/2013
• Issued: 02/28/2017
• Est. Priority Date: 09/26/2013
• Status: Active Grant

• Current Assignee: Greenfield Advisors LLC
• Original Assignee: Greenfield Advisors LLC
• Inventors: Krause, Andy, Lipscomb, Clifford A., Kilpatrick, John A.
• Primary Examiner(s): Chen, Alan
• Application Number:
  US14/038,627
• Publication Number:
  US 20150088766A1
• Time in Patent Office: 1,251 Days
• Field of Search: None
• US Class Current: 1/1
• CPC Class Codes: G06Q 30/0278 Product appraisal
• CPC Class Codes: G06Q 50/16 Real estate

Abstract

To optimize training data used by a predictive real-estate valuation model, a search space having multiple dimensions may be defined. Each search dimension corresponds to a range of candidate values for a search criterion for selecting subsets of sales-transaction records. The multiple dimensions include a temporal dimension and a geographic dimension. An accuracy-optimized subset of a multiplicity of sales-transaction records is identified by evaluating points that vary along each dimension within the multi-dimension search space. A statistical measure of model accuracy is used to evaluate each candidate point. The accuracy-optimized subset of the multiplicity of sales-transaction records is provided to a predictive model to generate an automated value prediction for a subject real-estate property as of an effective date.

Claims

1. A server-device-implemented method for optimizing training data for developing a predictive model to automatically value a subject real-estate property, the method comprising:

obtaining, by one or more server devices, an indication to provide an automated value prediction for the subject real-estate property as of an effective date;

in response to obtaining said indication, defining, by said server device, a search space having multiple dimensions, each corresponding to a range of candidate values for a search criterion for selecting subsets of a multiplicity of sales-transaction records, said multiple dimensions including at least a temporal dimension that measures distance in time before the effective date, and a geographic dimension that measures distance in space away from the subject real-estate property;

evaluating, by said one or more server devices, a multiplicity of points within said multi-dimension search space, said multiplicity of points varying along at least said temporal dimension and said geographic dimension;

selecting, by said one or more server devices based at least in part on evaluating said multiplicity of points within said multi-dimension search space, a first majority subset of said multiplicity of sales-transaction records as a first statistical accuracy data structure according to a first search criteria set;

developing, by said one or more server devices, the predictive model according to said first majority subset of said multiplicity of sales-transaction records;

obtaining, by said one or more server devices, a first statistical measure of the predictive model applied to said first majority subset of said multiplicity of sales-transaction records;

selecting, by said one or more server devices based at least in part on evaluating said multiplicity of points within said multi-dimension search space, a second majority subset of said multiplicity of sales-transaction records as a second statistical accuracy data structure according to a second search criteria set;

developing, by said one or more server devices, the predictive model according to said second majority subset of said multiplicity of sales-transaction records;

obtaining, by said one or more server devices, a second statistical measure of the predictive model applied to said second majority subset of said multiplicity of sales-transaction records;

determining that the second statistical measure of the predictive model applied to said second majority subset of said multiplicity of sales-transaction records indicates greater accuracy than the first statistical measure of the predictive model applied to said first majority subset of said multiplicity of sales-transaction records; and

developing, by said one or more server devices in response to said greater accuracy, the predictive model according to said second subset of said multiplicity of sales-transaction records to generate said automated value prediction for the subject real-estate property as of the effective date.

2. The method of claim 1, wherein defining said multi-dimension search space comprises:

sampling a temporal range to obtain a plurality of candidate temporal-distance values; and

sampling a geographic range to obtain a plurality of candidate geographic-distance values.

3. The method of claim 2, wherein evaluating a candidate point of said multiplicity of points comprises:

determining said first search criteria set comprising a candidate temporal-distance value selected from said plurality of candidate temporal-distance values and a candidate geographic-distance value selected from said plurality of candidate geographic-distance values;

selecting said first majority subset of said multiplicity of sales-transaction records corresponding to said first search criteria set;

developing the predictive model according to at least said first majority subset of sales-transaction records; and

statistically measuring a first accuracy of the predictive model thereby developed, said first accuracy being less than the greater accuracy indicated by the second statistical measure of the predictive model applied to said second majority subset of said multiplicity of sales-transaction records.

4. The method of claim 1, wherein said first and second statistical measure each comprises a forecast standard deviation statistical measure.

5. The method of claim 1, wherein said first and second statistical measure each is selected from a group consisting of:

a mean average percent error statistical measure;

a mean absolute error statistical measure; and

a root mean square error statistical measure.

6. The method of claim 1, wherein the predictive model is selected from a group consisting of:

a regression model;

an adaptive estimation model; and

a neural network model.

7. A computing apparatus for optimizing training data for developing a predictive model to automatically value a subject real-estate property, the apparatus comprising a processor and a memory storing instructions that, when executed by the processor, configure the apparatus to:

obtain an indication to provide an automated value prediction for the subject real-estate property as of an effective date;

in response to obtaining said indication, define a search space having multiple dimensions, each corresponding to a range of candidate values for a search criterion for selecting subsets of a multiplicity of sales-transaction records, said multiple dimensions including at least a temporal dimension that measures distance in time before the effective date, and a geographic dimension that measures distance in space away from the subject real-estate property;

evaluate a multiplicity of points within said multi-dimension search space, said multiplicity of points varying along at least said temporal dimension and said geographic dimension;

select, based at least in part on evaluating said multiplicity of points within said multi-dimension search space, a first majority subset of said multiplicity of sales-transaction records as a first statistical accuracy data structure according to a first search criteria set;

develop, by said one or more server devices, the predictive model according to said first majority subset of said multiplicity of sales-transaction records;

obtain, by said one or more server devices, a first statistical measure of the predictive model applied to said first majority subset of said multiplicity of sales-transaction records;

select, by said one or more server devices based at least in part on evaluating said multiplicity of points within said multi-dimension search space, a second majority subset of said multiplicity of sales-transaction records as a second statistical accuracy data structure according to a second search criteria set;

develop, by said one or more server devices, the predictive model according to said second majority subset of said multiplicity of sales-transaction records;

obtain, by said one or more server devices, a second statistical measure of the predictive model applied to said second majority subset of said multiplicity of sales-transaction records;

determine that the second statistical measure of the predictive model applied to said second majority subset of said multiplicity of sales-transaction records indicates greater accuracy than the first statistical measure of the predictive model applied to said first majority subset of said multiplicity of sales-transaction records; and

develop, in response to said greater accuracy, the predictive model according to said second subset of said multiplicity of sales-transaction records to generate said automated value prediction for the subject real-estate property as of the effective date.

8. The apparatus of claim 7, wherein the instructions that configure the apparatus to define said multi-dimension search space further comprise instructions configuring the apparatus to:

sample a temporal range to obtain a plurality of candidate temporal-distance values; and

sample a geographic range to obtain a plurality of candidate geographic-distance values.

9. The apparatus of claim 8, wherein evaluating a candidate point of said multiplicity of points comprises:

determine a multi-dimension search criteria set comprising a candidate temporal-distance value selected from said plurality of candidate temporal-distance values and a candidate geographic-distance value selected from said plurality of candidate geographic-distance values;

select a distinct subset of said multiplicity of sales-transaction records corresponding to said multi-dimension search criteria set;

develop the predictive model according to at least a majority of said distinct subset of sales-transaction records; and

statistically measure an accuracy of the predictive model thereby developed.

10. The apparatus of claim 7, wherein the instructions that configure the apparatus to said first and second statistical measure further comprise instructions configuring the apparatus to a forecast standard deviation statistical measure.

11. The apparatus of claim 7, wherein said first and second statistical measure each is selected from a group consisting of:

a mean average percent error statistical measure;

a mean absolute error statistical measure; and

a root mean square error statistical measure.

12. The apparatus of claim 7, wherein the predictive model is selected from a group consisting of:

a regression model;

an adaptive estimation model; and

a neural network model.

13. A non-transitory computer-readable storage medium having stored thereon instructions that, when executed by a processor, configure the processor to:

obtain an indication to provide an automated value prediction for a subject real-estate property as of an effective date;

in response to obtaining said indication, define a search space having multiple dimensions, each corresponding to a range of candidate values for a search criterion for selecting subsets of a multiplicity of sales-transaction records, said multiple dimensions including at least a temporal dimension that measures distance in time before the effective date, and a geographic dimension that measures distance in space away from the subject real-estate property;

evaluate a multiplicity of points within said multi-dimension search space, said multiplicity of points varying along at least said temporal dimension and said geographic dimension;

select, based at least in part on evaluating said multiplicity of points within said multi-dimension search space, a first majority subset of said multiplicity of sales-transaction records as a first statistical accuracy data structure according to a first search criteria set;

develop, by said one or more server devices, the predictive model according to said first majority subset of said multiplicity of sales-transaction records;

obtain, by said one or more server devices, a first statistical measure of the predictive model applied to said first majority subset of said multiplicity of sales-transaction records;

select, by said one or more server devices based at least in part on evaluating said multiplicity of points within said multi-dimension search space, a second majority subset of said multiplicity of sales-transaction records as a second statistical accuracy data structure according to a second search criteria set;

develop, by said one or more server devices, the predictive model according to said second majority subset of said multiplicity of sales-transaction records;

obtain, by said one or more server devices, a second statistical measure of the predictive model applied to said second majority subset of said multiplicity of sales-transaction records;

determine that the second statistical measure of the predictive model applied to said second majority subset of said multiplicity of sales-transaction records indicates greater accuracy than the first statistical measure of the predictive model applied to said first majority subset of said multiplicity of sales-transaction records; and

develop, in response to said greater accuracy, the predictive model according to said second subset of said multiplicity of sales-transaction records to generate said automated value prediction for the subject real-estate property as of the effective date.

14. The non-transitory computer-readable storage medium of claim 13, wherein defining said multi-dimension search space comprises:

sample a temporal range to obtain a plurality of candidate temporal-distance values; and

sample a geographic range to obtain a plurality of candidate geographic-distance values.

15. The non-transitory computer-readable storage medium of claim 14, wherein evaluating a candidate point of said multiplicity of points comprises:

determine a multi-dimension search criteria set comprising a candidate temporal-distance value selected from said plurality of candidate temporal-distance values and a candidate geographic-distance value selected from said plurality of candidate geographic-distance values;

select a distinct subset of said multiplicity of sales-transaction records corresponding to said multi-dimension search criteria set;

develop the predictive model according to at least a majority of said distinct subset of sales-transaction records; and

statistically measure an accuracy of the predictive model thereby developed.

16. The non-transitory computer-readable storage medium of claim 13, wherein said first and second statistical measure each comprises a forecast standard deviation statistical measure.

17. The non-transitory computer-readable storage medium of claim 13, wherein said first and second statistical measure each is selected from a group consisting of:

a mean average percent error statistical measure;

a mean absolute error statistical measure; and

a root mean square error statistical measure.

18. The non-transitory computer-readable storage medium of claim 13, wherein the predictive model is selected from a group consisting of:

a regression model;

an adaptive estimation model; and

a neural network model.