Good Morning
Dr. Michael Furick
–Faculty member at Georgia Gwinnett College, School of Business
–Teach Management Information Systems
and Marketing
Today’s Topic
Using neural networks to develop decision support systems to chose tenants for apartment rental.
Results of a pilot study
If we asked about rental property…
Owning rental property is the best financial decision you will ever make
and
Owning rental property is the worst financial decision you will ever make
Three levels of apartment tenants
Good tenants…….heaven on earth
Bad tenants……….hell on earthworry
No tenant………….empty unitworry and lose money
Fear of
this
Causes this
Picking “good” tenants is vital to rental business success
and sanity
Many decisions about tenants get made every year
34 million households live in rental housing (held steady due to immigration)
20% renters above $60k income 20% renters below $10k income 56% of rental units owned by
individuals
How do other industries pick “customers”
Most rely on credit reports and credit scoring to predict consumer financial behavior Banks Car dealers Mortgage brokers etc
What is a credit report and score?
Credit report- a multi page report that profiles a consumer’s financial transactions
Credit score- mathematical means of summarizing the credit report into a three digit number
Credit score is widely used because it is predictive and easy
Delinquency Rates by FICO Mortgage Risk Score
87%
71%
51%
31%
15%5% 2% 1%
0%
20%
40%
60%
80%
100%
below499
500 to549
550 to599
600 to649
650 to699
700 to749
750 to799
over 800
FICO Score Range
Rate of Delinquencies
Success has caused credit score use to spread to other industries
Auto industry uses credit score to Determine who gets auto insurance What price to charge for an auto policy
Two studies found that a lower credit score means
Up to 50% more accidents Bigger claims ($918 vs. $558)
Credit data and credit scores should work in apartment rentals
Two Part Study First part of study looked at tenant
performance vs. six commercially available credit scores (statistical analysis)
Second Part: If credit is not predictive then what is predictive? (neural network analysis)
Credit data and credit scores should work in apartment rentals
First part of study looked at tenant performance vs. six commercially available credit scores
22 different credit scores are available from Experian
Six scores tested from Experian
FICO Mortgage Risk Score FICO Advanced Risk Score
Derogatory credit in 24 months FICO Installment Loan Score
Repay short term loans auto etc. FICO Finance Score
Loans from non-traditional sources National Risk Score Sureview Non-Prime score
(non-prime bankcard applicants)
Correlation examined: credit score vs. tenant performance
Data collected One apartment complex 200 tenants that moved in during 2002 6 scores collected on each tenant Tenant performance followed in
satisfying lease over 12 months
Traditional statistical methods used to examine correlation
Results Part 1: credit data not predictive of tenant performance
No correlation between credit data and tenant performance in satisfying the terms of their lease
R square approaching zero
“We have as much trouble with people with good credit as we do with people with bad
credit” property manager quote
Why are commercial scoring models not predictive in selecting tenants?
? ? ? ? ? ? Many “good working” models filter
out consumers with Less job tenure High ratios of debt to income Older vehicles
What would be predictive in Part 2?
Hints from the decision process used in the apartment rental industry
Picking tenants more complex than picking customers
Financial consideration Non-financial considerations
Non financial consideration affected by Fair Housing Laws
Decision process mostly manual with a range of data and big dose of “gut feel”
96 units Baltimore Reject if landlord problems or criminal
Reject if bankruptcy
395 units Chicago Credit score in top 15%
Reject if landlord problems or criminal
264 units Chattanooga Reject is landlord problem or criminal
Reject if bankruptcy
210 units Athens, Ga. Income 3 times monthly rent
80% satisfactory accounts
Reject if landlord problems or criminal
68 units Washington D.C Reject if landlord problem or criminal
Reject if bankruptcy
Nationally, property managers make rental decisions on a range of items
33.8% ran criminal backgrounds 62.6% ran credit reports 65.5% called references
Rental Property Reporter
50.6% ran credit reports 52% verified income 75.5% relied on personal interviews
U.S. Census Bureau
Opportunity to standardize decision making with a Decision Support Model
Data to be a mix of financial and non-financial items (matching current decision process)
Apartment managers suggested 76 possible variables
Sample of data elements used in neural network model 7 From out-of-state (Application) 11 Size of employer (Chamber of Commerce) 12 Number of years with employer
(Application) 15 Income (Verification) 20 Number of people to occupy apartment
(Application) 34 Type of vehicle one (Application) 35 Age of vehicle one (Application) 48 Estimated monthly installment loan payments
(Credit Report) 68 Number of driving infractions (DMV report) 73 Information found on county criminal search
Data collection process
One apartment complex Data elements collected on 60
tenants as they moved in during 2004
Tenants lease performance tracked over 12 months
Why use neural networks to create this model
Neural network – an artificial intelligence system that is good at finding and differentiating patterns
modeled after the brain’s mesh-like network of interconnected processing elements (neurons)
Why use neural networks to create this model
NNs good with unstructured data how do data elements interact with
each other or with the output Analyze nonlinear relationships Learn and adjust to new
circumstances
Layers of a Neural Network
Input Layer Hidden Layer Output Layer
Why use Palisade’s NeuralTools®
Over 50 NN software packages Evaluated about a dozen Feature, function, benefit
Actual model creation details not covered here
Data divided into test and training data
Model run several hundred times using various combinations of variables
Prediction accuracy recorded and analysis completed
What did the model find?
Model accurately predicted 69.1% of tenants (good and bad )
Three data elements became most important in choosing tenants
1. Percent satisfactory accounts on credit report
2. Total applicant income3. Driving record of applicant
Comments on driving record as predictor in apartment rentals
Auto Industry
Credit Performanc
eDriving Record
Predicts
This Study
Credit Performance
Driving Record
Predicts
Limitations with this pilot NN study
Small data set Single geographic region (one
apartment complex) Data set of those who moved in
(sample selection)
Next Step for researchers
Proposal submitted to National Science Foundation to fund expansion of the study to the Southeastern U.S.
Thanks to Palisade Corporation for hosting the conference
Thank you for attending
Questions now and later
Dr. Michael Furick
Copies of the detailed result and model are available for purchase from
www.il.proquest.com Document UMI number: 3215298 Citation: Using neural networks to develop a new model
to screen applicants for apartment rentals. Furick, Michael T., PhD. 2006.