| City: |
Mexico, Mexico
|
| Organization: |
Data Science for Social Good (Eric & Wendy Schmidt, Summer Fellowship), The University of Chicago |
| Project Start Date: |
May 2015 |
| Project End Date: |
December 2015 |
| Reference: |
Eric & Wendy Schmidt:Data Science for Social Good: Summer Fellowship : The University of Chicago / https://dssg.uchicago.edu/2015/08/13/infonavit-project-reducing-home-abandonment KDD 2016: in:
· Proceeding
KDD '16 Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining
Pages 13-20
San Francisco, California, USA — August 13 - 17, 2016
|
| Problem: |
An increasingly prevalent problem of home abandonment in Mexico: when a homeowner decides to leave his property
and forego his investment. A major causal factor of this
outcome is a mismatch between the homeowner's needs, in
terms of access to services and employment, and the location
characteristics of the home. |
| Technical Solution: |
Overall, abandonment risk prediction
problem was formulated as a binary classi cation problem where the outcome variable is whether a person abandons his house. This outcome was modeled for each year
after the loan is granted,either until the end of the observation period or until the
home is abandoned.
An alternative is to model the percentage of abandonment
for a speci c colonia. This would not serve objectives
since the analysis would no longer be focused on the individual decision of a person to abandon his house. Estimation
at the colonia level would also require aggregating individual
level data to a representative average person, and the information would be lost from individual level variation. Furthermore, as the speci c addresses of houses are unknown
for a large portion of the data, attempts to geocode these
houses were unsuccessful due to the low quality of the address fields. Model was created to answer the following question:
What is the risk of abandonment for an existing loan in the
next year?
A variety of machine learning models were trained on different years and then tested on active loans in the following
year (e.g. training from 2008 to 2014 and testing on loans
active during 2015). The algorithms tested were Support
Vector Machines, Random Forests, AdaBoost, and Logistic
Regression. Best performing model was a Random Forest which
achieved an AUC of 0.70. With a 0.5
threshold, the model captured 55% of abandoned houses.
|
| Datasets Used: |
- Dataset 1: Loans Data, Infonavit, 2015 (includes personal info, loan info, and house characteristics)
- Dataset 2: Housing Survey Data, ECUVE (Quantitative Evaluation of Housing and its Environment), 2015
- Dataset 3: Municipality Data, INEGI (National In-
stitute of Statistics and Geography), 2015
- Dataset 4: Business, School and Hospital Location Data, DENUE (National Statistical Index of Economic Units), 2015
|
| Outcome: |
Current process to control home abandonment is reactive and often too late.; Using 20 years
of mortgage history data combined with surveys, census,
and location information, a model was developed to predict
the probability of home abandonment based on both individual and location characteristics. The model was used to
develop a tool that provides Infonavit(the largest provider of mortgages in Mexico) the ability to give advice to Mexican workers when they apply for a loan, evaluate
and improve the locations of new housing developments, and
provide data-driven recommendations to the federal government to in
uence local development initiatives and infrastructure investments. The result is improving economic outcomes for the citizens of Mexico by pre-emptively identifying
at-risk home mortgages, thereby allowing them to be altered
or remedied before they result in abandonment. |
| Issues that arose: |
Most of the data availableis collected annually,
except for the Population Census, which occurs every five
years. This represented a challenge for feature creation. The loans dataset spans the last
20 years, but house coordinates were only available for loans
granted after 2008. Given that location features were critical to acheive objectives, the scope of the project was limited
to loans granted from 2008 to 2015. |
| Status: |
Operational |
| Entered by: |
November 19, 2017: Moteen Butt, moteen.butt@mail.utoronto.ca
|