Presented by Scott Sambucci (Altos Research) at the IE Group’s Predictive Analytics Summit, "Predicting the Housing Market with Buyer & Seller Psychology" reviews real-time and leading indicators to predict future home prices.
Predicting the Housing Market with Buyer & Seller Psychology (Predictive Analytics Summit 2011)
1. PREDICTING THE HOUSING MARKET
WITH BUYER & SELLER PSYCHOLOGY
PREDICTIVE ANALYTICS INNOVATION SUMMIT
Scott Sambucci, Market Analytics
Altos Research
February 2011 | San Diego, CA
2. WHO, WHAT, WHERE, WHEN, WHY & HOW
• Altos Research
• Real-time housing market analytics
• 200+ metros, 20,000 zip codes
• Not what happened but what is happening & what
will happen
• Wisdom of crowds
• Using Active Market to forecast future market
activity
3. TODAY’S ACTIVE MARKET SIGNALS FUTURE
TRANSACTION PRICES
Home listed
$429,000 Buyer financing fails,
Inventory 49 Neighbor home listed Property relisted
$409,000 $394,000
Deal closed
Price reduced $389,000
$398,000
Inventory 69
Offer made Transaction Recorded
$391,000
March May July Sept Nov Jan
Closed transaction = 1 data point, months too late
Active Market = 9 months of pricing, price changes, supply and demand, leading indicators
7. SELLER RESPONSE TO MARKET DEMAND:
LISTING PRICE HISTORY - 3204 TRAWLER PLACE
8. MIAMI CONDOS: MARKET IMPROVED IN 2009?
Did values really
move higher when
the actives knew the
market was weak?
9. DEVELOPING A FORWARD VALUATION MODEL™:
ENSEMBLE OF REGRESSION TREES WITH VARIABLE
IMPORTANCE
• Non-parametric
• Capture non-linearity & variable interactions
• Embarrassingly parallel
• Train quickly using the cloud
• Bootstrap holdouts
• Use all of our data for training
• Information gain feature selection
• Most important active market statistics
• How entropic is the distribution of the dependent variable with each
variable, and then without?
• Modified ARCH model – hybrid of transaction and active market prices
10. BENCHMARK MODEL RESULTS
• Average change in median price over the period was -0.53%
• Second pass, multivariate OLS regression. Model
had R2 = 29.4%, and mean absolute error (MAE) = 0.60%
• This error completely swamps the signal, but why?
• Residuals (errors) show bias and extreme non-linearity:
3.0%
2.0%
1.0%
0.0%
-1.0%
-2.0%
11. ALTOS LPI™ TIME-SERIES RESULTS
Out-of-Sample-Out-of-Time results for Mesa, AZ:
0.0%
Change in Median Price (+3M)
Forecast Change in LPI (+3M)
-2.0%
-4.0%
-6.0%
-8.0%
-10.0%
-12.0%
-14.0%
12. FVM™ IN-SAMPLE SCATTERPLOT
San Mateo, CA 94401 / Single-family-home / All Quartile / 6M
x-axis is +6M change in LPI, y-axis is forecast +6M change in LPI
8.0%
6.0%
4.0%
2.0%
0.0%
-10.0% -8.0% -6.0% -4.0% -2.0% 0.0% 2.0% 4.0% 6.0% 8.0% 10.0%
-2.0%
-4.0%
-6.0%
-8.0%
13. FVM™ OUT-OF-SAMPLE SCATTERPLOT
San Mateo, CA 94401 / Single-family-home / All Quartile / 6M
x-axis is +6M change in LPI, y-axis is forecast +6M change in LPI
8.0%
6.0%
4.0%
2.0%
0.0%
-10.0% -8.0% -6.0% -4.0% -2.0% 0.0% 2.0% 4.0% 6.0% 8.0% 10.0%
-2.0%
-4.0%
-6.0%
-8.0%