Monday, January 07, 2013

Fun With Predictors

Here is a fun predictor I use to guess near-term (i.e. 4-6 months out) price movements on the Teranet HPI. The model works by finding peak correlation between price changes and months-of-inventory, the best is shown below:



I then use a simple regression fit (I'm ignoring log fit for now) and play around with the regression window length and get minimized error post-2009 with an 8-month window. Since peak cross correlation is shifted we have the ability to "predict" price changes some time into the future, in this case 3 months ahead:
This is then translated onto a year-on-year prediction:
On the graph above I get an estimate of close to -5.5% YOY change in the March Teranet HPI. I have been calling for between -6% and -4% on the Teranet HPI in February-March of 2013.

4 comments:

Maggie May said...

What is the confidence interval around that point estimate?

Etienne de Cochon said...

Jesse, I was going to call you the Nate Silver of Van RE pricing but that would be a disservice to you. Clearly, Nate Silver is the Jesse of US Election Polling.

Etienne de Cochon said...

Jesse, I was going to call you the Nate Silver of Van RE pricing but that would be a disservice to you. Clearly, Nate Silver is the Jesse of US Election Polling.

jesse said...

Shock, not super high, I haven't run confidence intervals on the regression. But to eyeball it I'll say +/-1% at 80%.

Correlation is -0.9 on HPI-MOI, if that helps.