Free Inflection Point Forecasting Model

[admin]

Maybe worth no more than its price!

The purpose of this model is to attempt to predict when resale homes in San Diego hit their trough price in the far-in-the-future upturn.

Caveats:

1.  I developed this model using data from 1988 to 1997. San Diego's last upturn occurred in late 1995. Thus, I only had one upturn to work with.

2.  The economic and housing circumstances that underlay the last upturn may (probably will) be different from the economic and housing circumstances that underlay the next upturn.

3.  This model is useful for predicting an upturn in prices, only.  The predictors for a downturn in prices are different.

Median price of a San Diego resale home =

$90,216

– (10.4 x notices of default 30 months prior)

+ (94.5 x employment (in thousands) 12 months prior)

+ (5.3 x sales of San Diego resale homes 3 months prior)

– (9.3 x notices of default 36 months prior)

Sources of data:

Notices of default: ideally, San Diego County Recorders Office. Alternatively, DataQuick quarterly news releases, divided by 3.

http://www.dqnews.com/CHFor0806.shtm

Employment: U.S. BLS

http://data.bls.gov/PDQ/servlet/SurveyOutputServlet;jsessionid=f0308f85b680$3F$3F$3

Sales of San Diego resale homes:  San Diego Union-Buffoon monthly news release

Note on notice of default data:  I think that this is the source, but I can't figure out how to properly select for homeowners notices of default under "Advanced Search."  Help, please!

http://arcc.co.san-diego.ca.us/arcc/services/grantorgrantee/search.aspx

 

Potential method of use:

1.  Record NODs, sales, and employment monthly (quarterly for NODs if relying on DataQuick data).

2.  Generate predictions for resale home prices.

3.  When the predicted prices consistently start moving up, buy.

Happy forecasting!

Sincerely,

Puritanical Bible-Beater

[powayseller]

When does this model show the inflection point? Does it provide a number or a buy/sell signal? I understand the inputs, but not how to use it.

[Anonymous]

PS, it provides the predicted price for a resale home. When the predicted price consistently moves up, buy, because it means you just passed the predicted trough.

Note — given the signs in front of the coefficients, for prices to be moving up, NODs will have to be coming down, sales moving up, and employment moving up.

See the red symbols? Those are the predicted prices using the formula/model. They fit reasonably nicely with the actual prices, in black.

[img_assist|nid=1671|title=
Prices: Actual vs. Fitted|desc=|link=node|align=left|width=400|height=267]

[powayseller]

jg, using this model, what are the buy/sell dates that it gives? Or am I misunderstanding what it does? According to the model, when would it have signaled that a person should be selling their home? 2004, 2005, or 2006? What month? When should a person have bough in the 90’s? 1995? 1996?

[Anonymous]

In my model, one of the factors is ‘sales of San Diego resale homes 3 months prior.’ I left out ‘deseasonalized.’ To deseasonalize raw sales data:

Monthly sales / monthly adjustment factor

Monthly adjustment factors:
January .758
February .730
March 1.073
April 1.029
May 1.091
June 1.136
July 1.117
August 1.182
September 1.029
October 0.971
November 0.923
December 0.962

Example: per DataQuick, May sales of resale San Diego homes comes in at 3,000. To deseasonalize such (as, historically, May sales are high):
3,000 / 1.091 = 2,750

Deseasonalizing data allows you to compare it month to month. If you didn’t deseasonalize the raw data, August sales would difficult to compare to September sales; August sales are almost always higher than September sales, historically. If you deseasonalize the August and September sales, you can make an apples-to-apples comparison.

[Anonymous]

PS, this model will do one thing, and one thing only: give you a buy signal. The buy signal is when the predicted prices arising from the model are consistently increasing.

Look at the graph. See how the predicted prices (red squares) are generally increasing in the second half of 1995/first half of 1996? That’s the signal to buy.

The actual trough price for resale homes occurred in December 1995. The predicted trough price for resale homes in my model occurred earlier, in July 1995. Given how both actual and predicted prices bounce around, it makes eminent sense to wait a bit to ensure that predicted prices are consistenly increasing. Using my model, an aggressive person may have felt comfortable buying in April ’96, when predicted prices had increased in 5 of the 9 past months (note: actual prices, over those same 9 months, increased in only 3 of those 9 months! Lots of bouncing around, up and down, in actual prices. Buying in the period immediately after the trough will be a leap of faith, because the actual sales prices may still be bouncing around up and down, and not consistenly moving up, yet).

The real value of this model, in my mind, is to bring out into the clear the three factors that may underly future firming of prices: increasing employment, decreasing NODs, and increasing (deseasonalized) sales.

[(former)FormerSanDiegan]

Hey, we closed on our first house in April 1996 and I used a different model.

Here it is:
My analysis : We bought because it was only about $200 per month (cheap car payement) more to buy a SFH in Clairemont compared to renting a townhome in Point Loma/OB. Much simpler model and it’s what the first-time buyer uses.

Still, I love the work your doing with this type of analysis.

[powayseller]

jg, good stuff. When was the sell signal in this cycle?

FormerSanDiegan, you’ve corroborated that prices move up when fundamentals are back in place: when home price = 8-10x income. It would be interesting to make a model of that, but the home rent data is probably not available.

[(former)FormerSanDiegan]

PS – Re: rent data. Just use your rent as a proxy. A renter is much more in tune with price of rents than data services.

JG – The “noise” on the filter appears to be about +/- 5000 out of 180K or about 3%. TO avoid falling victim to noise, one would have to wait for a significant move in the prediction relative to the noise. E.G. 2x the noise level or 6%. One could reduce the noise by averaging, but would give up the look-ahead capability. How far ahead is the prediction ahead of prices, assuming I need a 6% move that is confirmed over more than 1 month ?

Anyway – I don;t think hitting the exact bottom is important. Getting in a little late (e.g. 1997-1999) would not have been too bad, would it ?

[Anonymous]

Glad that you like the work, FSD. Good that you quickly recognize ‘manna from heaven’ during a pricing trough (i.e., that a $200/month premium is nothing for owning a home compared to renting a townhome; we moved from renting a condo to renting a home, and the rental premium is $1,000/month).

PS, I hate repeating myself, repeatedly; please see Caveat 3 and the first line of my response to you this morning, both above. For the third time: this model predicts the future trough price, only. The predictors for peak prices are different. We’re past the peak. I don’t care about the peak. The peak is ancient history. I’m solely interested in identifying the next trough, though it is years away.

FSD, I agree that the predicted prices are noisy, because the data that the model uses — NODs, employment, and deseasonalized sales — are noisy (i.e., move up and down month to month). Averaging, or buying on a 5%/6% trigger, as you suggest, may be useful. My model shows:
— Predicted trough price occurs in July 1995 at $165K. Actual trough price occurred in December 1995 at $163K.
— 5% above predicted trough price (of $165K) occurs in November 1995 (predicted November price = $173K, actual November price = $167K).
— 6% above predicted trough price (of $165K) occurs in August 1996 (predicted August price = $176K, actual August price = $174K).

[powayseller]

jg, it’s clever to use a different set of variables for the peak vs. the trough. Maybe you could put both models into one post, and explain it so that even dense people like me can understand it, and give the buy and sell dates so we can see how accurate it is.

When you say the actual price trough was in 12/95 at $168K, are you referring to the median price? Median price is a lagging indicator. Or are you using OFHEO data for that?

I think the inputs to this trough may be different from the last time, and I’d like your thoughts on that. I’m thinking of comments by the builders, who say that in past housing downturns, unemployment or high interest rates were the culprit, but this is the first housing downturn in an environment of low interest rates and high employment. We’ve also got a unique lending environment, where someone straight out of bankruptcy is getting a 60% DTI loan, 0% down, stated income, qualified based on the teaser rate, and whose payment can exceed their gross monthly income at reset. How will this unique lending environment affect the trough? Perhaps these loans will show up in the NOD and employment data that you already use.

[Anonymous]

PS, I abandoned work on a ‘predict peak’ model. It’s not difficult to do (I had a model that nicely modeled both peaks and troughs, except for the highest ‘greed’ peak, and I’d merely need to refocus and refine it), but I don’t need it, so I’m not going to work on it. I know the predictors for a peak model are different, because the predictors for my nicely fitting ’88-’03 up/down/up model were different from those in my ’90-’97 up only model.

Median price for San Diego County resale homes troughed in December 1995 at $163K ($162.5K, specifically), not $168K. I don’t concern myself with the fact that prices at close reflect transactions that were negotiated 30-60 days earlier. Maybe I should, but it would only confuse things (heck, some are having a hard time understanding this model!).

OFHEO data would not solve the ‘lagging by 30-60 days’ problem, PS. It’s based on closed sales and refinancings on conforming transactions compiled by FNMA/Freddie Mac. Thus, it, too suffers from the 30-60 day lag problem.

I’d forget about OFHEO for forecasting purposes; median prices for resale homes from DataQuick come out monthly, 15 days after month end. OFHEO gets updated quarterly, two months after quarter end. Per my graph, the median prices for resale homes per DataQuick and the index per OFHEO correlate very well.

I agree wholeheartedly, PS, that this will be a unique downturn (they all must be, I think, truly). I have no idea whatsoever how, mechanically, the personal debt (mortgage, HELOC, credit card), Federal deficit and debt, unfunded/underfunded state and city promises on pensions and benefits, war on terrorism, etc. will impact this downturn. But, I’m guessing it will be really, really bad.

I believe that buying at absolute trough will not be a major worry in the depths of the depression. Protecting one’s savings, and one’s family from Al Qaeda nuts coming in via Mexico with Iranian nukes and chemical weapons, will be the major concerns.

Repent, rent, save, and arm, people!

[Anonymous]

I slightly revised my analysis of the last upturn in San Diego resale home median prices (using employment lag 24 instead of employment lag 12), and present to you the graphs of the predictors of increasing home prices that I found in my number crunching.

Last time around, resale home median prices (black) didn’t go back up consistently until notices of defaults (NODs) (red) had been going down for 30-to-36 months:
[img_assist|nid=2372|title=
|desc=|link=node|align=left|width=466|height=350]
Last time around, resale home median prices (black) didn’t go back up consistently until employment (red) had been going up for 24 months:
[img_assist|nid=2373|title=
|desc=|link=node|align=left|width=466|height=350]
Last time around, resale home median prices (black) didn’t go back up consistently until sales (red) had been going up for three months:
[img_assist|nid=2374|title=
|desc=|link=node|align=left|width=466|height=350]
[img_assist|nid=2375|title=
|desc=|link=node|align=left|width=466|height=350]
So, my read is that until all three of these factors — NODs, employment, and sales — are going in the right direction, consistently, there may not be support for home prices to rise, consistently. But, when all three do turn in the right direction, consistently, it’s a signal to buy. That’s the theory, at least, based on the last upturn.

[citydweller]

jg, there was a link to this thread over at Calculated Risk. Have you updated these charts lately?

[citydweller]

jg, there was a link to this thread over at Calculated Risk. Have you updated these charts lately?

[Author]

Posts