Forecasting Model

[admin]

[img_assist|nid=1569|title=San Diego Resale Home Median Price: Actual vs. Fitted/Predicted|desc=Robert Campbell, in his book, laid out that sales, interest rates, notices of default, etc. can be used to pick the best time to buy a house. Someone on the forum stated that RC did not use regression analysis to analyze his data. I built on RC’s approach, using Minitab’s graphing, cross correlation, stepwise regression, and ordinary regression analysis functions, to arrive at a model for predicting San Diego resale home prices. The graph above shows actual San Diego prices vs. those predicted by my model. What I found is that price is a function of (1) time, (2) sales 6/18/24/30 months prior, (3) notices of default 18/24/36 months prior, 30 year fixed mortgage rates 6/24/30/36 months prior, and Fed funds rate today and 18/36 months prior. Without a doubt, given the steep run up in prices, things may very well be different on the way down. But, I feel comfortable that the predictor variables make sense, and may give a ‘heads up’ that prices are on their way back up (5-10 years from now!), and that folks should get back in at the trough. And, technically, if I was back in school, I would be in trouble for using all of the data that I had (monthly over ’88-’03) to arrive at my model, instead of using the first 80% of the data and then seeing if it fit using the remaining 20%. But, this is the real world, with money involved, so to heck with saving data! What do you think?|link=node|align=left|width=400|height=267]

[lindismith]

I think you are 10x smarter than me!

So, what’s the future look like?

[(former)FormerSanDiegan]

At least your period covers both down and up markets.

However, all you’ve shown is that you can fit a curve, with a set of input parameters based on other data. If you have enough input parameters and knobs you can fit anything.

An experiment to test your model is to predict the value from 1982 to 1990 (if you can find the input data) and then post that, without tweaking parameters.

How far forward do you project ? Since you have “Fed funds rate today” as pone of the input parameters does that imply you can’t project forward ?

[(former)FormerSanDiegan]

My Forecast Model says …

[img_assist|nid=1571|title=Housing Forecast Model|desc=|link=node|align=left|width=400|height=388]

[sdrealtor]

Forecast be damned, I love the model!

[Anonymous]

Chris Johnston

Show us a forward projection

[powayseller]

First, kudos to jg for trying what few have done. How many people have ever tried to make a real estate forecasting model? So my hat’s off to you.

However, I’ve got be honest and point out the model needs a lot of work. I want to be very careful to show the weakness in the model while still holding jg’s work in high regard.

FormerSanDiegan is right. You’ve got a bunch of data, and made an equation that fit the prices. The model seems to break down in the early 90’s, a period of falling prices. Since it breaks down when prices fall, how useful will it be when prices fall this time? Furthermore, NODs are a lagging indicator, so by the time NODs rise, prices have changed for a year already.

The model does not use ALL sales, i.e. resales plus new home sales? As builders are discounting heavily now and have overbuilt, new home sales are a big proportion of all sales; later in the cycle, as builders have stopped building, new homes sales will be negligible. Any model using only new home sales or only resales is using only half the data, but skewed data: it is weighed toward lower sales in the up cycle (bec. new home sales are a bigger percentage) and skewed down in the down cycle (as new homes are a smaller portion of sales).

The biggest weakness in the model is that it uses only demand, while ignoring supply.. Econ101: price is determined by demand and supply. How can you predict prices knowing only demand? How can you predict oil prices without knowing the OPEC output? If you are given only world demand,but not world supply, how can you? The same is true for real estate, and it is Campbell’s greatest flaw, and the reason I believe he has so many inaccuracies in his model. Campbell has several incorrect sell and buy signals, yet he was satisfied with his model enough to publish it? The guy should refund me the money for the book.

jg, I want to commend you for taking a stab at it, and I encourage you to project it forward, and find out exactly how far off you are. Then refine it by adding more data on sales and inventory. Perhaps unemployment, new home permits, consumer confidence, have an impact too. I admire your tenacity. It was a big task.

[Anonymous]

FSD, I don’t have monthly data for anything earlier than ’88; if you can supply it, I’m happy to crunch it.

The data that I have came from RC’s book (NODs), DataQuick ($300 for San Diego and La Jolla data on prices and sales for resale homes), and the St. Louis Federal Reserve site (interest rates). DataQuick has nothing earlier than ’88.

FSD, fun bubble model! But, now I can’t show my graph on the website to my kids (or wife, especially!).

Thanks for the compliment, LS.

CJ, when I get some missing data to fill in ’04-’06, I’ll lay out some projections.

PS, as my professors showed me in graduate school, simpler models often fit and predict better. With more inputs, there’s more room for something to go amiss. I vividly remember being shown that a simple autoregressive model was a better predictor of GDP than one based on employment, trade deficits, industrial production, etc. Hard to believe, but true.

[Anonymous]

Chris Johnston

There is only one error in his model just to be straight for the record, not several. It has 7 trades and 6 of them have been profitable pending the Aug 2005 sell signal. Do you think that one might be correct making it 7 out of 8?

Do not over complicate models, they become data mining once you do that and prone to failure. The supply is a part of his model because existing home sales which is supply is it’s most heavily weighted component.

[powayseller]

He does have several wrong signals, namely the buy in 1/94, sell in 8/95, buy in 8/96, sell in 12/01, buy in 1/03. He gets all confused: what is it, buy or sell? He’s got people buying and selling every 18 months between Jan 94 and Dec 96. He is definitely wrong on the buy signal in January 2003! Again he had a wrong sell signal in December 2001.

Following his method, you’d be losing money in real estate, because he told people to be out of the market from 12/01 – 1/03, at time of great appreciation. In addition, the real estate commission incurred by buying and selling every 18 months is prohibitive. He has 3 buy/sell signals in 2 years between 1994 and 1996, and 2 buy/sell signals in 2001 – 2003. I only see his signals through 1/04, the latest date in his book which I just purchased.

Let’s not confuse the fact that Campbell is a nice guy with the fact that his model is poor at forecasting real estate cycles. Kudos to him for trying, but shame on him for misleading people into thinking that he has it figured out.

[Anonymous]

Chris Johnston

PW – He is not wrong on the 2003 buy, the median was higher Aug of 2005 when the sell was generated. A buy in 2003 and sell in Aug of 2005 basis the median was a profitable trade.

[powayseller]

Chris, he was wrong to have a sell signal in 12/01. Then a buy signal in 1/03. He should not have sold in 12/01. What good is that? He had people out of the market during one of the greatest boom periods in San Diego real estate.

[Author]

Posts