Details on the Last Upturn

[admin]

Sales and prices are moving downhill fast. Given such, I’m interested in understanding the predictors of firming sales and prices (though such is far, far down the path).

In my analysis of the ’88-’03 data from Robert Campbell’s book, I found that two variables — sales and notices of default — were the strongest leading indicators of home prices, both up and down. But, when I looked in detail at the last upturn — in ’96-’97 — I found something disturbing:

[img_assist|nid=1668|title=
Prices and Sales|desc=|link=node|align=left|width=400|height=267]

[img_assist|nid=1669|title=
Prices and NODs|desc=|link=node|align=left|width=400|height=267]

See how sales and NODs march up smartly, fall off sharply, then march back up again simultaneous with increases in price? Not good, if we are to rely upon sales and NODs as leading indicators.

I wasn’t here in San Diego in the mid ’90s, but it appears that San Diego suffered losses in employment and income in the mid-’90s; I mistakenly thought that all of that stuff had happened in the early ’90s.

[img_assist|nid=1670|title=
Income and Employment|desc=|link=node|align=left|width=400|height=267]

While personal income is the better predictor, probably, I can only find annual personal income data for the San Diego SMSA, while employment is available monthly for the San Diego SMSA. Hence, I use employment as a proxy for San Diego’s economic status.

So, after graphing, cross-correlating, stepwise regression, and regression, I arrived at the following as predictors of increasing prices in ’95-’96, from most to least important:

Notices of default 30 months prior
Employment 12 months prior
Sales of resale homes 3 months prior
Notices of default 36 months prior

Here’s how these factors fit the actual price data:

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

This makes sense to me: NODs and sales do put pressure on prices. But, after an economic dislocation, employment and income must be there to support prices, and only increases in such will support increasing prices.

What do you think?

Sincerely,
The Puritanical Bible Beater

[powayseller]

Why do we need employment and income data? If unemployment is high and income is low, then wouldn’t sales also be low? So sales would be a proxy for employment and income, just as they are a proxy for inventory (as you explained to me before, thank you again).

The first chart shows a false recovery, something that asianautica was referring to before. What exactly happened to make sales increase for a few years, just to turn back down? Do we need to consider tax policies, inflation?

I mentioned in a post yesterday or Friday, that my friend bought a killer house in Poway in late 1995 for half the price/sq ft of her neighbor w/ the idential lot size and square footage just 18 months prior. So the market was still weak in late 1995.

Perhaps some old newspaper articles, or the memory of an appraise or realtor could be called upon, to explain the 3 year surge in sales.

I’m also wondering if anybody who bought in the early 90’s would have lost money if they sold in the mid-90’s.

[Anonymous]

PS, I can’t use sales alone, as sales gave a false harbinger of rising prices, fell off, then only moved up again just shortly before/simultaneous with prices moving up. Thus, sales needs help as a predictor.

To me, it appears that SMSA-wide income leads sales up, then down, then up:

[img_assist|nid=1674|title=
Sales and Personal Income|desc=|link=node|align=left|width=400|height=267]

County-wide income moved up, county-wide sales moved up two years later; income moved down, sales moved down three years later; income moved up, sales moved up one year later. Seems logical to me.

Does anyone know of monthly or quarterly indices of county-wide income? Otherwise, I have to rely on monthly employment as a proxy for county-wide economic health.

[powayseller]

I get it now. The real estate market picked up in late 1992 for one year, but rising unemployment nipped that recovery in the bud, making the market go down in late 1993 to late 1995.

Despite the up and down of sales, median price did not start rising until 1996. Perhaps inventory would explain some of this; maybe sales were increasing in 1992, making it appear that there was a recovery, but rising unemployment was causing inventory to spike as well, making months inventory stay high. Perhaps using months inventory as a predictor would not have given a false signal that the market had turned. I am back to wondering about inventory as a predictor, since the data doesn’t seem to be available.

So my theory is that on the way down, inventory is not as important, as slowing sales indicate the market is turning, and the slower sales pace causes inventory to rise. As the cycle turns down, inventory rises not only due to a fall in sales, but also due to distress and REOs. When sales pick up, there is no correlation with price as long as REOs are still high.

Maybe none of this makes sense, since you already accounted for NODs.

[Anonymous]

PS, your summary is spot on.

Maybe inventory would better predict ups and downs; I don’t know unless I crunch the numbers.

But, as you say, NODs capture financial distress and may capture overhang, too.

Regression is nothing more than weighting of factors that go up and down; to me, it’s amazing how careful manipulation of the weights (by the software) yields a fit as close as it does from independent factors such as employment, NODs, and sales.

Get some sleep!

[powayseller]

NODs stayed above 800 in every month from from 3/91 to 8/97. If we use NODs as a predictor, then we would stay out of real estate until 9/97, the month in which NODs went below 800 and stayed there (only one month, 3/98 was slightly above 800).

It looks like the median turns up in January 1996. However, two things we have to keep in mind about the median:
1) it lags by 12-18 months
2) on the way down, the median overstates the housing prices, and on the way up, it understates it I think.

I think at the trough, sales pick up on the low-end, as investors get back into the market, and first time home buyers once again start the chain of sales. Once the starter homes are sold, those starter home sellers can go and buy the mid-priced home, and those sellers can get their high end homes. So I’ve been saying that the increased activity of lower end homes at the trough, would actually make the median go down, even though the price of each indiviual home would be rising.

OFHEO data, which tracks the price of the SAME house, indicates that prices dropped from 1989 Q2 to 1995 Q2, rose from 1995 Q3 until 2004 Q3, and fell from 2004 Q4 on. Except for some minor zig zagging in 1992 – 1995, the price trends are fairly clear. I had discarded OFHEO data, bec. it only tracks homes with conforming mortgages, but the data shows they are accurate on the pricing trends. Although only a small percentage of San Diego homes are tracked with OFHEO, the price movement of those homes is accurately captured.

[Anonymous]

PS, NODs are one of three factors in the model. If one looked at NODs, exclusively, one would be in trouble. In my model, one must also look at sales and employment.

To me, the OFHEO data looks amazingly like the median sales price data from DataQuick. So, I felt comfortable using the DataQuick information, which comes out monthly (OFHEO is quarterly) with a lag of 15 days (OFHEO has at least a 45 day lag).

[img_assist|nid=1685|title=
Median Prices vs. OFHEO Index|desc=|link=node|align=left|width=400|height=267]

[powayseller]

jg, the OFHEO index peaked in 2004 Q3, so I wonder if we used the same data series. Also, OFHEO data is not lagging, so I don’t understand why these two series are so closely correlated. I expected median to lag OFHEO by 1 – 1.5 years.

Another factor we haven’t considered is population change. In the early to mid-90’s, as unemployment rose due to aerospace and defense cutbacks, many people left San Diego.

Building permits indicate builder optimism; builders respond to the turning market faster on the upturn, since they can just ramp up production. On the downturn, they have to finish out their phases, which can take up to a year? Building permits increased in October 96, marking the begining of builder optimism.

[Anonymous]

Building permits were not a useful predictor; they lagged prices on the way back up (i.e., prices started moving up in ’96; permits only started moving up in ’97):

[img_assist|nid=1691|title=
Prices and New Home Permits|desc=|link=node|align=left|width=400|height=267]

Population is nice, but total employment, which I use, certainly reflects population and is available monthly. County wide personal income would be better, as it would reflect population, employment, and wherewithal to pay, but it’s only available annually.

The OFHEO data peaked in Q2 ’06 at 324.56, not Q3 ’04, when it was 280.46. You are confusing peak rate of change (Q3 ’04) with peak in absolute value (Q2 ’06). Prices for resale median homes were higher in June ’06 than in October ’04.

OFHEO data comes from (1) closed sales and (2) refinancings, both on resale homes. Given such, it makes sense that it is consistent with the DataQuick data on closed sales for resale homes.

I’m confused; (1) I’ve laid out three factors — NODs, sales, and employment — that tightly fit the data from the last upturn. (2) The three factors, together, give a three month heads up of increasing prices. (3) Two of the three factors — NODs and employment — give 12-36 months heads up of potentially increasing prices. (4) The three factors individually and together make economic sense. (5) The three factors are available free of charge from sources on the internet.

What exactly is there to modify here, and why?

[powayseller]

I prefer using rate of change over absolute value. When I chart % change quarter over quarter, I can clearly see the trend reversed in 04. You are much better at statistics than I am, so perhaps you can explain why the derivative is a better predictor than the absolute value. Actually, I see the second derivative as important: the point where the slope is zero. It is usually at those points, that indicate a reversal in the cycle. If the data point reverses only a teeny bit, or for only one or two times, then it turns out to be noise. That is my conclusion from looking at the second derivative, visually.

Likewise, building permits clearly peaked 6 months earlier if you use % change. I think looking at the rate of change more clearly indicates a reversal in the trend. So using this method, I get a sell signal as I noted above when the building permits and sales peak at the top or bottom, and change direction.

Maybe there is nothing to modify. I’m just not clear when your model generates a buy or sell signal.

[Author]

Posts