Forecasting Model — Not!

[admin]

[img_assist|nid=1574|title=San Diego Resale Home Median Price: Actual, Fitted, and Predicted|desc=

Oh well, the best analyzed data of mice and men…

I arrived at a simpler model, which well fitted the data over ’88-’03. The simpler model only had four variables: sales 24 mos. and 30 mos. prior, notices of default 36 mos. prior, and the Fed funds rate 18 mos. prior. The model fitted the data well: r-squared, adjusted, of 95%, with all predictors having p-values of 0.000.

With this simpler model, I had sufficient historical data to predict prices over ’04-’06 and compare it to actuals. It appears that I’m missing some “irrational exuberance” factor over ’04-’06.

Any suggestions on what variable might help shift up the values over ’04-’06?|link=node|align=left|width=400|height=267]

[lindismith]

You’re clearly missing the g-factor. Greed.

And yes, the p-factor. Psychology.

Have no idea how you measure those.

[Anonymous]

LS, I agree with you. I’ve seen similar things captured via Nexis counts of newspaper articles. Anybody have access to Nexis and willing to capture quarterly or annual counts of articles on “getting rich quick in real estate,” or whatever captures LS’ greed factor, over ’88-’06?

[vcguy_10]

The in-sample fit is awesome! Have you looked at the time series literature on cointegration? Problem with time series data is that many variables trend upwards over time, resulting in high R2 regardless of the “true” causal mechanism. See if you can request your software to print the Durbin-Watson statistic, which would give a measure of serial (i.e., over time) correlation. Again, great job!

[vcguy_10]

When did psychology and greed become an issue in this boom period? Sometime between the 2003 and 2004 high selling seasons (late spring, early summer of each year). And why? Two (maybe three) factors made the boom into a bubble:

(1) House appreciation had been so fast that people started to think “I can’t lose no matter what I pay for RE”.
(2) Proliferation of exotic loans
(3) Mortgage interest rates bottomed in June 2003

Factor (2) above is not measurable, but (1) is. What we need for (1) is not only prices, but the change in prices, either as a difference or as a growth (percentage) rate. The same for (3), you need an indicator of how fast interest rates were dropping, not just their levels, but also the speed of change.

Finally, there’s no way a statistical model can capture the irrational aspect of bubbles and manias, so even if the % change in prices and interest rates turn out to be significant, you may still have a gap between actual and fitted for 2004-2006. You may want to try including a binary (dummy) variable for the bubble years (say, 2003Q2 through 2006Q1) and also its interaction with some other variables, especially % changes in house prices and interest rates.

Powayseller’s comment about missing the supply side is completely off. You are tracking the history of housing prices: each data point is already given by the market (both supply and demand, and any imbalance therein).

Great job JG!

[sdrealtor]

Jg
Just curious? What is the percentage difference between where we are and where the model predicted we should be? It looks to be somewhere between 25 and 35%.

[Anonymous]

Chris Johnston

It is better not to curve fit it as tight and use standard deviations from the regression line as fade entries. This is what I was trying to get Robert to do with his model and he did not want to spend the time doing it. In other words once it deviates 2 or more standard deviations for example upward you sell, and downward you buy etc..

I think it is ok to dial the curve fit in for trend purposes tighter but not actual buy and sell signals. This could also help forecast a low in terms of time, but actual price entries could be tied in with both that time horizon and standard deviation variance from the regression line.

I hope this makes sense. Here is an example, now we are in an uptrend but we have gotten probably 3 or 4 std deviations above a regression line which by basic mathematical probabilities gives a approx 99% + chance of a regression to the mean which were are seeing in prices right now. Also look for how many std deviations below it has gotten at market lows as a trigger for buys. If you could tie those two together you could greatly improve on Roberts model IMO.

If you curve fit a model too tightly to the widest variation it will never pick up normal cycles which is the overall goal. We have to accept through regression analysis that things will be stretched beyond the normal parameters and this is actually what we want for entry points.

Sorry for the trader talk, but hopefully it is helpful. Great work so far!

[Anonymous]

Thanks for the nice comments, VC and CJ.

Durbin-Watson statistic = 0.89602.

SDR, for Aug. ’06, the predicted price ($426K) is only 77% of the actual price ($555K).

VC, like you, I thought about putting in a dummy variable for ’03-’06; but, as I was only using data through ’03 to estimate the model, and as the frothiness really seemed to take off in ’04, such wouldn’t fully capture things. Maybe I’ll use data through the end of ’04 to fit the model, and only predict against actual for ’05-’06.

[Anonymous]

Rich/Prof. P., with your Voice of San Diego connection, could you get us an annual count over ’88-’06 on articles on ‘getting rich in real estate’? That way, I don’t have to cheat on the indicator variable for greed!

[powayseller]

vcguy-10, Do you know of any product whose price is determined only by demand, where supply is irrelevant? If you do, I will accept your comment that I am “completely off” 🙂

[Anonymous]

High sales correlate with (and causes) low inventory.  Low sales correlate with (and causes) high inventory.  Thus, sales alone may capture the effect of inventory.

I think it was Jim the Realtor (http://www.bubbleinfo.com/journal/) who showed that the increasing inventory in North County was due to slowing sales, not more homes coming on market.  If something well correlates with another predictor variable, use one or the other, but not both (cardinal statistical sin!).

[powayseller]

I haven’t seen a model yet that actually works, and they all miss the supply issue. Maybe the data is too hard to get.

What I’m looking for in a model is the inflection points. I’d like to know within 1 quarter if the market is turning. What is the leading indicator to show the market is going from high to low, or low to high? Once the market has turned, it keeps going in its new direction for 5 – 10 years, so the need for the model fades.

Let’s look at the inflection point [inflection point = time when prices reversed] for this current cycle. Prices peaked for condos in spring 2004, when inventory was 3,000. By June, inventory was 6,000, and by Sept 04, it was 12,000. The ripple effect took a little over one year to work to SFH, which peaked in August 05.

jg, could you check into this? Did prices fall as inventory rose? Did sales fall at that time also? If sales is a proxy for inventory, the summer of 04 should prove it.

My guess is that sales do not change enough at inflection points to signal the price reversal. I can think of situations where prices are pressured down to increased inventory, even though sales stay flat (such as rising foreclosures, unemployment). I know the inventory data is hard, if not impossible to get, and that’s why nobody is using it. However, my theory is that the model’s accuracy could be improved by using supply.

Another question about the model: what does it tell us that is better, or more timely, than the info we can get from asking SD Realtor or sdrealtor the following question: “What is your observation of price trends today?” These guys are going to be the first to know when prices change, even before my model or Campbell’s model gets a whiff of it. They will know within 2-3 weeks when the market has shifted, as they will see increased buyer interest, rising prices, fewer seller concessions. I wish we could backtest that question. Maybe we can: realtors, when did you first notice the market shifted?

Could the “ask the realtor” test be superior to any model?

[FutureSDguy]

“Could the “ask the realtor” test be superior to any model?”

Using RAND() in Excel would be superior to asking a realtor. 🙂 J/K (no truth in this joke, mostly anyway… I’m too much of a newbie to be making jokes, aren’t I) Realtors specialize in making commissions more than they specialize in following the market, so their answers tend to be rather biased. I’ve had short email exchanges w/ some agents in the area. One of them was “But San Diego is a very attractive place to live, so prices will stay high.” I told her that frequent piggington.com and she never replied back.

[powayseller]

Not any realtor, I specifically mentioned SD Realtor and sdrealtor. I should include Bugs to the list.

[Anonymous]

Chris Johnston

It gives a QUANTIFIABLE way of making a buy or sell decision. If you just read this blog there is so much information that is available. Having a measurable way of synthesizing all of it into a fixed model is the whole school of thought behind systematic investing. Discretionary investing takes you into paralysis of over analysis that leads to very random and arbitrary decisions.

This in and of itself is problematic. However, add the emotional influences that can also creep in to these comprehensive analysis decisions, and what you have is a chaotic decision making process. Just think about all of the different opinions on certain topics everyone has. How can you ever know whose opinion is the correct one? This leads to very inconsistent results in investments. This is the reason that I make my decisions off of models and trading systems that I develop. The emotion and random influence of the process is minimized. I do not want alot of stress over buy and sell decisions and which of the hundreds of possible considerations might have been evaluated incorrectly.

What if you made it this simple. Create a regression model and buy whenever price was extended 2 or more std deviations below the line and sell when it is extended 3 or more above? Nothing else. Why 3 above and not 2? Prices are upwardly biased and skewed so that window should be broader. Very simple, and just a basic example of how to cut alot of the noise out of all of this. This is not my model, just an example to illustrate my point.

This is all that this model helps you do. It is a tool not the holy grail. If you can create a better one you should, but I believe as a tool his works pretty well.

[Author]

Posts