When are we ever going to learn

[Austin]

Tom: You open the door to an interesting discussion that would be very lengthy. I wish we could post graphs, that would save a lot of time as one picture is worth a thousand words. How do you know when you have accounted for all value influencing factors? The answer is that you have to know where you are going and most importantly, recognize when you get there. When the variance between actual and predicted prices falls within a 10% or less deviation, it is time to stop adjusting. Then you have entered the range of variance zone and can't tell what is up or down.
Try this method just for fun the next time you do a residential marketing grid: Go into your Excel spreadsheet and program in a marketing grid just like it is programmed in your residential software. At the bottom of the sheet make a graph of GLA vs. actual sale prices. Then make a second graph of adjusted sale prices with GLA vs. adjusted prices. Then make adjustments for physical property features only in the order of their dollar value, not including GLA. Generally in this order: Basements with any finished areas, garages, porches, fire places, etc., just as if you were doing the cost approach.
If the sales you selected are truly comparable, their trendline should be generally linear with a positive slope. Not true in older properties. Size in then meaningless.
Now make a size adjustment using this method. Do a graph of the GLA vs. adjusted sale prices per square foot, calculate the trendline and regression equation, and using the slope of this trendline, adjust the sales GLA to the subject's GLA by multiplying the difference by the slope and adding or subtracting.
Now look at the graph of the GLA vs. adjusted prices. If you did a perfect job, the resulting trendline will be flat with a zero slope and the data points will be within a 10% range of the trend line. If they are not, then something else is affecting price. Until you have accounted for the physical differences, it is not possible to know if anything else affects price. Each time you make an adjustment watch the graph of GLA vs. adjusted prices change. It is very rarely that I need to go beyond this level. This is essentially doing a stepwise linear regression. If either graph does not form a generally linear pattern, your sales are not comparable, or either there is no market correlation between price and price influencing factors, in which case, there is no reason to be doing an appraisal anyway. If one sale is out of line, then you start looking to find out what the next value influencing factor is.

 

[Tom Hildebrandt]

Austin

I understand the thrust of your arguments regarding identifiable physical characteristics. I agree, that broadly speaking, an individual line adjustment that is flat (or negative) and less than 10%, it is not relevant statistically so long as the individual item can be accurately desribed empirically.

Essentially the question I pose is how do you quantify subjective issues, such as design and appeal, in an empirical manner? We know they have an impact on value. If the best you can get empirically is a 60-70% R squared grading strictly empirical data (a number I heard thrown out at a seminar the AI taught regarding AVMs) there must be something else adding value.

The problem is where the line adjustment is definitively positive, but the variance (or is it deviation I am speaking of, hell I do not know I am a layperson on this regression stuff) is 20, 30 or 40%, perhaps even 100%. This often happens for items like design and appeal. The purely emperical model can not describe this market variable as it can only be quantified subjectively and there are a whole host of such variables.

I am just a layman at this statitical stuff, but it seems to me you can only go so far with a pure empirical model. You then have to make some assumptions regarding the subjective issues. They can be quantified by establishing a grading scale, and tested against known sales.

I think we are talking on the same page on this topic, I just don't know enought to explain it better.

Regards

Tom Hildebrandt GAA

 

[Austin]

Tom: Lets take it one step at the time. We don't know if design and appeal is a factor until we have solved for the physical differences. Not ten minutes ago I did a marketing grid and clearly demonstrated a design and appeal /quality of construction factor of about $40,000. You could see it in the pictures of the comps and the subdivisions. After I had solved for everything else, it was obvious what the difference was. Design and appeal and quality of construction or inseparable covariant variables. I could demonstrate this to you with data out of Marshall & Swift, but the steeper the trend line, the higher the quality of construction. Take a range of sizes with the same quality of construction-design & appeal, and the trend line is perfectly linear. It is in the cost data that way, so it has to be that way in the market.
Second: You mentioned that models like AVM can only account for 60 to 70% value influencing factors. Believe it or not, that is darn good. Probably much better then most appraisers are getting now. The reason for this is that we are explaining 60 to 70% of the factors affecting the price of selected properties, not the entire population. Also, the R^2 accounts for number of degrees of freedom which lowers it considerably, but again, we have accounted for that by carefully selecting the sale data. For example you can have a narrow value range of around 6% variance of predicted prices and the R^2 factor can be around 70%. The test statistics don't tell the whole story.
PS: I just thought of another factor about the R^2 which shows the percentage of how much we have explained with the regression. Built into this test stastic is a safety factor. When you see an AVM with an R^2 (coefficient of multiple determination) of 60%, the actual percentage is almost always much better because in AVM and other non stepwise methods, have a number of independent variables that may not be tell us anything. This R^2 statistic tells us four things: 1. One of the independent variables is not telling us any thing. 2. We may have to many or not enough independent variables. 3. We may not have enough degrees of freedom. 4. The amount of variance that we have explained. All of these factors can be over come with the model design. In other words, we design our way around them. The bottom line is that we are doing much better than the test statistics are telling us if we feed the model truely comparable sales. I have another model to pick and verify the comparability of the comparable sales as part of the process.