Start Date

23-11-1988 4:30 PM

End Date

23-11-1988 6:30 PM

Description

The paper tests the efficacy of feedback models in two ways. First, to see whether feedback models can arrive at known values. Second, to see whether feedback models can "automatically" time adjust residential models. Two feedback algorithms are used: MicroCAMA's and Lehigh County's hard-coded version. The experiment uses three data sets, each with a given value generator and a distribution. We specify a model of the form (Gm(Lm(La)+Bm(Ba)))e, where Gm, Lm and Bm are each a product of multipliers and La and Ba are the sums of a set of additive coefficients multiplied by property characteristics. The variable e represents an error term drawn from three known distributions: the Normal, Gamma, and an asymmetric distribution of the Beta family with two fixed terminuses (at .5 and 2.25) and positive skewness. The structure of the property generating procedure is identical in each case; only the stochastic parameter will vary. The data set includes about 2,000 parcels, and the distribution of characteristics reflects the actual distribution of characteristics in Lehigh County. The first test involves the ability of feedback models to discover a known model in environments with different error structures. This is a key issue in evaluating the feedback approach. The other major valuation procedure, regression analysis, has well-defined statistical properties, and its appropriateness can be tested. Feedback requires no distributional assumptions and is said to arrive at optimum values regardless of error distributions. The advantage to using a feedback model is that it allows the estimation of values in any circumstance and does not require the strong (and generally unmet) distributional assumptions required to allow testing of the statistical validity of the ordinary least squares coefficient estimates used in regression analysis. The disadvantage is that the experimenter knows nothing about the distributional character of the feedback coefficients. A necessary condition of the claim that feedback reaches the optimal coefficients is that it not be sensitive to the distribution of the property prices and characteristics. That is, the models should arrive at trivially different estimated values for the three data sets. Another claim of feedback proponents is that feedback automatically adjusts for time. By ordering the data by sale date, there is a time-related increase in coefficient values during inflationary periods (and a corresponding decrease during deflationary periods). A specific time-related change will be built into the model, using the rate of time change discovered by Michael Skaff in a study of sales in Detroit, Michigan, as recorded in the Assessment Digest. The feedback values should automatically correct for both the direction and the magnitude of the change. The results of the two parts of the experiment will be valuable for researchers assessing the value of non-parametric, iterative methods that are similar to feedback.

Publication Date

November 1988

Download

Share

COinS
 
Nov 23rd, 4:30 PM Nov 23rd, 6:30 PM

Feedback: A Monte Carlo approach

The paper tests the efficacy of feedback models in two ways. First, to see whether feedback models can arrive at known values. Second, to see whether feedback models can "automatically" time adjust residential models. Two feedback algorithms are used: MicroCAMA's and Lehigh County's hard-coded version. The experiment uses three data sets, each with a given value generator and a distribution. We specify a model of the form (Gm(Lm(La)+Bm(Ba)))e, where Gm, Lm and Bm are each a product of multipliers and La and Ba are the sums of a set of additive coefficients multiplied by property characteristics. The variable e represents an error term drawn from three known distributions: the Normal, Gamma, and an asymmetric distribution of the Beta family with two fixed terminuses (at .5 and 2.25) and positive skewness. The structure of the property generating procedure is identical in each case; only the stochastic parameter will vary. The data set includes about 2,000 parcels, and the distribution of characteristics reflects the actual distribution of characteristics in Lehigh County. The first test involves the ability of feedback models to discover a known model in environments with different error structures. This is a key issue in evaluating the feedback approach. The other major valuation procedure, regression analysis, has well-defined statistical properties, and its appropriateness can be tested. Feedback requires no distributional assumptions and is said to arrive at optimum values regardless of error distributions. The advantage to using a feedback model is that it allows the estimation of values in any circumstance and does not require the strong (and generally unmet) distributional assumptions required to allow testing of the statistical validity of the ordinary least squares coefficient estimates used in regression analysis. The disadvantage is that the experimenter knows nothing about the distributional character of the feedback coefficients. A necessary condition of the claim that feedback reaches the optimal coefficients is that it not be sensitive to the distribution of the property prices and characteristics. That is, the models should arrive at trivially different estimated values for the three data sets. Another claim of feedback proponents is that feedback automatically adjusts for time. By ordering the data by sale date, there is a time-related increase in coefficient values during inflationary periods (and a corresponding decrease during deflationary periods). A specific time-related change will be built into the model, using the rate of time change discovered by Michael Skaff in a study of sales in Detroit, Michigan, as recorded in the Assessment Digest. The feedback values should automatically correct for both the direction and the magnitude of the change. The results of the two parts of the experiment will be valuable for researchers assessing the value of non-parametric, iterative methods that are similar to feedback.