## Savvi's Analytic Engine

Regression techniques have long been central to the field of economic statistics (“econometrics”). Increasingly, they have become important to appraisers in their data analysis. In an appraisal regression model, the dependent variable sale price is “regressed” on a selected set of property characteristics to determine how much of the variation in sale prices of geographically comparable properties is due to variation in the predictor characteristics included in the regression model. The higher the percentage of variation in the sale prices of like properties that can be “explained” by the predictive set of property characteristics, the more accurate your prediction of the property value. A thorough regression analysis will therefore improve your ability to accurately appraise property values beyond what can be done by “guesstimating” or looking at only one or a few property characteristics.

The SAVVI Analytics Regression Modeling Approach

The SAVVI modeling approach is called “Hedonic Regression” because it uses a set of property characteristics to estimate the value of a marketed good; in this instance the value of a property. There are 5 steps in the process:

To begin, a group of properties in geographical proximity to the property whose value is to be appraised is selected from the appraisers' MLS database.

The number of properties selected for analysis should be large enough (no fewer than 200 properties) to allow for variation on the property characteristics to be examined as well as on the sales prices. At this point, we are after variation and not strict comparability, because having variation to statistically model actually improves the estimates of the value of each characteristic as well as sales price of the appraised property. This step includes calculating a standard deviation score for each property to identify which have extremely high or low sales prices relative to the average of other selected properties and removing those with extreme prices from the database.

The second step involves calculating a time adjustment for each property in the data base to account for monthly fluctuations in sales price contingent upon when each selected MLS property was sold over the preceding few years. A smoothing algorithm that uses 5 month averages is used to reduce extreme price fluctuations on the trend line.

Third, a set of bivariate regression (i.e., two-variable) models is run using the selected properties. Bivariate regression assesses the association between a single dependent variable such as sale price and a single independent or predictor variable such as square footage. A separate bivariate regression is run for each property characteristic that is being examined. Each simple regression is run on the full data set containing information on the 200+ properties. The purpose of this step is to screen for those property characteristics that are reliably associated with variations in the sales prices of the properties in the data set.

In the fourth step, those property characteristics found to be reliably associated (i.e., statistically significant) with the sales prices of the examined properties are included in a stepwise multivariable regression (i.e., many variables) model. Using the full data set, the multivariable regression assesses the association between sales price, the single dependent variable, and a set of multiple predictors such as square footage, lot size, and age of the property.

This model determines how well the remaining property characteristics individually and as a set predicts variation in sales price. Importantly, the model examines the associations among the property characteristics themselves. Only those characteristics that uniquely predict sales price are retained in the model. Characteristics that overlap substantially with other characteristics already included in the model (in terms of predicting sales price) are dropped. Only the best predictor set is retained.

In the fifth and final step, the appraiser selects a small subset (3) of properties comparable to the property being appraised. A map displaying properties by their proximity to the appraised property assists this step. Using the predicted sales prices for these properties and the valuation for each property characteristic in the final multivariable model, a final sales price is obtained for the appraised property.