EXTRACTING LINE-ITEM ADJUSTMENTS

Transcription

1 EXTRACTING LINE-ITEM ADJUSTMENTS 1

2 Automated Valuation Technologies, Inc. This introductory workbook was created by Automated Valuation Technologies, Inc. (AVT). The purpose of AVT is to fill the voids in appraisal practice that result from the rapidly changing appraisal environment. Appraisers often find themselves engaged in new activities which require the use of technology that has not yet reached their computer desktops. This is both unfortunate and unacceptable. It is unfortunate because appraisers are not fully effective in carrying out their duties. It is unacceptable because it compromises the vital role appraisers perform in the safekeeping of their country s greatest wealth: real property. It is AVT s mission to provide technologies real estate appraisers require to fulfill their duties. AVT operates under the belief that there is no substitute for the Neighborhood Appraiser. Their knowledge of the local market is unique and cannot be duplicated by remote computer analysis. These local appraisers are hardworking and dependable. Without question, these gritty individuals will carry out their duties as long as they have the knowledge, training, and tools to do so. This material is for educational purposes only This Workbook is copyrighted Automated Valuation Technologies, Inc. 215 Ellis Avenue, Maryville, TN

3 David A. Braun, MAI, SRA David A. Braun, MAI, SRA, has been an appraiser in the Greater Knoxville area, since He was born in St. Louis Missouri, and later moved to Knoxville, Tennessee. He is a 1976 graduate of the University of Tennessee with a Bachelor s of Science degree in business administration. David received the Appraisal Institute s SRA in 1980 and MAI in David has been appraising most types of properties since He is also a licensed real estate broker with sales and management experience. David founded Braun & Associates, Inc. in 1983 in Maryville, Tennessee. Braun & Associates, Inc. is a medium-sized appraisal firm with 12 employees and handles most types of assignments and property types. He is also president and founder of Automated Valuation Technologies, Inc. (AVT). AVT produces unique appraisal seminars and software for the practicing appraiser. David has had how to articles published in the Real Estate Valuation Magazine, The Working RE Magazine, and presented white papers on appraisal topics. He has twice served as President of his local Chapter of the Society of Real Estate Appraisers (now a part of the Appraisal Institute). He has served on the Appraisal Institute s Instructor Subcommittee at a national level. David is currently an approved Appraisal Institute Instructor and is certified by the Appraisal Foundation as a USPAP Instructor. David has been a pioneer in developing and reporting the scope of work. He developed and taught the first scope of work seminar, and was the first to use a computer program to aid in the development and reporting of the scope of work. He authored, APPRAISING IN THE NEW MILLENNIUM; Due Diligence & Scope of Work. David has developed numerous applications that run on macros in MS Excel some of which are free and some are for purchase. These can be found at AVT s website. He has designed, but did not program other applications such as the USL Documenter and the Regression Plus, and the Fusion. David has made presentations to the Federal Financial Institutions Examination Council (FFIEC), and to the Association of Appraiser Regulatory Officials (AARO) on the topic of scope of work. 3

4 Contents FORWARD... 5 INTRODUCTION... 6 A VERY POWERFUL TOOL... 8 THE TWENTY-ONE MARKET STUDY... 9 Thresholds Model (A) Model (B) Model (C) Model (D) Model (E) Model (F) Model (G) SCORE CARDS Model (A) Model (B) Model (C) Model (D) Model (E) Model (F) Model (G) Coefficient Confidence Intervals Interpretation of the Score Cards CONCLUSION

5 FORWARD After trying several training methods we have concluded that workbooks supplemented with videos is the best avenue for success. This is an inexpensive and flexible learning environment. This is the third of a five part series: 1. An Introduction to Market Modeling 2. Introduction to Regression Analysis 3. Trending Values over Time 4. Extracting Line-Item Adjustment Rates 5. Reporting Your Regression Analysis The first three workbooks will empower the valuator to form credible value opinions with regression analysis. The fourth workbook explores the phenomena of extracting lineitem adjustments. Finally, the fifth workbook shows the valuator how to report the regression analysis in a USPAP compliant manner. 5

6 INTRODUCTION In this series of workbooks we have differentiated between two methods of the sales approach to value; the traditional comparison method and the regression analysis method. We identified some differences and explored that the ideal comparables employed in these methods differ and discussed under which circumstances one method is more reliable than the other. In this Workbook we will consider two cases; theoretically and real world as we attempt to answer some important questions..some of the questions we will address are: Is regression a useful tool in extracting line-item adjustment rates? Do the regression model and the adjustment grid have to consider the same variables (property characteristics) to be in sync? Typically, how accurate are the coefficients? What are coefficient confidence intervals? The only reason people get lost in thought is because it s unfamiliar territory. ~ Paul Fix One of the reasons that regression is so hard to learn is that most books make very broad statements about it which are often related to a very specific situation. In reality the answer to most questions about regression begin with; It depends In addition to this ambiguity, the subject of extracting line-item adjustment rates is really intermediate, rather than novice level material. The goal of this Workbook is to provide some answers to the question: How dependable is regression analysis for extracting line-item adjustment rate? 6

7 To definitively answer this question we will run 2,000 regression analyses on Twenty- One individual markets (42,000 analyses in all). These markets will be based on hypothetical data so we will know the correct final values and the coefficients of the variables. We will basically build a score card of how well regression performs as we move from good data-sets to weaker data-sets. This will provide quantitative answers to the question above. This laboratory approach will engage our minds in ways that are different than researching books and published articles alone. Let s begin by clarifying what we mean by extracting line-item adjustment rates. The variable s coefficients typically represent the line-item adjustment on the comparative grid. In the following example the subject property has 1,800 square feet, is 12 years old, has a double garage, and an inground pool. 7

8 A VERY POWERFUL TOOL Premise 12 In theory regression analysis will return the correct line-item adjustment rate even when all of the property characteristics are not considered. 8

9 THE TWENTY ONE MARKET STUDY The novice analyst can benefit from regression analysis many situations. What we need to determine is which situations allow for the novice user to be successful, and which situations do not. I have created a Twenty-One market study with that goal in mind. The models go from a strong data-set to progressively weaker data-sets. The market models are hypothetical allowing the correct answers to be known. Then a Monte Carlo simulation is run while the parameters of each model are randomly changed to give a more complete outlook. There are seven basic models; (A) through (G). Each model is divided into three sub-models based on the amount of random variance built into the sales prices. The variances ranges are +-8%, +-12%, and +-16%. Each of the 21 submarkets is run 2,000 times resulting in a total of 42,000 runs. The following items randomly change within each model in each run: The number of sales. The variance (within the three levels above) built into each sale price. The size of the variable s (property characteristics). The value of the hypothetical coefficients which the sales prices are dependent on. 9

10 The following information is monitored: The coefficient of determination (R 2 ). The absolute correlation when only two variables are considered. The absolute residual as a percentage of the predicted value to the actual value. The absolute residual as a percentage of each variable considered in the analysis. The number of times in the 2,000 runs that; o The value has a residual > absolute 10%. o The coefficients of each variable has a residual > absolute 33%. o The number of times the actual coefficient is outside of the confidence interval. o The number of times the confidence interval is greater than +-30% of the predicted coefficient (a total interval size of 60% of the predicted coefficient). 10

11 The model is built backwards for the subject and each sale used in the analysis. For example in Model (A) if the subject (or sale) has 2,000 square feet, is 10 years old, has +5% random variance, the assigned coefficient for GLA is $50.00 per square foot, and the coefficient per year old is -$1,000 then the value (sales price) would be: Step 1: Find the Value with no random variance = [(2,000 * $50.00) + (10 * -$1,000)] = ($100,000) + (-$10,000) =$90,000 Step 2: Calculate the amount of the random variance =5% * $90,000 =$4,500 Step 3: Find the actual value (sale price) =$90,000 + $4,500 =$94,500 If the regression model returned a value of $99,000 then the absolute residual as a percentage would be: = Absolute ($94,500 - $99,000) / $94,500 = Absolute (-$4,500) / $94,500) = (4.76%) 11

12 If the regression analysis returned a coefficient for the GLA of $40.00 then the absolute of the GLA residual as a percentage would be: = Absolute ($ $40.00) / $50.00 = Absolute ($10,000) / $50.00) =20% The subject property s characteristics do not change like the sales do. They are fixed and are identified in the sections describing each of the seven master models. This was done so that the accuracy of the value predictions could be measured based on the particular model, not based on the subject s physical make up. Keep in mind that the more similar the subject is to the sales the more accurate the value opinion will be. Each of the seven models is run 2,000 times at three different levels of random variance; +-8%, +-12%, and +-16%. The variable values and the coefficients for the sales randomly change within specific ranges. So while each model is structured to be progressively weaker, each model is considered at three levels based on the amount of random variance built into the model. Random Variance is built into the models by adding a column of data that randomly changes between the +- range used for each property sale (observation). This is achieved by the RandBetween function within MS Excel. The random variance is what makes the hypothetical data more realistic. It substitutes for a whole host of things that occur in real markets such as: Variation from market inefficiencies. This is best described by the range that is common in a market in the sales prices. For example, it may be typical for a property worth $100,000 to sell within the range of $90,000 to $110,000. This would be a range of +-10%. Errors in the data such as bedroom counts, sales price, age, land size. 12

13 Inconsistencies in qualitative data such as; quality, visual appeal, functionality of the floor plan, quality of the view, location rating, etc. Changes in market conditions such as; interest rates, sudden changes in demographics, etc. that are not easy to apply to specific submarkets within the subject property s over-all market. Not considering all of the predictive variables in the analysis. This is a partial list, perhaps you can think of other things. In each model a run that results in any variable having a P-factor greater than 0.1 is aborted and another run is started. In this situation the aborted run is not counted as one of the 2,000 runs. The following items are monitored: 1. The actual and predicted values. 2. The actual and predicted coefficients for each variable considered. 3. The absolute residual in the value as a percentage. 4. The absolute residual of each variable considered as a percentage. 5. The coefficient of determination (R 2 ). 6. The correlation coefficient, when only two variables are considered. 7. The maximum absolute residual in each run is recorded as a proxy for random variance. 8. The number of times residual of the value is greater than 10%. 9. The number of times residual of the value is greater than 15%. 10. The number of times the variable coefficient s absolute residual is greater than 33%. 11. The number of times the actual variable is outside of the 90% confidence interval. 12. The number of times the 90% confidence interval is greater than a +-30% of the predicted coefficient. 13

14 This data will permit an understanding of how effective regression analysis is when using varying degrees of quality in the data-sets. This effectiveness is considered for predicting value, and estimating the coefficients of the variables (extracting line-item adjustment rates). It will also allow for the identification of quantifiable regression outputs that will gauge the quality of the value and coefficient outputs. Thresholds Numbers 8, 9,10, and 12 in the above list of monitored items lists the thresholds of 10%, 15%, 33% and 30% respectively. These thresholds or limits were chosen somewhat arbitrarily, but based on practical logic. These are the points that the information or returns are not useful. For example in most low to medium scope of work assignments a residual for value of 15% or less is acceptable. The number of times the residual exceeded 10% added useful information to the study. A wider range of 33% was chosen for the variable coefficient residuals. Consider the following examples for the GLA and a swimming pool: Actual Low High GLA $50.00 $33.50 $66.50 Pool $12,000 $8,040 $15,960 In this example if the returned GLA coefficient is above $66.50 or below $33.50 then that run would be counted as the actual coefficient being over 33% off. When the pool has a value of $12,000 any return above $15,960 or below $8,040 would be counted. If the predicted coefficient is farther from the actual then this would not be useful to the analyst and could even cause more harm than good. This analysis allows us to gather information about the confidence interval (CI). The study counts the number of times the actual coefficients are outside of the CI, and the times the CI is greater than a +-30% of the predicted value (a total range of 60%). This 14

15 30% is arbitrary, but meant to approximate the point that the CI is no longer useful. At +-30% range the analysis may begin to lose credibility. Imagine your family is on a road trip and you have a quarter of a tank of gas. You are in a very remote area and stop at the only gas station nearby. They are out of gas, but the tanker should arrive within four hours with more gas. You can travel another miles on the gas you have. Your legitimate question to the attendee is, how far to the next gas station? If his answer is that the next station is somewhere between 75 and 150 miles then the range is too wide to be useful. In valuing real property there is a point that a CI is too wide to be useful. I have chosen +-30% (60%) interval as that point. The 90% CI was chosen over the 95% because the 90% has more narrow ranges. At the 95% range the intervals often become very large. Only two or three variables are considered in these models. This is because using few variables allows for monitoring the correlation coefficients, and because the Monte Carlo simulation is set up as an automatic loop which does not allow for a step-wise analysis which would be necessary with additional variables. QUESTION: What percentage of times should the actual coefficient be outside of the 90% CI? 15

16 Model (A) In this model two of two predictor variables are considered; [GLA] and [AGE]. These variables are the only property characteristics that have an influence on the value. Having only two variables allows us to monitor the correlation coefficient. This model is designed for the correlation coefficients to be low (typically below 0.3). The following table shows the information for this model. Subject Variable Range Coefficient Range GLA 1,800 1,600 2,400 $40.00 $80.00 AGE $2,000 -$5,000 This model is considered to be a very good data-set based on a very simple market. Real property markets and the associated data-sets will never be this good even when +-16% random variance is added in to the sales prices. Model (B) This model is very similar to Model (A) except it has more correlation built in between the two variables. To achieve this increased correlation the [AGE] ranges from 15 to 30 for the sales. The subject s value is based on being 25 years old. This model is slightly more complex than Model (A) based on the added correlation, but is still better than real property markets. Subject Variable Range Coefficient Range GLA 1,800 1,600 2,400 $40.00 $80.00 AGE $2,000 -$5,000 16

17 Model (C) In this model a third variable [VIEW] is added to the market, but this predictor variable is omitted in the model. [VIEW] is considered in calculating the real value, but it is not considered in the regression analysis. This is realistic in that most regression analyses omit one or more variables (property characteristics). In this model there is a high degree of correlation between the [GLA] and the [Age], but the [VIEW] is purely random which usually results in a low correlation. Subject Variable Range Coefficient Range GLA 1,800 1,600 2,400 $40.00 $80.00 AGE $2,000 -$5,000 VIEW $500 $2,000 This model is more realistic than Models (A) and (B), but when the +-12% and +-16% random variance is added in begins to be more realistic. Model (D) In this model there are three variables considered in the market and the model. The third variable is a [POOL]. In this model none of the variables have a correlation coefficient above 0.2 with one another. Subject Variable Range Coefficient Range GLA 1,800 1,600 2,400 $40.00 $80.00 AGE $2,000 -$5,000 Pool $8,000 $16,000 This model like Model (C) is somewhat realistic when the +-12% and +-16% random variance is added in. 17

18 Model (E) This model is much like Model (D) except there is more correlation between [GLA] and [AGE]. In this model the correlation coefficient between the Pool and the other two variables cannot exceed 0.2. Subject Variable Range Coefficient Range GLA 1,800 1,600 2,400 $40.00 $80.00 AGE $2,000 -$5,000 Pool $8,000 $16,000 Like Model (D), this model is somewhat realistic when the +-12% and +-16% random variance is added in. Model (F) In this model the market is made up of two variables; [GLA] and [AGE], however, the regression model only considers the [GLA]. The [GLA] is the dominant variable (has more influence on the value than [AGE]. There is a low degree of correlation between the two variables. Subject Variable Range Coefficient Range GLA 1,800 1,600 2,400 $40.00 $80.00 AGE $2,000 -$5,000 This is considered a very weak data set as it only considers one of two predictor variables. 18

19 Model (G) This model is very similar to Model (F) except it has more correlation built in between the two variables. Subject Variable Range Coefficient Range GLA 1,800 1,600 2,400 $40.00 $80.00 AGE $2,000 -$5,000 This is considered an unacceptable data set because it only considers one of two predictor variables and these variables are highly correlated. SCORE CARDS The seven models have simple markets and data-sets that range from unrealistically good to unacceptable. Let s see how the regression analysis faired in returning the over-all property value and the individual coefficients. 19

20 Model (A) The analysis did well predicting the coefficient for the [GLA]. Even at +-16% variance it only missed by more than 33% (six times out of 2,000). The analysis did not do so well for the [AGE] coefficient missing >33%, 10% and 16.8% of the time for the variance ranges of +-12% and +-16% respectively. The confidence interval (CI ) for [AGE] exceeded +-30% of the absolute predicted coefficient 23.2%, 47.4% and 62.0% of the time. Note that the actual coefficients fell outside of the CI from 8.4% to 11.1% of the time. This is exactly as expected with a 90% confidence interval. The analysis never missed the actual value by more than +-10%. 20

21 Model (B) In this analysis the [GLA] coefficient was accurate, they were within the CI about 90% of the time, and the CI was not too wide. The values predicted were very accurate with only one sale out of the 6,000 runs having a residual greater than 10% and none being greater than 15%. However, the residual for the age [AGE] coefficient was often greater than 33%, and CI was typically greater than a +-30% wide. 21

22 Model (C) In this model the related [GLA] and value information and returns were good even at the +-16% random variance level. The information and returns for the [AGE] were similar, but a little worse in this model than in Model (B). 22

23 Model (D) Again, in this model the information and returns for the [GLA] coefficient and the value fared well across the board. However, the related information for the [GLA] and the [POOL] coefficients were not reliable. 23

24 Model (E) This model had the similar results as Model (D) except the model tended to do a little better when predicting the [POOL] coefficient. 24

25 Model (F) In this model the subordinate variable [AGE] influenced the sales prices, but was not considered in the model. It had a low degree of correlation between the variables. This is a very poor dataset, but the model did fairly well. 25

26 Model (G) This model is very similar to Model (F), but has significant correlation between the two variables. It is considered to be a very bad data-set. The information returned on the [GLA] coefficient is generally not reliable. The value was also stressed with many residuals exceeding 10%. However, even in this bad data-set the value residuals seldom exceeded 15%. Note that this is the first time that the number of times the actual [GLA] coefficient fell outside the CI a significant number of times (in fact most of the time). Coefficient Confidence Intervals A 90% confidence interval was used in the Twenty-One Market Analysis. This tool was very accurate in that only about 10% of the actual coefficients values were outside of the interval returned. However, many times the CI became too wide to be useful. Further, my studies indicate that there is no correlation between the size of the interval and the accuracy of the coefficient returned. These are the main reasons that the confidence intervals for the coefficients are not utilized in the Regression Plus at this time. However, confidence intervals do provide some useful information and as the 26

27 appraisal industry evolves in its understanding of regression analysis this tool will then do more good than harm. 27

28 Interpretation of the Score Cards Premise 13 The returned coefficients for subordinate variables are often inaccurate. Premise 14 The confidence intervals for the coefficients are often so wide they are misleading. Premise 15 The predicted values are typically more accurate than the coefficients. These premises are borne out by the data presented in the Score Board section above. An indication of the accuracy of the value opinion can be provided by a residual analysis of the actual sales, and then measuring the fit of the subject property to the data. These analyses cannot be performed on the predicted coefficients. The question is: What single statistical output returned by the analysis can be relied on to determine if the coefficients returned are indeed accurate? To answer this question let s examine the correlation matrixes that correspond to Models (A) through (G) in search of an output that will alert the analyst to an inaccurate coefficient. The correlation coefficients that relate the coefficient of determination, variance, correlation, and number of sales to the value residuals and the variable 28

29 coefficient variables are highlighted in yellow. A correlation coefficient of 0.3 or greater is chosen as being significant. Correlation Matrix for 12,000 runs of Markets (A) & (B) Abs Correl Residual GLAco Residual Value Residual AGEco Num Sales AdjR2 AdjR2 1 Abs Correl Residual GLAco Residual Value Residual AGEco NumSales None of the relationships have very high correlation coefficients. In these two models the correlation matrix indicates that the adjusted R 2 has a relationship with the value residuals and the residuals of the coefficients. The number of sales and the correlation do not seem to be related to the residuals. Correlation Matrix for 6,000 runs of Market (C) Abs Correl Residual GLAco Residual Value Residual AGEco Num Sales AdjR2 AdjR2 1 Abs Correl Residual GLAco Residual Value Residual AGEco NumSales This matrix indicates the same thing as the matrix for Models (A) and (B) except to a little less degree. 29

30 Correlation Matrix for 12,000 runs of Markets (D) & (E) Residual GLAco Residual AGEco Residual POOLco Residual Value Num Sales AdjR2 AdjR2 1 Residual GLAco Residual AGEco Residual POOLco Residual Value NumSales This matrix shows a weak relationship between Adjusted R 2 and the residuals. It does not indicate any other relationships with the residuals. Correlation Matrix for 12,000 runs of Markets (F) & (G) Abs Correl Residual GLAco Residual Value Num Sales AdjR2 AdjR2 1 Abs Correl Residual GLAco Residual Value NumSales This matrix indicates a relationship between the absolute correlation (between the [GLA] and [AGE] prior to the analysis) and the residuals. This makes sense considering that the two variables were highly correlated and the [AGE] was omitted from the model. 30

31 Premise 16 A study of the correlation matrixes indicates that while the coefficient of determination, the amount of variance, and the correlation all tend to influence the quality of the predicted coefficients, none of them definitively identify when the returned coefficient is accurate. Number of Sales used in the Analysis The matrixes do not indicate a relationship between the number of sales used and the residuals. However, in my first model designs I used a fixed number of 30 sales. I later set the number of sales to vary between 30 and 100 sales. There was a distinct drop in the number of times the absolute residuals for the variable coefficients were greater than 33%. This indicates that the number of sales does influence the quality of the model. 31

32 CONCLUSION Regression analysis is appropriate when the expectations (based on due diligence) does not exceed the limitations imposed by: 1. A very inefficient market 2. An inactive market 3. Inaccuracies in the information on the sales data 4. Extent of multicollinearity. 5. Having particular property characteristics that are too similar. 6. Not having enough hits for a particular variable. For example if the model considers five variables and the expectations are to extract the line-item adjustment rates to a statistical certainty; then for most real property data-sets the analysis would not be effective enough to meet the expectations. Let s address the question put forward earlier: What single statistical output returned by the analysis can be relied on to determine if the coefficients returned are indeed accurate? 32

33 There is no one single statistical output that is meaningful for every situation. The best way to get reliable coefficients is to create the best market model possible. This is accomplished by: Including all of the comparable sales available. Assuring the data-set is of reasonable quality. o Ample number of sales o Reasonably accurate data o Inclusion of all predictor variables o Proper scrubbing and presentation of the data. Assuring that a properly performed regression analysis is performed: o That all of the regression assumptions are met. o A step-wise or similar process is properly employed after each modification to the model. o Experience in valuation techniques and market behavior. o Experience in the subject s over-all market. o Applying all of the various testing methods available. o Meeting the applicable regression assumptions. Linear relationship between the variable and value. Residuals are normally distributed, and have a zero mean. There is no heteoscedasticity. There is no auto correlation. The X variables are non-stochastic. There is no multicollinearity. 33

34 The conclusions of the Twenty-One Market Study indicate that a much greater due diligence is required for extracting line-item adjustment rates than to simply form an opinion of the value of the property. This due diligence within the realms of actual market behavior and the limitation of actual data-sets will not only require much more field work, but also an understanding of intermediate to advanced analysis techniques. Two advanced statistical techniques that might be used to aid the analyst in returning better quality coefficients are variance inflation factors and ridge regression. The problem with these and other methods, as well as many testing techniques, is that they require a good data-set to be effective. If the valuation analyst had a good data-set to begin with he/she would typically not need these techniques to begin with. Real property data-sets are typically not high quality. They will definitely improve over the next decade. It is important to understand that some functions in regression such as extracting line-item adjustment rates, are limited by imperfections in the data-set; and frankly by the nature of real estate markets in general. 34

35 It is the author s opinion that the novice level analyst with valuation training and experience will be able to successfully perform the following analysis with regression analysis: Form a low to medium scope of work opinion of a property s value, quickly and with a minimum of effort- most of the time. Form a high level scope of work opinion of value within certain markets. This process will require an extensive amount of effort and time in: o data gathering and verification; o data scrubbing and design and form of the data presentation; o model building analysis, and o testing methods and techniques Extract line-item adjustments: o To a limited degree (number and accuracy) when a high scope of work has been performed. Even then many adjustment rates cannot always be extracted. Some of the reasons are: There are is an insufficient number of sales. The ratings in the data-set are too similar. There are too few hits in the data-set. The property characteristic is not a predictor variable (there is no relationship between the variable and value). The variable has very little weight in the market model. 35

36 Many people mistakenly believe that all regression analysis return coefficients that are a statistical fact. In theory, regression analysis can return perfect coefficients when its analysis is based on perfect data. In the real world the data most valuation analysts have access to is, let s just say, far from perfect. Yet, I have heard many appraisers say that extracting line item adjustment rates with regression is not an opinion, it is a statistical fact. These appraisers are not only ignorant about what regression analysis can and cannot do; they do not know the secret statistical nerd club code. I will reveal it to you now. It is the acronym of Regression- Line-item adjustment rates (answer presented upside down below): This really applies more to those who ignorantly profess regression analysis as a means of producing statistical facts, rather than to the returned coefficients. On a more serious note, regression analysis does not provide facts it provides additional evidence of a credible analysis. It is an additional tool of extracting line-item adjustment rates; not the end all- be all. Even with regression analysis the conclusions (value and coefficients) are opinions not facts. Therefore, the qualifications of a valuation analysis are: An understanding of valuation principles. Training in valuation techniques, including regression analysis. Experience in performing all aspects of valuation including, problem identification, scope of work analysis, data-gathering, a wide spectrum of analysis techniques, and reporting. Geographical competency (experience in the subject s market). 36

37 I am curious of what I.Q. you think the appraiser in the following story has: There was an appraiser who needed to perform a property inventory from the curb of the road (a drive-by) in the metropolitan city some 50 miles away. However, the car would not start and according to the mechanic it would not start until he got a replacement part from that very metropolitan city. The appraiser s client was in a hurry so he immediately took his wife s car and drove to the city and retrieved the part. He took the part to the mechanic and pleaded with him to make the repair as quickly as possible so he could drive to the city, make his property inventory and complete the appraisal. What did this appraiser do wrong? He should have gone by the subject property while he was in the city picking up the part as he was already there. Well, that story is very much like the appraiser who was asked by his client to form an opinion of a property s value. The appraiser performed a regression analysis and extracted two line-item adjustments. He then set up three sales on an adjustment grid and proceeded to form an opinion of value based the comparison method. As these things go he ended up trying at least 6 sales before choosing the best three. He was not sure about all of the adjustment amounts, but did like the ones he extracted with regression. A few hours later he was done. What did this appraiser do wrong? According to the Twenty-One Market Study regression analysis tends to return a reliable indication of value, often more reliable than the returned line-item adjustment rates. This appraiser should have stopped after he ran the regression analysis as he had the value opinion at that point. Finding the value with regression and then extracting adjustments to re-find the value by the comparison method is a lot like the appraiser who drove to the city for the car part and then drove back to the city later to see the property. In both of these scenarios he was already there, and then went back a second time. 37

38 While I profess that for many assignments there is no need to perform an additional quantitative (URAR style adjustment grid), I do recommend the appraiser perform a very limited qualitative analysis in a narrative format. This separate analysis includes: 1. Identifying three sales from the sales used in the regression analysis that bracket the subject property. 2. Research the properties for sales concessions etc. 3. Include photos of the sales. 4. Locate the sales and subject property on a map. 5. Then comment on why one property is superior, one is about equal, and one is inferior. 6. Identify the value range of the subject property. 7. Report this analysis in a narrative format in from four to eight sentences. 8. Reconcile the two methods used (regression and comparative methods). This dual method provides a very dependable analysis that will be accepted by many clients. Obviously this cannot be done for secondary market providers, FHA, and VA; at least not now. 38