Extrapolation: Concepts and Techniques

Pam Perlich Urban Planning 5/6020 University of Utah

Learning Objectives
1.

Extrapolation
Concepts, assumptions, limitations Alternative functional forms linear and nonlinear

2.

Simple linear regression

Computational basis Alternative techniques in Excel

3.

Calculating forecasted values from fitted functions in Excel

Part 1: Extrapolation

Forecasting Context
Uncertainty (forecasting error) increases with
Longer forecast horizon Smaller areas

Extrapolation techniques have a higher probability of success in

Short time horizons Large areas

Extrapolation Technique
Fit function to a set of observations and extend this pattern into the future Use the function that
Is the function of best fit
(Least squares or regression)

Approximates our best understanding of future conditions

Incorporate growth constraints or known conditions

Assumptions
Use of aggregate data, generally across time (population, employment, etc.) Future movement of the data series is determined by past patterns embedded in the series The essential information about the future of the data series is contained in the history of the series Past trends will continue into the future

Advantages / Benefits
Computational simplicity Transparent methodology Ease of application May work for
Large areas Short time horizons Slow grow areas

Disadvantages / Risks
Does not account for underlying causes / structural conditions
Example: Cohorts are invisible

Ignores structural / systemic context Current trend often do not continue Excludes any external considerations

Alternative Functional Forms (Klosterman)

Linear constant increments of change Unbounded Geometric constant rate of change Parabolic Accelerating growth rate Modified Exponential growth limit Often preferred Gompertz growth limit for small area Logistic growth limit forecasts Explore these in spreadsheets:
http://home.business.utah.edu/bebrpsp/URPL5020/Trend/

Klostermans Technique
Klosterman
Transforms curves into lines Performs linear regression

Some functions are not available in Excel (e.g., Gompertz, modified exponential, logistic) His approach can be applied in Excel so that these alternative functions are available.
Transform data according to his technique Fit a trend line using the Excel function Reverse the transformation to compute forecasted values.

Linear Function
Linear Function: Y=10X+100
450 400 Dependent Variable 350 300 250 200 150 100 50 0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Inde pe nde nt Variable

= a + bX

where a Y is the intercept and b is the slope

Constant increments of growth

Geometric Function
Geometric Function:
200 180 160 140 120 100 80 60 40 20 0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 Inde pe nde nt Variable

= 10(1.1X ) YC

Dependent Variable

= ab X YC

where a is the intercept b is the growth rate plus one

Constant rate of growth

Parabolic Function
Parabolic Function
2,000
Dependent Variable

= 10 + 1.5 X + 2 X 2 YC

1,500 1,000 500 0

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Independent Variable

= a + bX + cX

where a is the Y intercept and b is the slope

Constantly changing slope If b>0 Growth is accelerating

Modified Exponential
Modified Exponential Function:
250 200
Dependent Variable

= 200 100 * (0.8) X

150 100 50 0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Independent Variable

= c + ab X YC

where a is c minus the Y intercept b is the ratio of successive growth increments (constant) c is the asymptotic value

Gompertz Function
Gompertz Function
120 100
Dependent Variable

= 100 * (0.9)

0.8 X

80 60 40 20 0
0 2 4 6 -8 -6 -4 -2 8 10 12 16 18 14 -1 4 -2 0 -1 8 -1 6 -1 0 -1 2 20

Independent Variable

Y C = ca

bX

If ln(a)<0 with 0<b<1 C is the upper limit Ratio of the logarithms of successive observations is constant

Logistic Function
Logistic Function
0.120 0.100
Dependent Variable

YC =

1 10 + (0.5) * (0.5) x

0.080 0.060 0.040 0.020 0.000

-8 -6 -4 -2 0 2 4 6 8 10 12 14 16 18 -2 0 -1 8 -1 6 -1 4 -1 2 -1 0 20

Independent Variable

1 Y C = c + ab x

If 1. b is between 0 and 1 and 2. a < 0 Then 1. Curve takes the S shape and 2. 1/c is the asymptotic value (upper limit) and 3. 0 is the lower limit

Excel Tool to Fit S-Curve to Data

Developed by Stephen R. Lawrence of University of Colorado
http://leeds-faculty.colorado.edu/Lawrence/Tools/SCurve/scurve.xls

Algorithm for fitting a logistic function to a set of data.

Extrapolation Method - Summary

Simple technique that may be the most appropriate for
Tight time constraints Slow changing conditions Short time frames General trend identification

Judgment must be exercised or results may potentially be absurd

Part 2: Simple Linear Regression

Regression

Fitting Equations

The regression (least squares) technique is used to:

Establish a trend in time series data Extend this pattern into the future

Given the existence of a time trend, fitting equations enables us to identify the mathematical function that best captures the relationship

Meanings of Coefficients
R2 is the regression coefficient 0<R2 <1
0 1 No relationship Perfect fit

R : correlation coefficient
Square root of R2 and signed according to the direction of the relationship -1<R<1
1 -1 0 Perfect fit, positive relationship Perfect fit, inverse relationship No relationship

Fit a Function to the Data

Fit a Function to the Data

Deviation = Observed Minus Fitted

Fitted Value

Deviation

Observed Value

Least Squares Method

Find the line that minimizes the squared differences between the observed dependent variable and the calculated dependent variable. y=ax+b is the linear function (x,y) is the observed pair. Minimize the sum of all (y-y)2

Solve these Simultaneous Equations

a x i + bn =a y i
i =1 i =1

a x i + b x i = x i y i
2 i =1 i =1 i =1

Example: Scatter Plot

X - Y Scatter of Data
12 10 8 6 4 2 y 0 0 5 10 15 20 Linear (y) y = 0.5147x + 1.2794 R = 0.9577
2

From: Gottfried, page 106

Matrix Algebra Solution: Covered in the Next Section of Course

In Matrix Form 41 479 5 41 a b 27.5 299

Inverse

a b a b

-0.0574 0.6709 0.514706 1.279412

0.0070 -0.0574

27.5 299

From: Gottfried, page 106

Goodness of Fit Measure

Sum of Squared Errors

SSE = [ y i f ( x i )]
i =1

Goodness of Fit Measure

r-squared

SSE r = 1 SST
2
Where:

SST = [ y i y ]
i =1

Values of r2
0<r2<1 As r2 approaches 1, the fit is better As r2 approaches 0, the fit is worse

Calculating r2 in Excel
I 1 2 3 4 5 x 2 4 7 11 17 y 2.00 3.50 4.50 8.00 9.50 5.5 Error Error^2 f(x) y-f(x) (y-f(x))^2 y-(AveY) (y-(AveY))^2 2.308824 -0.308824 0.095372 -3.50 12.25 3.338235 0.161765 0.026168 -2.00 4.00 4.882353 -0.382353 0.146194 -1.00 1.00 6.941176 1.058824 1.121107 2.50 6.25 10.029412 -0.529412 0.280277 4.00 16.00 1.67 39.50 SSE SST

Average of Y

R^2 = 1 - (SSE/SST) 0.957743857 is the R squared

From: Gottfried, page 106

Spreadsheet online: http://home.business.utah.edu/bebrpsp/URPL5020/Matrix/LinearRegression.xls

Add a Trend Line in Excel

Plot the data in an x-y scatter Right click the series on the graph Select Add a Trend Line Select Linear from the Type tab Select Display Equation and Display r squared from the Options tab Note: for other applications you may select Forecast as well

Using the Analysis Tool Pack

Enter the data into the worksheet From Tools menu, select Data Analysis/Regression Tool. Complete the required selections.

Results from Analysis Tool Pack

x 2 4 7 11 17 y 2 3.5 4.5 8 9.5 SUMMARY OUTPUT Regression Statistics Multiple R 0.978643887 R Square 0.957743857 Adjusted R Square 0.943658476 Standard Error 0.745903847 Observations 5 ANOVA df Regression Residual Total 1 3 4 SS MS F Significance F 37.83088235 37.83088 67.99559 0.003734402 1.669117647 0.556373 39.5 Lower 95% Upper 95% Lower 95.0% Upper 95.0% -0.664886955 3.223710484 -0.664886955 3.223710484 0.316059694 0.71335207 0.316059694 0.71335207 12 10 8

X Variable 1 Line Fit Plot

6 4 2 0 0 5 10 X Variable 1

Y Predicted Y

15

20

Intercept X Variable 1

Coefficients Standard Error t Stat P-value 1.279411765 0.610944132 2.094155 0.127272 0.514705882 0.062419278 8.245944 0.003734

Simple Linear Regression - Summary

Simple linear regression fits a line to a set of x,y coordinates. This procedure minimizes squared errors. r2 is a measure of goodness of fit
The better the fit, the closer r2 is to 1

There are multiple ways to compute linear regression in Excel.

Part 3: Calculating Forecasted Values from Fitted Functions in Excel

http://home.business.utah.edu/bebrpsp/URPL5020/Trend/CalcExtrap.xls

Worksheet examples

Forecasted Values in Excel

Select data series Right click Add a trendline Select Type
Linear

Go to Options tab

Forecasted Values in Excel

Options Menu
Forecast period Display equation Display R-squared Note: You can customize trend label Click OK

State of Utah Population

3,500,000 3,000,000 2,500,000 2,000,000 1,500,000 1,000,000 500,000 0 1940

Tip: Select the formula label, then format, and increase the number of digits to the largest possible. This will result in a more precise computation.

y = 29481.450370525x + 316996.533799534 2 R = 0.962332124

State of Utah Trend Forecast

1950

1960

1970

1980

1990

2000

2010

2020

State of Utah Population

3,500,000

y = 29,481.45x + 31,6996.53
3,000,000

2,500,000

2,000,000

Tip: For final presentation purposes, reduce the number of digits displayed in the equation.

R = 0.962

1,500,000

1,000,000

State of Utah Trend Forecast

500,000

0 1940

1950

1960

1970

1980

1990

2000

2010

2020

Calculating Forecasted Values

Calculate the forecasted population for the year 2020. Equation: y = 29481.450370525x + 316996.533799534
Y = population X = time marker substitute the year (?)

29481.450370525*(2020)+ 316996.533799534

= 59,869,526
This is much too high.

Calculating Forecasted Values

Data series is 1940 through 2020. If the actual year does not work, create an index for each year, incrementing by 1.
1940 =1, 1941 = 2 , etc. 2020 = 81

29481.450370525*(81)+ 316996.533799534 = 2,704,994 This is the correct formula. You can determine how your version of excel interprets the x in your equations by experimenting. See example spreadsheet. CalcExtrap.xls

Residuals: Forecast Minus Actual

200,000 150,000 100,000 50,000 0 -50,000 -100,000 -150,000 -200,000 -250,000 -300,000 -350,000

1940

1950

1960

1970

1980

1990

2000

Residuals: Forecast Minus Actual

200,000 150,000 100,000 50,000 0 -50,000 -100,000 -150,000 -200,000 -250,000 -300,000 -350,000

Trend line: y = 0.00x + 0.00

Linear regression on residuals collapses to the x axis. The sum of the residuals is zero.

1940

1950

1960

1970

1980

1990

2000

Ratio Methods
Smith, Tayman, Swanson Chapter 8 Smaller region (city) is contained in larger region (county or state) Projection of larger region projection of smaller region Depends upon a preexisting forecast / projection of the larger region These will be used in the economic models section of the course.

Types of Ratio Methods

Constant share: small area maintains same growth rate as larger area Shift share: trend in small areas share of region is extended into the future
Observed differential growth rates are maintained

Share-of-growth: small areas share of larger regions growth is maintained.

Extrapolation - Summary
Use with care.
Just because a function fits (high r2) does not mean that the extrapolation is reasonable. Make your assumptions explicit Generally there are growth limits at some point

Explore various approaches.

Create your own functions in excel based on your knowledge of the area (growth limits, etc.) Use Excel to fit a trend line and extrapolate it into the future

Calculation of the forecast value when using Excel may require the construction of an index.
Use the reported equation and substitute either the year or index number into the formula for x. If you create an index, the beginning value should be 1.