Demand Estimation & Forecasting

EC611--Managerial Economics
Dr. Savvas C Savvides, European University Cyprus

Marketing Research Approaches (1)

Consumer Surveys: Questioning a sample of consumers to gauge their Observational Research: This is a supplementary method to the
response to changes in the explanatory vbls. (Px, Py, I, Features, etc)not very reliable. Yet, sometimes they may be the only option!

Consumer surveys using product scanners and people meters at homes (e.g. for TV viewing) to observe changes in consumer behaviors.

Consumer Clinics: Laboratory-style experiments, where a sample of

consumers are given a budget and are asked to simulate their spending/buying habits in a simulated store or market. The objective is again to investigate how they respond to changes in P, Product features and promotional alternatives). This can also be simulated on computer via virtual shopping.

Market Experiments: Similar in concept to the consumer clinics, but the

experiments are in the actual marketplace. Different markets with similar demographic and socioeconomic characteristics are chosen. In each market only one controlled variable is changed (P, packaging, promotional tricks). They are expensive (large scale), but useful and effective because they are effective in determining best pricing strategies, and for testing different product features and promotional activities.
Managerial Economics DR. SAVVAS C SAVVIDES 1

Marketing Research Approaches (2)

Virtual Management
This is a very sophisticated method for simulating consumer behavior using computer models based on the theory of complexity, without relying on actual consumers to simulate actual shopping behavior. The systems relies on consumer behavior parameters that are programmed in the computer model, based on information from consumer surveys and other database information This is potentially a very powerful tool that can allow managers to test the impact of various managerial decisions (e.g, changes in P, product features, outlet layout, customer service personnel, etc)
Managerial Economics DR. SAVVAS C SAVVIDES 2

Estimation Techniques
These are various quantitative methods to find the exact relationship between the dependent variable and the independent variable(s). The most common method is regression
Simple (bivariate) Regression: Y = a + bX Multiple Regression: Y = a +bX1 + c X2 +dX3 +...
Managerial Economics DR. SAVVAS C SAVVIDES 3

analysis

Regression Analysis
Adv. Sales

Year 1 2 3 4 5 6 7 8 9 10

X 10 9 11 12 11 12 13 13 14 15

Y 44 40 42 46 48 52 54 58 56 60

Scatter Diagram

Managerial Economics

DR. SAVVAS C SAVVIDES

Regression Analysis
Regression Analysis: This is the statistical technique used to obtain an imaginary line that best fits the data points (the Regression Line or Line of Best Fit). It is used to:
derive estimates of the slope coefficients (a, b, c, ) estimates of how changes in each independent (explanatory) variable impacts on the dependent variable Conduct test of significance (hypothesis testing) Construct confidence intervals Test for the overall explanatory power of the regression

Ordinary Least Squares (OLS): This is the method which minimizes the sum of the squared vertical deviations or residuals (et) of each actual observation point from the regression line.
Managerial Economics DR. SAVVAS C SAVVIDES 5

Regression Analysis
Regression Line (Line of Best Fit)

Y1 = actual observation of sales ($44)

(Y-hat) = estimated value of Y at X=$10

e1 = Y 1

the residual or error term

Managerial Economics

DR. SAVVAS C SAVVIDES

Ordinary Least Squares (OLS)

Lets start with a simple two-variable model (say the sales-advertising model examined before) shown in the general form below:

Yt = a + bX t + et
Our objective is to find estimates of the slope coefficients (a and b) of the regression line

Yt = a + bX t where

, and are the estimates of the dependent variable, the intercept and the slope, respectively.

The error term (deviation) of the actual Y from its et = Yt estimated value is given by the equation
Managerial Economics DR. SAVVAS C SAVVIDES

Yt
7

Ordinary Least Squares (OLS)

Recall that the OLS is the method which minimizes the sum of the squared vertical deviations or residuals (et) of each actual observation point from the regression line. It enables us to determine the slope ( ) and intercept ( ) that minimize the sum of these squared errors. The formula for deriving the sum of squared residuals is:

et2 = (Yt Yt ) 2 = (Yt a bX t ) 2

t =1 t =1 t =1
Managerial Economics DR. SAVVAS C SAVVIDES 8

Ordinary Least Squares (OLS)

Estimation Procedure
The estimated value of the b coefficient is:

b=

(X
t =1 n t =1

X )(Yt Y )

( X t X )2

The estimated value of the a coefficient is:

a = Y bX

Managerial Economics

DR. SAVVAS C SAVVIDES

Ordinary Least Squares (OLS)

Estimation Example
Tim e
1 2 3 4 5 6 7 8 9 10

Xt
10 9 11 12 11 12 13 13 14 15 120
n X 120 X = t = =12 10 t=1 n

Y t
44 40 42 46 48 52 54 58 56 60 500
n Y 500 Y = t = = 50 10 t =1 n

Xt X
-2 -3 -1 0 -1 0 1 1 2 3

Y Y t
-6 -10 -8 -4 -2 2 4 8 6 10

(Xt X)(Y Y) t
12 30 8 0 2 0 4 8 12 30 106

(Xt X)2
4 9 1 0 1 0 1 1 4 9 30

n = 10

106 b= = 3.533 a = Y bX = 50 3.533 (12) = 7.60 30 Managerial Economics DR. SAVVAS C SAVVIDES

10

Ordinary Least Squares (OLS)

Therefore, the equation of the regression line is:

Yt = a + bX t

= 7.60 + 3.53 (Xt)

Substituting alternative observed values for Xt we derive corresponding values for Plotting these values for Xt and regression line. With Xt = 10
Managerial Economics

we obtain the = 7.60

For example, with Advertising (Xt) = 0, = 42.90 (vs. actual of 44)

DR. SAVVAS C SAVVIDES

11

Ordinary Least Squares (OLS)

The estimated regression line can be used to estimate the value of Y that may result from a value of X within or near the range of values used to derive the regression line. If Xs are much higher or lower that this range, then we cannot attach much confidence to the value of the estimated parameters, and ultimately on the estimated value of Y. The value of the slope coefficient of X ( ) measures the change in the dependent variable (Y) as a result of a change by one unit in the value of X. It is a marginal measure of the effect of X on Y = Y/X
Managerial Economics

= dY/dX (in derivative form)

DR. SAVVAS C SAVVIDES

In this sense,

is a slope!

12

The Standard Error

The regression results of our example are based on a sample period of ten years. How confident are we that the relationship we found is true for larger or different time periods? To test whether the values of the parameters we estimated indicate that Y and X are significantly related (in a statistical sense), we need the Standard Error (or Deviation) of the Slope Estimate (SEE). The formula of the standard error of is:

sb =

( n k ) ( X t X )

(Yt Y ) 2

( n k ) ( X t X ) 2

et2

The smaller the value of this error term in relation to the estimated value of the larger is our confidence that the value of is in fact significantly different from zero. In other words, that there is indeed a significant relationship between the actual Y and the actual X. TheManagerialindicates how DR. SAVVAS C the model can predict Y. 13 SEE Economics precisely SAVVIDES

Tests of Significance
Hypothesis Testing Null Hypothesis: The hypothesis that the value of b = 0. Alternative Hypothesis: The hypothesis that the value of b is non-zero. If we reject the null hypothesis (i.e., that b = 0), then we accept the alternative hypothesis that b is different from zero. This process is called Hypothesis Testing (or Significance Test).
Managerial Economics DR. SAVVAS C SAVVIDES 14

Tests of Significance
Example Calculation
Time
1 2 3 4 5 6 7 8 9 10

Xt
10 9 11 12 11 12 13 13 14 15

Yt
44 40 42 46 48 52 54 58 56 60

Yt
42.90 39.37 46.43 49.96 46.43 49.96 53.49 53.49 57.02 60.55

et = Yt Yt
1.10 0.63 -4.43 -3.96 1.57 2.04 0.51 4.51 -1.02 -0.55

et2 = (Yt Yt )2
1.2100 0.3969 19.6249 15.6816 2.4649 4.1616 0.2601 20.3401 1.0404 0.3025 65.4830

( Xt X )2
4 9 1 0 1 0 1 1 4 9 30

e = (Y Y )
t =1 2 t t =1 t t

= 65.4830

(X
t =1

X ) = 30
2

sb =

(Y Y ) (n k ) ( X X )
2 t t

65.4830 = 0.52 (10 2)(30)

15

Managerial Economics

DR. SAVVAS C SAVVIDES

Tests of Significance
Calculation of the t-Statistic
The basic test of significance is the t-test. This is done by dividing the estimated coefficient by its standard error

3 .5 3 b t = = = 6 .7 9 0 .5 2 sb
Degrees of Freedom = (n-k) = (10-2) = 8 Critical Value of t-statistic at 5% level =2.306 Rule of thumb: t-value of >2 If the estimated coefficient of a variable passes the t-test we can be confident that X (advertising) truly has an impact on Y (sales). In this case, where t=2.306, we have a 5% chance that in reality things are not like this. Managerial Economics DR. SAVVAS C SAVVIDES 16

Goodness of Fit
Coefficient of Determination (R2 ): This is a statistic to test the overall explanatory power of the regression, or the goodness of fit of the observed data around the regression line. It shows how well the regression explains changes in the value of the Y. It is defined as the proportion of total variation in Y that is explained by the full set of independent variables. This is shown by the formula:

ExplainedVariation R = = TotalVariation (Yt Y )2

2
Managerial Economics DR. SAVVAS C SAVVIDES 17

(Y Y )2

Goodness of Fit & Correlation

R2 can range from zero to 1. A value of 1 indicates that all the
variation in Y is explained by variations in the explanatory variables. The closer the observed data points fall on the regression line, the smaller the error terms (deviations); and the greater is the proportion of the variation in Y explained by the variation in X, the larger the value of R2. The R2 for the sales-advertising example is calculated as: 373.84

R2 =

440.00

= 0.85

Coefficient of Correlation: This is a measure of the degree of

association or correlation (co-variation) between X and Y

r = R 2 with the sign of b 1 r 1 In our example r is calculated as: r = 0.85 = 0.92

that X and Y vary (or move) together 92% of the time.
Managerial Economics DR. SAVVAS C SAVVIDES

This means
18

Regression An Example
Estimating the Consumption Function of Cyprus
Yt = a + bXt
where Yt is consumption, and Xt is income. b is the slope coefficient (or Marginal Propensity to Consume) which measure the impact (change) on Consumption per unit change in Income a is the vertical intercept (Consumption at zero Income)the constant! Regressing Consumption on Income by OLS we get the following results: = - 14.85 + 0.66 Xt t-value ( 74.45)

R2 = 0.9948
t =

n=31

b 3 .5 3 = 0.66 / 0.008819= 74.447 = 6 .7 sb 0 .5 2

The value of = 0.66 means that (on average) Cypriots spend (consume) 66% of any change in income. A 1m in income leads to 0.66m in Cons. Since the t-value of 74.45 is much greater than the critical value of t (2.045 with n-k = 29), we conclude that is statistically significant (5% level)
Managerial Economics DR. SAVVAS C SAVVIDES 19

Multiple Regression Analysis

Model:

Y = a + b1 X1 + b2 X 2 + L + bk ' X k '

To estimate multiple regression equations, we use essentially the same procedure as in the simple two-variable case. Here, however, the complexity of the calculations necessitate the use of the computer.

Analysis of Variance and F Statistic

Explained Variation /(k 1) F= Unexplained Variation /(n k )

R 2 /(k 1) F= (1 R 2 ) /(n k )

The F-stat is used to test the hypothesis that the variation in the independent vbls. as a set (the Xs) explains a significant variation in the dependent vbl. Y. Null Hypothesis: All regression coefficients are equal to zero.
Managerial Economics DR. SAVVAS C SAVVIDES 20

Multiple Regression Analysis

Adjusted Coefficient of Determination
R2 may be artificially high for small samples.

For meaningful, reliable estimates, the sample size must be sufficiently large (usually over 25-30 observations) to get enough degrees of freedom (n-k). Because R2 always approaches 100% as (n-K) 0, statisticians use a method for adjusting R2 to account for the degrees of freedom. The formula for the adjusted R2 ( denoted as R2bar) is the following:

( n 1) R = 1 (1 R ) (n k )
2 2
Managerial Economics DR. SAVVAS C SAVVIDES 21

Multiple Regression Example (1)

The Aggregate Consumption Function of Cyprus Revisited
(1) Specifying the Consumption Equation

Ct = a + b1Y + b2Ct-1+ b3P + b4DUMMY1974.

Y is Income (measured by GDP), Ct-1 is previous periods private consumption spending (to capture habit-forming behavior), P stands for inflation (as measured by the Retail Price Index), and DUMMY1974 is a dummy variable to capture the impact of the Turkish invasion (2) Estimating the Consumption Equation Regressing Consumption on Income, Inflation and Consumption lagged one period by OLS we get:

Ct = - 22.55 + 0.25 Y + 0.67Ct-1+ 5.59P 92.2DUMMY1974

( 4.13) (6.78) (1.57) (-2.67) n=31 Critical t-value = 2.056 n-1= 4 n-k = 26
22

R2 = 0.9982
F-stat = 3624.5
Managerial Economics

R2-bar = 0.9979

DR. SAVVAS C SAVVIDES

Critical F-value (3, 27) = 2.99

Multiple Regression Example (2)

Interpreting the Results R2 = 0.9982. This means that almost all (100%) of the variation in C is explained by s in Y, P & Ct-1 The value of 0.67, the coefficient of Ct-1, means that for every 1m spent last year leads Cypriots to spend 0.67m on current Consumption. Also, 1% increase in inflation leads to an increase in C of 5,59 million. However, since the t-ratio is below 2, we are not statistically confident that this is true. The betas for Y and Ct-1 are statistically significant (at the 5% level) > critical value of t (2.05) High value of F indicates that the regression as a whole explains a statistically significant portion of the variation in C. The critical value of F (3, 27) = 2.99
Managerial Economics DR. SAVVAS C SAVVIDES 23

Multiple Regression Example (3)

Forecasting Consumption in Cyprus Ct = - 22.55 + 0.25 Y + 0.67Ct-1+ 5.59P
Lets say that the Government estimates that private final consumption expenditures (excluding government consumption expenditures) last year (2003) were 4800 million, this year (2004) Gross Domestic Product is expected to reach 7200 million and this years inflation rate is expected to be 3%. When we substitute these values in the estimated equation for Consumption we get:

C2004 = -22.55 + 0.25 (7200) + 0.67 (4800) + 5.59 (0.03) = 5161 million (for a 7.5% growth) Income Elasticity = 0.25 (7200 / 5161) = 0.35 Managerial Economics DR. SAVVAS C SAVVIDES 24 aggregate consumption is regarded as a necessity (!!??)

Problems in Regression Analysis(1)

Multicollinearity:
Some explanatory variables move together over the same period. We are not able to distinguish the individual impact on the dependent vbl. All we observe is their combined impact. If current income changes in the same way as inflation over time, then well not be able to separate their impact on current consumption If advertising is a constant % of total budget, then regressing Sales on Advertising and Total Budget will lead to multicollinearity.

This leads to low t-values (even though R2 is high).

How do we correct for multicollinearity? Increase the sample size (collect more data) Transform the functional form of the equation (e.g deflate data) Delete one of the collinear variables Managerial Economics DR. SAVVAS C SAVVIDES 25

Problems in Regression Analysis(1)

Autocorrelation:
Arises when consecutive error terms have the same sign or change sign frequently. It may be caused by cyclical trends in the Xs, from the exclusion of significant vls. or from non-linearities in the data. Autocorrelation produces artificially high t-values, which erroneously may point to statistically significant parameters. The obtained values for R2 F-stat will be unreliable in the presence of autocorrelation. Autocorrelation is detected by the Durbin-Watson statistic, routinely produced by the computer. As a rule of thumb, a value of d=2 is indicative of the absence of autocorrelation. How do we correct for autocorrelation? Include time as an additional variable and/or Omit the constant Transform the functional form of the equation in non-linear form Include significant variables previously omitted
Managerial Economics DR. SAVVAS C SAVVIDES 26

Steps in Demand Estimation

1. 2.
important determinants of the dependent variable (e.g. Ct = a + b1Y + b2Ct-1+ b3P. Use theory and market/empirical knowledge.

Specify the Model : Identify Variables that are believed to be

Collect Data: Accurate data are essential to get reliable

estimates. Proxy variables: used when data for actual variables are lacking or incomplete or inaccurate (e.g. stock prices may be used as a proxy for wealth). Time-Series or Cross-Sectional data may be used.

3. 4. 5.

slope coefficients.

Specify Functional Form: linear, non-linear, log-linear, etc. Estimate Function: Run the regression to get estimates for the Test the Results: Run all the necessary tests of significance to

test for the statistical significance of each explanatory vbl, the explanatory power of the regression, (t-ratio, R2, F-test, etc), test for Economics Managerial presence of problems and proceed to correct them. 27 DR. SAVVAS C SAVVIDES

Functional Form Specifications

Linear Function: QX = a0 + a1PX + a2 I + a3 N + a4 P +L+ e Y Power Function: Q X = a ( PXb )( PYb )
1 2

To estimate a power function using the standard regression techniques, we linearize the function; in other words, we transform it into a linear form. To do this, we take the logarithms of the variables as follows:

ln QX = ln a + b1 ln PX + b2 ln P Y
Managerial Economics DR. SAVVAS C SAVVIDES 28

Demand Forecasting

EC611--Managerial Economics
Dr. Savvas C Savvides, European University Cyprus

Need for Forecasting

Long Range Strategic Planning
Corporate Objectives: Profit, market share, ROCE, strategic acquisitions, international expansion, etc

Annual Budgeting
Operating Plans: Annual sales, revenues, profits

Annual Sales Plans

Regional and product specific targets

Resource Needs Planning

HRM, Production, Financing, Marketing, etc

Managerial Economics

DR. SAVVAS C SAVVIDES

30

Selecting Forecasting Techniques

desired/required.

Hierarchy of Forecasts: The level of aggregation

The macro-level (GNP, disposable income, inflation, plant & equipment, retail sales, etc) Industry sales forecasts Individual firm sales forecastsby product line. region, etc

Some are simple and inexpensive, others are complex, time-consuming and expensive. Some are appropriate for ST planning, others for long-range planning / forecasting. Factors that impact on the selection of technique(s) are:
Cost-benefit for developing forecasting model Complexity of behavioural relationships to be forecasted The accuracy of forecasts required The lead time required for making decisions dependent on results of the model
DR. SAVVAS C SAVVIDES 31

Criteria used in Selection of Forecasts:

Managerial Economics

Prerequisites for Good Forecasts

Consistency with other parts of the business
sales forecasts are as good as the ability of the firm to manufacture and distribute products to the market (on time, good quality, etc)

Based on knowledge of past data

Exceptions: marketing new products, erratic changes in sales due to exogenous factors Gut feeling and market experience are relevant here

Consider political and economic factors

Government policy (taxation, trade, profit repatriation, political instability, FX considerations)

Accuracy vs. Timeliness

There is usually a trade-off
Managerial Economics DR. SAVVAS C SAVVIDES 32

Qualitative Forecasts (1)

Survey Techniques -- Macro Level
Plant & Equipment Spending Plans: Surveys
(quarterly, bi-annual) for expenditures on fixed assets and R & D (McGraw-Hill, Dept of Commerce, Fortune Magazine, conference Board, etc)

Plans for Inventory Changes & Sales Expectations (McGraw-Hill, Dept of Commerce, Dunn &
Bradstreet, NAPA)

Consumer Expenditure Plans: Aims at showing

trends in consumer income, asset holdings and consumer intentions to purchase specific goods, such as appliances, cars, homes, etc (Census Bureau, U of Michigan).
DR. SAVVAS C SAVVIDES 33

Managerial Economics

Qualitative Forecasts (2)

Opinion PollsMicro Level (Sales Forecasting) Business Executives: subjective views of top Sales Force: being close to the market, sales

management (often used in conjunction with quantitative forecasts by trend analysis, etc) employees have significant insight of the state of consumer psyche, and thus future sales. Care should be used not to confuse forecasts with sales targets

Consumer Intentions: mail surveys to estimate

consumer intentions of purchasing replacement , upgraded, or complementary (component) products

Managerial Economics DR. SAVVAS C SAVVIDES 34

Variation in Time-Series Data

Secular Trend
Due to Long-Run changes in economic data series (e.g. population, age distribution, tastes, etc)

Cyclical Fluctuations
Due to major expansions or contractions in economic data (usually >year). Example: housing sales, car sales

Seasonal Variation
Regular/rhythmic fluctuations in sales due to weather, habit, or social custom. Example: Tourism

Irregular or Random Influences

Example: wars, natural disasters, extraordinary government action (imposition of new taxes (VAT), nationalization, trade embargo, etc)
Managerial Economics DR. SAVVAS C SAVVIDES 35

Time-Series Variation (Graphs)

Managerial Economics

DR. SAVVAS C SAVVIDES

36

Trend Analysis & Projection

Linear Trend: St = S0 + b t
St

Easy, simple, but inflexible and unsophisticated. b = Growth per time period

St = So + b t

So Time (t)
Managerial Economics DR. SAVVAS C SAVVIDES 37

Linear Trend ProjectionExample (1)

Sales Revenue for Microsoft Corp. (1984-2001)--$ millions
Year
1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001

Time Trend
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

Actual Sales
100 140 202 346 591 804 1183 1843 2759 3753 4649 5937 8671 11358 14484 19747 22956 25296

Fitted Sales (Linear)

- 5042 - 3633 - 2224 - 815 594 2003 3412 4821 6223 7639 9048 10457 11866 13275 14684 16093 17502 18911

Managerial Economics

DR. SAVVAS C SAVVIDES

38

Linear Trend ProjectionExample (2)

Sales Revenue for Microsoft Corp. (1984-2001)--$ millions
25000

Linear Fitted Line Sales= - 6451.2 + 1409.0t (- 3.48) (8.22)

R2 = 0.808

15000

Actual Sales

1984

1998

2001

Time

-5000

Managerial Economics

DR. SAVVAS C SAVVIDES

39

Linear Trend ProjectionExample (3)

Forecasts of Sales Revenue for Microsoft Corp. (2002-2005)
Year
2002 2003 2004 2005

Time Period
19 20 21 22

Trend Line
S= - 6451.2 + 1409(19) S= - 6451.2 + 1409(20) S= - 6451.2 + 1409(21) S= - 6451.2 + 1409(22)

Fitted Sales (Linear)

20,320 21,729 23,138 24,547

The linear trend line model implies that sales increase by a constant value each year. In the Microsoft example, the model projects sales revenue to increase by $1409 million per year. The projections should not be far from the current period. From the previous graph, we see that the true trend line for Microsoft sales is non-linear. Thus, the forecasts are poor estimates of actual values. It may be more appropriate to assume that sales increase at a constant growth factor (rate) rather than a constant amount (value)
Managerial Economics DR. SAVVAS C SAVVIDES 40

Growth Trend Analysis/Projection

Constant Growth Rate Trend: St = S0 (1 + g)t g = Growth rate
To estimate, we linearise this behavioral relationship by taking logarithms (natural logs): ln St = ln S0 + ln (1 + g) x t
St

St = S0 (1 + g)t
So Time (t)
Managerial Economics DR. SAVVAS C SAVVIDES 41

Linear Trend ProjectionExample (1)

Sales Revenue for Microsoft Corp. (1984-2001)--$ millions
Year
1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001

Time Trend
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

Actual Sales
100 140 202 346 591 804 1183 1843 2759 3753 4649 5937 8671 11358 14484 19747 22956 25296

Natural log of Sales

4.60517 4.941642 5.308268 5.846439 6.381816 6.689599 7.075809 7.51915 7.922624 8.230311 8.444407 8.688959 9.067739 9.337678 9.5808 9.890757 10.04133 10.1384

Managerial Economics

DR. SAVVAS C SAVVIDES

42

Growth Trend ProjectionExample

Using this linearized regression technique to the Microsoft sales data (transformed in logs), we get the following results (t-values in parenthesis): ln St = 4.5695 + 0.33603 t
(38.58) (30.71)

R2 = 0.983

Sales revenue forecasts can be obtained by transforming the equation back into its original form by taking antilogs: St = (antilog 4.57) x ( antilog 0.336)t St = $96.38(1.4)t Note that 1.4 in the parenthesis is (1+ g). Therefore, g = .4 or 40%

Forecasts of Sales Revenue for Microsoft Corp. (2002-2005)

Year 2002 2003 2004 2005 Time Period (t) 19 20 21 22 Constant Growth Trend Line St = $96.38(1.4)t St = $96.38(1.4)t St = $96.38(1.4)t St = $96.38(1.4)t Fitted Sales (Constant Growth) 57,596 80,639 112,896 158,054

Fitted Sales (Linear)

20,320 21,729 23,138 24,547
43

Managerial Economics

DR. SAVVAS C SAVVIDES

Seasonal Variation
Often the quantity demanded of (sales) of certain goods does not increase uniformly during the year, but exhibit seasonal patterns. For instance, the demand for ice cream, soft drinks, bottled water, beer, electricity, hotel beds, etc. The sale of all these are higher over the summer months and lower in winter.

Demand for (Consumption of) Electricity in Cyprus (1998.1 2001.4)

Year / Quarter 1998.1 1998.2 1998.3 1998.4 1999.1 1999.2 1999.3 1999.4 Time Period (t) 1 2 3 4 5 6 7 8 Quantity (000 kwh) 2330 2930 2885 2640 2470 3245 3050 2800 Year / Quarter 2000.1 2000.2 2000.3 2000.4 2001.1 2001.2 2001.3 2001.4 Time Period (t) 9 10 11 12 13 14 15 16 Quantity (000 kwh) 2660 3495 3250 2850 2775 3640 3350 2950
44

Managerial Economics

DR. SAVVAS C SAVVIDES

Seasonal Variation
Regressing electricity with linear time trend (t: 1 to16), we get: St = 2605.5 + 41.412t
S17 = 3309.5 S18 = 3350.9

R2 = 0.294

(t-ratio of t =2.42)
S20 = 3433.7

Forecasting into the following four quarters (2002) we get:

S19 = 3392.3 We see that the forecasts take into consideration only the long term trend factor in the data. Yet, the data show strong seasonal variation. One way to improve the forecasts is to adjust them with the ratio-to-trend method:

Managerial Economics

DR. SAVVAS C SAVVIDES

45

Seasonal Variation--Example
Ratio-to-Trend Method : Seasonality Ratio = Actual / Trend Forecast Seasonal Adjustment = Average of Ratios for Each Seasonal Period Seasonally Adjusted Forecast = Trend Forecast * Seasonal Adjustment Calculation for Quarter 17 (2002.1)
Year
1998.1 1999.1 2000.1 2001.1

Time Period (t)

1 5 9 13

Trend Forecast St = 2605.5 + 41.412t

2646.9 2812.6 2978.2 3143.9

Actual Sales
2330 2470 2660 2775

Ratio
0.8803 0.8782 0.8932 0.8827

Seasonal Adjustment Seasonally Adjusted Forecast for: 2002.1 2002.2 2002.3 2002.4 (t =17) = (3309.5)(0.884) = 2924.3 (t =18) = (3350.9)(1.132) = 3792.5 (t =19) = (3392.3)(1.053) = 3570.6 (t =20) = (3433.7)(0.932) = 3199.7 (vs. 3309.5 unadjusted) (vs. 3350.9 unadjusted) (vs. 3392.3 unadjusted) (vs. 3433.7 unadjusted)

0.8836

Managerial Economics

DR. SAVVAS C SAVVIDES

46

Seasonal VariationDummy Variables

Similar results to the ratio-to-trend method can be obtained by using seasonal dummy variables. Regressing electricity with dummy variables, we get the following results:

Managerial Economics

DR. SAVVAS C SAVVIDES

47

Seasonal VariationDummy Variables

KWH = 2405 129.75D1 + 598.5 D2 +364.25D3 +40.5 t (-2.38) (11.11) (6.83) (9.63) R2 =0.969 We can now use these regression results to forecast electricity consumption for each of the four quarters in 2002, as follows:
Adjusted forecasts using dummy variables: 2002.1 2002.2 2002.3 2002.4 (t =17) = 2405 129.75 +40.5 (17) = 2963.75 (t =18) = 2405 + 598.50 +40.5 (18) = 3732.5 (t =19) = 2405 + 364.25 +40.5 (19) = 3539.75 (t =20) = 2405 +40.5 (20) = 3210.0 Ratio-to-Trend 2924.3 3792.5 3570.6 3199.7 Unadjusted 3309.5 3350.9 3392.3 3433.7

Managerial Economics

DR. SAVVAS C SAVVIDES

48

Smoothing Techniques (1)

Moving Average Forecast This is another nave forecasting technique. Useful when there are no regularities in the data series (random variation). The moving average forecast is the average of data from w periods prior to the forecast data point.
For example, the twelve-month moving average forecast for sales of a product for March 2004 is the average of sales for the previous twelve months (March 2003February 2004). The greater the number of periods in the moving average, the greater is the smoothing effect since each new observation has a small weight. This is useful for erratic/random time-series data.

Ft =
Managerial Economics

i =1

Ati w
49

DR. SAVVAS C SAVVIDES

Smoothing Techniques (2)

Exponential smoothing Technique: The forecast for period t+1 (that is Ft+1) is the weighted average of the actual value from the prior period and the forecast (Ft ).

Ft +1 = wAt + (1 w) Ft

0 w 1

Root Mean Square Error : It measures the Accuracy of a Forecasting Method. The lower the RMSE, the better the forecast.
RM SE =
Managerial Economics

( At Ft ) 2 n
50

DR. SAVVAS C SAVVIDES

Exponential Smoothing

Managerial Economics

DR. SAVVAS C SAVVIDES

51

Barometric Forecasting Methods

Macro-Economic Indicators:
they show the state of the economy and the direction of short-term changes or turning points in business cycles. Such series of economic indicators are produced regularly by the NBER and the Conference Board.

Leading Indicators: show where general business activity is heading (examples: housing starts, stock market
index, money supply, index of consumer expectations, etc)

s in labor costs, interest rate, RPI for services, average duration of unemployment, etc)

Lagging Indicators: follow (lag) economic activity (e.g., Coincident Indicators: move in line (coincide with)

economic activity (e.g., disposable income, industrial production)

Composite Index (e.g of 10 leading indicators) Diffusion Index (breadth ( %) of them moving up or down)
Managerial Economics DR. SAVVAS C SAVVIDES 52

Econometric Models
Single Equation Model of the Demand For Good X
QX = a0 + a1PX + a2Y + a3N + a4PS + a5PC + a6A + e QX = Quantity of X PX = Price of Good X Y = Consumer Income N = Size of Population
Managerial Economics

PS = Price of Substitute PC = Price of Complement A = Advertising Budget e = Random Error

53

DR. SAVVAS C SAVVIDES

Econometric MethodsExample
Suppose that we estimate the sales of economy class tickets (in thousands) between New York and London and get the following results (variables are transformed in natural logs): ln St = 2.737 - 1.247 ln Pt + 1.905 Yt (- 3.73) (10.8) R2 = 0.975 Assume further that an airline company gets estimates for next year that Y (USA income) = $1480 (billion) and P (prices) = $550. Ln(550) = 6.31 and ln(1480) = 7.3 Substituting in the regressed equation above yields: ln St+1 = 2.737 - 1.247 (6.31) + 1.905 (7.3) ln St+1 = 8.7755 Antilog of 8.7755 is 6470. Therefore, sales of airline tickets next year will be 6.47 million. Since the variables are in logs, the estimated slope coefficients are elasticities. Price elasticity = 1.247 and Income elasticity = 1.905
Managerial Economics DR. SAVVAS C SAVVIDES 54

Econometric Models
Multiple Equation Model of GNP
Ct = a1 + b1GNPt + u1t I t = a2 + b2 t 1 + u2t GNPt Ct + I t + Gt

GNPt =
Managerial Economics

Reduced Form Equation Gt a1 + a2 b2 t 1

1 b1 +
1 1 b1 -

b1 +

1 b1
55

DR. SAVVAS C SAVVIDES