Chapter 2: Introduction to the Simple OLS Regression

A. Simple OLS Regression Coefficients

Regressions are used to determine relationships between a dependent variable and
one or more independent or explanatory variables. A simple regression is concerned with
the relationship between a dependent variable and a single independent variable; a
multiple regression is concerned with the relationship between a dependent variable and a
series of independent variables. A linear regression is used to describe the relationship
between the dependent and independent variable(s) as a linear function or line (or
hyperplane in the case of a multiple regression).
The simple Ordinary Least Squares regression (simple OLS) takes the following
form with n observations i measured by dependent variable yi and independent variable
xi, and two parameters describing a line, a vertical intercept term b0 and slope term b1 and
an error term i that represents the error or difference between the actual observation
value yi and the value i predicted by the line:

(1) yi = b0 + b1xi +i

Expressed with matrix notation (the x vector is augmented with a vector of ones to
accommodate the vertical intercept b0), the simple OLS with n observations and a single
independent variable takes the following form:

(1.a)

y = X b +

The ordinary least squares regression coefficients b0 and b1 are derived by minimizing the
variance of errors in fitting the curve (or m dimensional surface for multiple regressions
involving m variables, as we shall see later). Since the expected value of error terms
equals zero, this derivation is identical to minimizing error terms squared. Regression
coefficient b1, as we will derive later, is simply the covariance between y and x divided
by the variance of x; b1 and b0 are found as follows:

(x i x )( yi y )
(2) b1 x ,2y i 1
x n

(x x)
i 1
i
2

(3) b0 y b1x

Using matrix notation, parameters b0 and b1 are found as follows:

1
(2.a)

Appropriate use of the OLS requires the following assumptions:

1. The variance of x is approximately the same over all ranges for x.

2. The variance of error term values is approximately the same over all ranges of
x.
3. The expected value of each disturbance or error term equals zero.

Violations in these assumptions will weaken the validity of the results obtained from the
regression and may necessitate either modifications to the OLS regression or different
statistical testing techniques. Section C of this chapter provides discussion on how the
regression coefficient formulas are obtained.
There are numerous types of regressions depicting different types of relationships
among variables. Table 1 provides details on some of these different types of regressions.
A simple regression is concerned with the relationship between a dependent variable and
a single independent variable. Regression coefficients b0 and b1 represent the vertical
intercept and the slope in the statistical linear relationship between the dependent variable
yi and the independent variable xi. Thus the vertical intercept b0 represents the
regression's forecasted value for yi when xi equals zero and the slope of the regression b1
represents the change in i (the value forecast by the regression for yi) induced by a
change in xi. The error term i represents the vertical distance between the value i
forecasted by the regression based on its true value yi; that is, i = yi i. The OLS
regression minimizes the sum or average of these error terms squared. The size of the
sum of the squared errors (often called SSE or, when divided by (n-2), the variance of
errors 2 ) will be used to measure the predictive strength of the regression equation. A
regression with smaller error terms or smaller 2 is likely to be a better predictor, all
else held constant.

Table 1:
Classes of Ordinary Least Squares Regressions

By Number By Shape of Curve

of Variables Example
Simple Linear Yi b0 b1 xi i
Multiple Linear Yi b0 b1 xi b2 zi i
Simple Curvi-linear* Yi a xib i
Simple Log-linear* log( Yi ) b0 b1 log( xi ) i
Multiple Curvi-linear* Yi a1 xib (1) a2 zib( 2) i
Multiple Log-linear* log( Yi ) a1 b1 log( xi ) b2 log( zi ) i

2
 Simple Non-linear** Yi a0 a1 xib i
Multiple Non-linear** Yi a0 a1 xib(1) a2 zib( 2) i
* Curvi-linear regressions can be transformed into linear regressions. In these examples, the
transformation is completed by finding the log of both sides, while ignoring the error term, since its
expected value is zero. The simple log-linear regression is a transformation of a simple curvi-linear
regression.
** Non-linear regressions cannot be transformed into linear regressions.

Illustration 1: Calculating a Stock Beta

Historical time series returns for Holmes Company stock, along with those of the
market portfolio and Treasury Bill (T-Bill) rates rf are summarized in Table 2. This table
also computes risk premiums on Holmes Company stock and the market portfolio.
We can determine the relationships between risk premiums of the Holmes
Company shares and risk premiums on the market portfolio by the use of an ordinary
least squares regression (OLS). That is, if we were to regress risk premiums of the stock
(rH,t - rf) against risk premiums of the market (rm,t - rf), we would be able to obtain a
Capital Asset Pricing Model Beta for the Holmes Company stock. In this scenario, risk
premiums on Holmes Company stock (rH,t - rf) are the dependent variable yt and the
independent variable xt is market risk premiums (rm,t - rf). Our regression equation is
given as follows:

(rH,t - rf,t) = H + H(rm,t - rf,t) + H,t

where H designates the vertical intercept for the regression and bH is the regression
slope, or the stock beta for Holmes. Each data point is depicted in Figure 1. The slope of
the line is bH and its vertical intercept is H. The beta from the Capital Asset Pricing
Model is frequently used as a measure of a stock's risk relative to the risk of the market
portfolio.

Table 2:
Holmes and Market Returns

Year Holmes rH-rf Market rm-rf T-Bill

Co.
2006 12% .06 10% .04 6%
2007 18% .12 14% .08 6%
2008 7% .01 6% 0 6%
2009 3% -.03 2% -.04 6%
2010 10% .04 8% .02 6%

3
Figure 1:
Regression Line for Stock and Market Returns
0.14

0.12

0.1

0.08

0.06

0.04

0.02

-1E-17
-.04 .02 .04 .08
-0.02

-0.04

We compute the stock beta through the following by first calculating the variance
of risk premiums for market portfolio:

(.04 .02)2 (.08 .02)2 (0 .02)2 (.04 .02)2 (.02 .02)2

m2
5 1
=.008/4=.002

Next, we calculate the covariance between risk premiums on the Holmes Company stock
and risk premiums on the market portfolio:

m
( RH ,t E ( RH ,t )) ( Rm,t E ( Rm,t ))
H ,m
t 1 n 1

= {[(.06 -.04) (.04 -.02) (.12 -.04) (.08 -.02) + (.01 -.04) (0 -.02)
+ (-.03-.04) (-.04 -.02)+ (.04 -.04) (.02 -.02)] /(5-1)

(.02) (.02) (.08) (.06) (.03) (.02) (.07) (.06) 0

= 0.0025
4

We obtain the beta for the stock by dividing the covariance between the stock and market
risk premiums by the variance of premiums on the market portfolio:

H , m .0025
H 1.25
m2 .002

4
 This beta implies that one might expect that a 1% increase in the market's return
will lead to a 1.25% increase in the return of Holmes Company stock. However, this
interpretation does omit consideration of nonmarket influences or firm-specific
influences on its return. In any case, Holmes Company stock is regarded as having 1.25
times the level of systematic risk of the market or average security. The vertical intercept
H for the Holmes Company will equal .015. Normally, if we believe that the Capital
Asset Pricing Model holds, we believe that = 0 for all securities. If a does not equal
zero, we might conclude either that Capital Asset Pricing Model did not hold for the
testing period, or that our measuring or sampling techniques were insufficient or
inappropriate (note that we only have five data points in this example) or that the stock
consistently outperformed the market for the sampling period.

Illustration 2: The Jensen Alpha for Portfolio Performance Evaluation

Proper benchmarks for comparison must be established when evaluating portfolio
returns. Ideally, a benchmark should make allowances for portfolio risk. Here, we will
use the Jensen alpha measure to evaluate the performance of a portfolio:

[Rp - rf] = p + [p(rm-rf)]

Presumably, a positive alpha () which is statistically significant will indicated that

portfolio outperforms the market on a risk-adjusted basis. Consider Table 3, which
records returns over a twenty year period for a portfolio (p) and the market along with
riskless return rates:

Table 3:
Portfolio p and Market Performance

Year Rp Rm rf Rp-rf Rm-rf Year Rp Rm rf Rp-rf Rm-rf

1991 0.14 0.05 0.03 0.11 0.02 2001 0.11 0.07 0.05 0.06 0.02
1992 0.11 0.03 0.02 0.09 0.01 2002 0.22 0.17 0.06 0.16 0.11
1993 0.04 0.01 0.02 0.02 0.03 2003 0.22 0.16 0.07 0.15 0.09
1994 0.16 0.11 0.03 0.13 0.08 2004 -0.01 -0.05 0.06 -0.07 -0.11
1995 0.03 -0.12 0.02 0.01 -0.14 2005 0.04 -0.08 0.05 -0.01 -0.13
1996 0.14 0.09 0.03 0.11 0.06 2006 0.28 0.21 0.06 0.22 0.15
1997 0.26 0.13 0.04 0.22 0.09 2007 0.22 0.11 0.07 0.15 0.04
1998 0.26 0.18 0.05 0.21 0.13 2008 0.21 0.11 0.08 0.13 0.03
1999 0.13 0.04 0.05 0.08 -0.01 2009 -0.04 -0.11 0.07 -0.11 -0.18
2000 0.08 -0.11 0.04 -0.12 -0.15 2010 0.18 0.12 0.05 0.13 0.07

To compute the Jensen alpha measure, we first convert returns to risk premiums by
subtracting riskless rates from returns on the portfolio and the market. We then run a
regression of portfolio risk premiums against market risk premiums to obtain the
following:

[Rp - rf] = p + [p(rm-rf)] = .073509 + .951512(rm-rf)

5
 First, we have computed the portfolio beta to be p = .951512. The systematic risk
of this portfolio is slightly less than that of the market. More importantly, we have
computed Jensen's alpha, which is .0735 and exceeds zero. This suggests that this
portfolio outperformed the market on a risk-adjusted basis over the twenty year-period.
These results are depicted in Figure 2.

Figure 2: Jensens Alpha Scatter Diagram

B. Simple OLS Regression Inference

Relationship Strength
Once we have determined the statistical relationship between yi and xi based on
our OLS, our next problem is to measure the strength of the relationship, or its
significance. One of the more useful indicators of the strength of the regression is the
coefficient of determination or 2 statistic. The coefficient of determination (often
referred to as r-square) represents the proportion of variation of variable y that is
explained by its regression on x. It is determined as follows:
n
[ ( xi x )( yi y )]2
( x , y ) 2
(4) x2, y i 1
x2 y2 n n

( xi x )2 ( yi y )2
i 1 i 1

This coefficient of determination might also be expressed as either of the

following:

6
 Total Variation in y Explained by theRegression
(5) x2, y
Total Variation in y

n n

(6) ( yi y )2 i2
x2, y i 1
n
i 1

( y
i 1
i y)2

The sum (yi -y)2 represents total variation in y; the sum 2 represents the variation in
y not explained by the regression on x.
Assume that there exists for a population a true OLS regression equation yi =0
+1xi + i representing the relationship between yi and xi, without measurement or
sampling error. However, we propose the regression yi = b0 + b1xi + i, whose ability to
represent the true relationship between yi and xi is a function of our ability to measure
and sample properly. Our sampling coefficients b0 and b1 are merely estimates for the
true coefficients 0 and 1 and they may vary from sample to sample. It is useful to know
the significance of each of these sampling coefficients in explaining the relationship
between yi and xi.

Inference from a Regression

Our estimate 1 for the slope coefficient b1 may vary from regression to
regression, depending on how our sample varies. Our estimates for b1 will follow a t-
distribution if our sample of yi's is large or normally distributed; if our sample is
sufficiently large, our estimates for b1 may be characterized as normally distributed. One
potential test of the significance of our coefficient estimate b1 is structured as follows:

H0: b1 0
HA: b1 > 0

Our null hypothesis is that y is unrelated or inversely related to x; our alternative

hypothesis is that y is directly related to x. The first step in our test is to compute the
standard error se(b1) of our estimate for b1 as follows:1

i 1
i
2

1 2 n2
(7) se(b1 )
n x2 n

(x x)
i 1
i
2

1
Standard errors for b0 and b1 are obtained by premultiplying the 2 X 2 matrix by 1/(n-m-1)eTe.
The square roots of elements in the principal diagonal of this 2 X 2 matrix are the standard errors for b0 and
b1. We will discuss this computation further in Chapter 3.

7
 The standard error for b1 is, in a sense, an indicator of our level of uncertainty
regarding our estimate for b1. The numerator within the radical indicates the variability in
x not explained by the regression; the denominator indicates the total variability of the
independent variable. Our next step is to find the test statistic for b1, which will be related
to both the standard error for b1 and the b1 estimate itself. This is analogous to
standardizing or finding the normal deviate in our earlier hypothesis tests:

b1
(8) t (b1 )
se(b1 )

We next compare this test statistic to a critical value from a table representing the t-
distribution such as Table 2.A.2 in the appendix to this chapter.
The process for determining the statistical significance of the vertical intercept 0
is quite similar to that for determining the statistical significance for 1. We first
designate appropriate hypotheses, such as those that follow:

H0: b0 = 0
HA: b0 0

The primary difference in the process is in determining se(b0):

i
2
n
i 1
xi2
n2
(9) se(b0 ) n
i 1

n ( xi x ) 2
i 1

Next, we find our t-statistic as follows:

b0
(10) t (b0 )
se(b0 )

We then compare our t-statistic to the appropriate critical value just as we did when
testing the significance of the slope coefficient. This particular test involves two tails,
since our alternative hypothesis is a strict inequality. Be certain to make appropriate
adjustments to the critical value (for example, divide the critical value by 2 for two tailed
tests) when making comparisons.

Illustration 1: Calculating a Stock Beta, Continued

Suppose that we merely wish to test the hypothesis that our beta is greater than
zero. Our null hypothesis and alternative hypotheses are given as follows:

H0: H 0
HA: H > 0

8
Assume we wish to test at a 95% level of confidence; that is, we wish to conclude that
there is at least a 95% probability that we are correct if we reject our null hypothesis.
Since we are working with a sample set of 5 data points, we will perform a one-tailed test
of our null hypothesis with 3 (5-2) degrees of freedom assuming that H follows a
student-t distribution.
Using the data for the Holmes Company, determine the statistical significance of
the Holmes Company stock beta. We need to compute a standard error so that we may
compute a test statistic. Our first step is to compute 2 = SSE and divide by its number
of degrees of freedom, n-2, as illustrated by Table 4. Now that we have computed the
standard error for the estimate of H = .064549, we can compute the test statistic for H
using:
H 1.25
t ( H ) 19.3649
se( H ) .064549

Table 4
Holmes Stock Beta Significance

Year rf rH-rf rm-rf H(rm-rf) +H(rm-rf) t

2

2006 .015 .06 .04 .050 .065 -.005 .000025

2007 .015 .12 .08 .100 .115 .005 .000025
2008 .015 .01 0 0 .015 -.005 .000025
2009 .015 -.03 -.04 -.050 -.035 .005 .000025
2010 .015 .04 .02 .025 .04 0 0
SSE = .000100

n 5

SSE t 1 i2
.0001 i2
Average Squared Error = = = t 1 = .000033333
n2 n2 52 3
5

(r m,t rf )
(rm rf ) = i 1
.02
5
5

[(r
i 1
m,t rf ) (rm rf )]2 = .08

i 1
i
2

1 2 n2 .000033333
se( H ) .004166 .064549
n m2 n

[(r
.08
m,t rf ) (rm rf )]2
i 1

9
Since we can assume that H follows a student t distribution, may compare our t-statistic
to a critical value found in the t-distribution table, given the appropriate level of
confidence and degrees of freedom. In this case, the significance of our test is = 1 - .95
= .05 and the number of degrees of freedom k equals n - 2 = 3. Our critical value for the
test is found:

= 2.353

Since 19.36492 > 2.353, we reject the null hypothesis in favor of the alternative
hypothesis. Therefore, it appears that market risk premiums are significant at the 95%
level in explaining risk premiums of Holmes Company stock.
The process for determining the statistical significance of the vertical intercept H
is quite similar to that for determining the statistical significance for H. Again, our null
and alternative hypotheses are:

H0: b0 = 0
HA: b0 0

and the primary difference in the test is in determining se():

i
2
n
i 1
xi2
n2 .00003333 .01
se( ) i 1
.002886
n
5 .008
n ( xi x ) 2
i 1

Suppose we are interested in testing whether the risk premium on Holmes (rH-rf)
tends to be consistently different from the risk premium of the market (rm-rf). In effect,
we are testing whether H 0. More formally, we are interested in testing the following at
the 95% level of confidence:

H0: H = 0
HA: H 0

Given our H value of .015 and our standard error of .002886, we find our t-statistic
using:
H .015
t ( H ) 5.19615
se( H ) .002886

Next, we find our critical value in this 2-tail test at t = t.025. Note that our significance
level is divided by 2 since we are performing a two tailed test. Our critical value is 3.182.
Since our t statistic exceeds this critical value, we reject our null hypothesis; we cannot
conclude that H = 0. Therefore, based on our data, we must reject the applicability of the
Capital Asset Pricing Model. However, we should stress at this point that our data set is
far too small to reach meaningful conclusions about the Capital Asset Pricing Model.

10
 Finally, we can determine the coefficient of determination (r-square) from
Equations 4 or 5 above. Based on the coefficient of determination of .992 determined
from Equation 4, we conclude that 99.2% of all of the variation in risk premiums for
Holmes Company stock can be explained by risk premiums on the market portfolio. A
typical reporting format with t-statistics in parentheses for an OLS regression such as that
we have just completed follows:

yi = .015 + 1.25 xi r-squared = .992

(5.196) (19.365) d.f. = 3

Illustration 2: The Jensen Alpha for Portfolio Performance Evaluation, Continued

The following represent the regression results and diagnostics from the portfolio
performance evaluation illustration from the previous section:

[Rp - rf] = p + [p(rm-rf)] = .073509 + .951512(rm-rf) r-squared = .865127

(8.505677) (10.7452) d.f. = 18

The t-statistics are reported in parentheses below the regression coefficients. First, we can
conclude that the portfolio beta (.951512) is statistically significant at the 1% level. The
portfolio has slightly less systemic (market-related) risk than the market itself. The r-
squared value suggests that slightly more than 86% of this funds risk premiums are
driven by market risk premiums. But most importantly, we can conclude that Jensen's
alpha is statistically significant at the 1% level. Thus, the fund outperformed the market
on a risk-adjusted basis for the reported period. Although we have concluded statistical
significance for alpha, we should question the relevance of data as old as twenty years to
evaluation of current fund management.

Illustration 3: Testing Momentum versus Mean Reversion

In this illustration, we will provide an illustration of a momentum/mean reversion
test. Essentially, this means that we will test whether the return for a stock on a given day
can be used to predict the return on that stock for the subsequent day. If returns seem to
be positively correlated from one day to the next, this would evidence momentum in
stock prices over time. That is, high returns in one day suggest high returns the following
day; low returns suggest low returns the following day. A negative correlation would
evidence mean reversion. Mean reverting returns tend to revert back to their mean value
over time.
Again, suppose an analyst was interested in whether there existed a relationship
between the daily return on a security and its return on a prior day. For example, consider
the following sequence of daily closing prices in Table 5 for a given stock Q, from which
we compute returns:

11
 Date t Pricet Returnt Returnt-1
02/01/10 1 49
02/02/10 2 50 .020408

02/03/10 3 51 .020000 .020408

02/04/10 4 52 .019607 .020000

02/05/10 5 55 .057692 .019607

02/06/10 6 57 .036363 .057692

02/07/10 7 58 .017543 .036363

02/08/10 8 59 .017241 .017543

02/09/10 9 58 -.016949 .017241

02/10/10 10 55 -.051724 -.016949

02/11/10 11 53 -.036363 -.051724

02/12/10 12 52 -.018867 -.036363

Table 5: Stock Price and Return Data

The prices given above assume that the stock traded each day, including weekends. We
will attempt our test illustration based on an extremely small sample to simplify
computations. Although there are many ways to determine the nature of the relationship
between the return on a security and its prior day's return, we will examine whether there
exists a linear relationship based on a simple ordinary least squares regression of the
form: rt = a + brt-1. We report regression results as follows:

r .0020142 .7717758 rt 1 r-square = .57

(-.25959) (3.259325) SSE = .00450311
n = 10 d.f. = 8

Note that t-statistics are given in parentheses. We note here that the slope term b is
.7717758, positive and statistically significant at the 5% level. We will return shortly to
explain and interpret these results. However, we will first work through the details on
obtaining the results of our simple Ordinary Least Squares (OLS) regression. First, our a
and b coefficients are determined following the simple OLS method as follows:

(1) b (r r )(r
t t t 1 rt 1 )

.00597425
.595638 .7717758
(r rt 1 t 1 )
2
.01003

(2) a rt brt 1 .0020142

To obtain regression diagnostics, we compute error terms squared as follows in Table 6:

12
 t Returnt Returnt-1 a + brt-1 i i2

3 .020000 .020408 .013735 .00626 .000039239

4 .019607 .020000 .013420 .00618 .000038278

5 .057692 .019607 .013118 .04457 .001986848

6 .036363 .057692 .042510 -.00614 .000037788

7 .017543 .036363 .026049 -.00850 .000072353

8 .017241 .017543 .011525 .00571 .000032673

9 -.016949 .017241 .011291 -.02824 .000797554

10 -.051724 -.016949 -.015095 -.03662 .001341651

11 -.036363 -.051724 -.041934 .00557 .000031029

12 -.018867 -.036363 -.030079 .01121 .000125692

SSE = .004503110

Table 6: Predicted Stock Return Data

Now we are able to compute regression standard errors:

SSE
se(b) n2
.000562888
.056069 .23679
(rt rt ) 2
.01003
n

i
2
n
i 1
xi2
n2 .000562888 .0107417
se(a) i 1
.000060228 .007760684
n
10 .01003911
n ( xi x ) 2
i 1

Given our standard error estimates se(a) = .007760684 and se(b) = .05606069533, we
find our t-statistics to be t(a) = -.25959 and t(b) to be 3.2593. Suppose our tests were
expressed formally as follows:

H0: a = 0 H0: b = 0
HA: a 0 HA: b 0

Since we will assume that a and b follow student t distributions, may compare our t-
statistic to critical values found in the t-distribution table, given the appropriate level of
confidence and degrees of freedom. We are performing two tail tests for significance.
Suppose we wish to test for significance at a level of .95 (.025 on each side of the tail). In
this case, the significance of our test is = .025 and the number of degrees of freedom k
equals 10 - 2 = 8. Our critical value for the test is found:

t10-2 = t.025 = 2.306.

13
While we fail to reject our first null hypothesis that a = 0, we reject our second null
hypothesis that b = 0. Thus, our test does provide evidence that the returns for this
security are linearly related to returns on prior days.

C. Deriving the Simple OLS Regression Equation

A simple ordinary least squares regression is used to fit a line to a series of data
points. In this section, we refer to 0 as the fitted value for the vertical intercept of the
regression line; 1 is the fitted value for the slope of this line. One attempts to maintain
small vertical distances (errors: i) of predicted dependent variable values ( i) on the line
from actual data point values (yi). The following equations represent the line fit by the
regression and the actual or empirical values of the dependent variable:

(10) Predicted yi value

i Empirical or actual yi value

Assume that each of the historical data points expressed in terms of y and x are
known. We define our line in terms of regression coefficients and . Our objective
when obtaining regression coefficients is to minimize the sum of error terms squared.
First, we solve Equation (10) for i, which is then squared and summed to obtain:

(11)

We rewrite Equation (11) as follows:

(12)

Note that if we were to attempt to minimize the sum of error terms themselves, we
would find that our forecasted values i equal infinity for all i; this causes each of our
error terms to equal minus infinity. We minimize squared error terms so that we do not fit
a line such that all of the error terms equal minus infinity; we want error terms to be close
to zero. Squaring the error terms results in positive values which we will want to
minimize. Also, we show later that the normal distribution is defined in terms of squared
variable values, adding to the attractiveness of working with least squares. Minimizing
the squared error terms is analogous to fitting the line as close as possible to the data
points. To minimize the sum of error terms squared, we find partial derivatives with
respect to and and set them equal to zero:2

(13.a)

2
Readers lacking backgrounds in calculus should review Appendix B to this chapter to understand how
derivatives are useful for finding minimum values for polynomial functions.

14
(13.b)

Next, we cancel 2's to simplify and solve equation (13.a) for yi and equation
(13.b) for xiyi to obtain our set of normal equations:

(14)
+

To finally obtain our regression coefficients, we solve the above normal equations
simultaneously for and . We can show algebraically that solving this system is
identical to solving the following for regression coefficients and :3

( x x )( y
i i y)
( x , y ) 2
(15) b1 i 1

n
x2
( xi x )
i 1
2

(16) b0 y bx

Similarly, in matrix format, Equation Set 14 is rewritten as Equation 17, which solves for
b as 2.a:

(17)

(2.a)

D. OLS Assumptions
The simple ordinary least squares regression (OLS) provides the best fit line
defined by slope and vertical intercept that minimizes squared errors. This means that the
OLS will fit a line to data points that minimizes differences between predicted values for
the dependent variable and actual values for the dependent variable. However, in order
for the OLS to provide the best fit line and to be useful for predicting out of sample
values for the dependent variable, the following assumptions should hold:

1. The expected value for each observation of the dependent variable y is a linear
function of a given value of the independent variable x (this implies linearity yi =

3
We divide both sides of the first part of Equation 14 by n and solve for to obtain = .
Substitute this result into the second part of Equation 14 and solve for ;
)=( - ), which is rewritten as Equation 15.

15
 b0 + b1xi and exogeneity where the errors have mean zero: E[] = 0, and that the
regressors and parameters are uncorrelated with the errors: E[x] = 0,
2. For each value of x, the value of y is normally distributed (while convenient,
normality is not crucial, and can be replaced with identically and independently
distributed, or i.i.d.) about its expected value, with a probability or frequency
distribution that has the same variance 2 for each observation (This means that
the error terms exhibit homoskedasticity; that is, E[i2|X] = 2.)
3. The sample values of y are all uncorrelated with one another or are statistically
independent, and
4. Variable x must take at least two different values; that is, the sample set must
exceed one.

These assumptions (aside from normality) underlie the Gauss-Markov Theorem,

which holds that linear regression estimators b0 and b1 have the smallest variance of all
linear and unbiased estimators of the true 0 and 1. These estimators are the Best Linear
Unbiased Estimators (B.L.U.E.) of 0 and 1.

E. Functional Form and Transformations

To this point, we have focused on linear relationships between independent and
dependent variables. However, accounting for nonlinearities might improve the
descriptions of these relationships. One such scenario might be where change in the
independent variable better describes the change in the dependent variable than the
independent variable itself:

(1) ln(yi) = b0 + b1xi +i

Anti-logs on both sides of (1) leads to:

(2) yi = e(b0+ b1xi) if i = 0

Differentiating y with respect to x leads to:

(3)

Substituting y for leads to:

(4)

Assuming small changes in x and y (sufficiently small ), dividing both sides

by y, and multiplying both sides by leads to:

(5)

(6)

16
 Suppose, for example, that for a particular pharmaceutical company, a given
increase in the number of drugs approved by the FDA leads to some constant times that
increase in the percentage change in firm equity value (% returns). Thus, according to
Equation (6), , or %r =
. Equation (5) implies that Equation (2) describes the relationship
between firm equity value y and the number of FDA approvals x. We would prefer not to
use the OLS regression technique on Equation (2) to test the relationship between y and x
because this relationship is clearly not linear. However, it is easy to take natural logs of
both sides of (2) to obtain Equation (1), which states that the natural log of y is linearly
related to x. Thus, for the pharmaceutical firm, the log of a return would be linearly
related to the number of FDA approvals as follows:

(7) ln(ri) = b0 + b1(# FDA approvals)i +i

Actually, because equity returns can be negative, but never less than -1, or -100%, it is
practical to add 1 to the return as the dependent variable:

(7) ln(1+ri) = b0 + b1(# FDA approvals)i +i

In addition, log transforms are useful to obtain meaningful measures or

comparisons of mean values when observations are skewed. For example, when
frequency distributions for an independent variable have a single fat tail or are skewed,
log transforms of data are more likely to be approximately normally distributed. These
log transforms of independent variables are more likely to provide for meaningful results
in an OLS regression. This situation can hold even where there are a very small number
of extreme observations on one side of the mean. Thus, for example, with many cross
sectional samplings of publically traded corporations, there are likely to be a very small
observations of very large firms (e.g., over $100 million capitalization), and the largest of
this small number of firms might be several times larger than the smallest in this sub-
sample. However, there may be dozens or even hundreds of firms in the $1 to $10 million
range. Thus, the left tail of the frequency distribution for firm sales observations is likely
to be much fatter than the right tail. Transforming these observations to logs of their
actual values is likely to lead to more symmetry in the distribution, and more likely to
yield a frequency distribution closer to that of the normal distribution.

Illustration 4: Slippage and the Perfect Foresight Half Spread

Brokers have a responsibility to provide their clients best execution. Best
execution is often estimated as the sum of order processing costs and market impact or
slippage. Measuring order processing costs is normally fairly straightforward. These costs
will normally include brokerage fees, which might account for other processing costs as
brokers normally provide these services as well. However, market impact or slippage
costs measurement is more problematic. The price or market impact of a transaction, the
effect that a given transaction has on the market price of a security, is particularly
important for a large transaction or in an illiquid market. When institutions execute large
transactions, their buy or sell activities affect security prices, typically in a manner that
increases execution costs. Large buy orders will tend to force the security price up against

17
the buyer and large sell orders will typically force the selling price down against the
seller. While substantial, especially for large orders, these slippage costs are not explicit,
and might be considered to be hidden. However, we generally expect that a single large
transaction will cause more slippage than several small orders totaling to the same
number of shares (e.g., a single 10,000 share purchase causes more slippage than ten
1,000 share purchases). One measure of market impact or slippage is the perfect foresight
half spread, the absolute value of the difference between some future price and the trade
price. Here, the assumption is that the subsequent or future price is the fundamental value
of the security and the prior price reflects the price impact of the transaction.

Perfect
Subsequent Foresight
Purchase Transaction Half- Transaction
Purchase Price Price Spread Size
1 50.00 50.30 0.30 1000
2 50.26 50.30 0.04 200
3 50.26 50.35 0.09 400
4 50.17 50.33 0.16 600
5 50.32 50.33 0.01 100
6 50.41 50.46 0.05 300
7 50.51 50.55 0.04 200
8 50.44 50.47 0.03 300
9 50.71 51.12 0.41 1200
10 51.03 51.04 0.01 100
11 50.97 50.99 0.02 200
12 50.57 50.75 0.18 700
13 50.71 50.74 0.03 200
14 50.92 50.96 0.04 200
15 50.91 50.93 0.02 200
16 51.18 51.32 0.14 600
17 51.71 51.76 0.05 300
18 52.04 52.12 0.08 400
19 52.03 52.34 0.31 1000
20 52.30 52.34 0.04 300
Table 7: Transaction Size and Slippage

18
 Figure 3: Slippage and Transaction Size Plot

Table 7 depicts a series of 20 historical transactions prices that the brokerage firm
paid for a given stock's purchases. Table 7 also depicts subsequent transactions prices,
from which perfect foresight half-spreads are computed. The sizes of the purchases
follow, so that the broker can analyze the relationship between the perfect foresight half-
spread and the transaction size based on 20 of the broker's own purchase transactions for
the same stock. For example, the first transaction has the broker purchasing 1000 shares
for $50.00. The next transaction, by another investor, was at a price of $50.30, implying a
perfect foresight half spread of .30. Presumably, this transaction by the broker increased
the price of the security by $.30, and this increases the purchase price for any subsequent
purchases that the broker might wish to execute. Purchases by other investors are omitted
in the "Purchase Price" column.
Notice that larger purchase orders tend to lead to higher perfect foresight half-
spreads in Table 7. This suggests that larger orders lead to greater slippage, just as we
would expect. However, a close examination of these half-spreads and transactions will
reveal that the relationship is not linear (Notice the slight concave-up relationship in
Figures 3 and 4). In fact, most brokers would expect that this relationship would not be
linear because the effect of transaction size on the perfect foresight half spread tends to
be concave up. That is, for example, a broker might expect that a single purchase
transaction for 1000 shares will cause more slippage than 10 purchase transactions for
100 shares each. To analyze this non-linear relationship, we will conduct a log-log
regression of the following form:

log(Perfect Foresight Half-Spread)t = logb0 + b1log(Transaction Size) + t

When t = 0, this log-log relationship would imply that the following estimated
relationship between the perfect foresight half-spread and Transaction Size:

Perfect Foresight Half-Spread = b0(Transaction Size)b(1)

Perfect Foresight Half-Spread = v(Transaction Size)m

19
where b0 serves as an estimate for v and b1 serves as an estimate for m. These parameters
b0 and m characterize slippage, which is positive when v is positive and m exceeds 1.
Higher values for b0 and m indicate increased slippage. Table 8 presents log values (base
10) for the perfect foresight half-spreads and transaction sizes, and presents regression
results. The log-log regression implies that b0 = -4.96 and b1 = 1.48; that is, v is
approximately equal to .00001 and m is approximately equal to 1.5 (use anti-logs of b0
and b1 to obtain v and m). Thus, the perfect foresight half-spread or slippage is S = vcm =
.00001c1.5 where c is transactions size. We see in Table 8 that the regression t-statistics,
the F-statistic and the adjusted r-square value all suggest that these results and parameters
are statistically significant.

Figure 4: Slippage and Transaction Size Curve Plot

Log Perfect Log

Foresight Transaction
Half-Spread Size
-0.522879 3.00000 SUMMARY OUTPUT
-1.397940 2.30103
-1.045757 2.60206 Regression Statistics
-0.795880 2.77815 Multiple R 0.977851577
-2.000000 2.00000 R Square 0.956193706
-1.301030 2.47712 Adjusted R Square 0.953760023
-1.397940 2.30103 Standard Error 0.101914919
-1.522879 2.47712 Observations 20
-0.387216 3.07918
-2.000000 2.00000 ANOVA
-1.698970 2.30103 df SS MS F Significance F
-0.744727 2.84510 Regression 1 4.0809 4.0809 392.9 1.12419E-13
-1.522879 2.30103 Residual 18 0.18696 0.010386651
-1.397940 2.30103 Total 19 4.26787
-1.698970 2.30103
-0.853872 2.77815 Coefficients Standard Error t Stat P-value
-1.301030 2.47712 Intercept -4.96055 0.189596533 -26.2 8.9E-16
-1.096910 2.60206 X Variable 1 1.4805 0.074692199 19.82 1.1E-13
-0.508638 3.00000

20
-1.397940 2.47712
Table 8: Slippage and Transaction Size Regression

F. The Single Index Model

Simple observation of security markets reveals a strong tendency for security
returns to be affected by one or more common factors, particularly the market portfolio.
From a mathematical perspective, these factors represent a source of covariance or
correlation between returns of pairs of securities. The single index model specifies a
single source of covariance among security returns Ri,t, and denotes security returns as a
linear function of this factor or index It:

(1) Rit i i I t it

where i represents that portion of the return of security i which is constant and
independent of the index It, i represents the sensitivity of security i to the single index I
and it represents the portion of security i's return which is security specific and unrelated
to the index or to returns of other securities. The index models are simply regression
models that presume that security returns are a linear function of one or more (in the case
of multi-index models) indices. If index models can be used to generate security returns,
then the process for obtaining security variances and covariances with respect to one
another will be much simplified. While returns on the market portfolio (e.g., S&P 500)
would normally serve as the single index I in this framework, in theory, the index could
take on any observable numerical value.
The Single Index Model is particularly useful to portfolio managers, with the
following applications:

1. To reduce the number of inputs and computations required for portfolio analysis.
In particular, the Single Index Model will be useful for deriving forecasts for
security and portfolio expected return, variance and covariance.
2. To build and apply equilibrium models such as the Capital Asset Pricing Model
and Arbitrage Pricing Theory.
3. To adjust for risk in event studies and back-testing programs.

In addition to the standard OLS assumptions that we introduced earlier, the Single
Index Model is based on the following series of assumptions:

1. As indicated above, security returns are linear in a common index as follows:

Rit i i It it

2. The parameters of the index model i and i are computed through a linear
regression procedure such that the risk premium is purely a function of the index,
not security specific risk. That is, E(it) = 0. Furthermore, it will be assumed that
security specific risk is unrelated to the value of the index; that is, E(itIt) = 0 =
Cov(itIt).

21
 3. The index represents the only source of covariance between asset returns. That is,
E(itjt)=0.

Based on the Single Index Model, we might reflect the expected return of a
security i or portfolio p as follows:

(2) E [ Ri ] i i E[ I ]

(2a) E [ R p ] p p E[ I ]

where the parameters for the portfolio are simply a weighted average of the parameters
for the individual securities. For sake of notational convenience, we use the expectations
operator E[*] to replace the summation notation i=1[*]Pi. That is, for expected security
return and variance, we have:

m m
(A) [ Ri ] Ri Pi i i I i Pi
j 1 j 1

(B) E( ( I EI ) 2i ( I EI ) i
i
2
i
2 2
i
2

We can use Equations 1 and 2 and our standard definition for security variance to express
security variance as a function of the index:

(A)
i2 E (i i1I i ) (i i EI ) Ei ( I EI ) i 2
2

Note that vertical intercept terms cancel. We can complete the square of Equation A and
write security variance as:

(B) i2 E(i2 ( I EI )2 i2 2i ( I EI ) i

Because the covariance between the index and firm specific returns is assumed to be zero
above (E[(it-0)(I-E[I])] = 0), the cross product terms drop out:

(C) i2 E(i2 ( I EI )2 i2 i2 E( I EI )2 E( i 0)2

Due to our definition of variance, and that the expected unsystematic risk premium
(error) equals zero, Equation C simplifies to:

(3) i2 i2 I2 2i

This expression has a particularly useful intuition: security variance is the sum of
systematic or index induced variance ii and firm specific variance i. Firm specific
risk i will tend towards zero in a well-diversified portfolio such that portfolio variance
is expressed:

22
(3a) p2 p2 I2

The Single Index Model can be used to substantially reduce the number of
computations for covariances required for portfolio risk analysis. We know that n2
covariance calculations are required to compute portfolio risk.4 For example, a ninety-
security portfolio will require 8100 covariance calculations. Thus, it is very useful to
limit the number of calculations required for each covariance. Using the expectations
operator notation, we can define covariance as follows:

(4) i , j E( Ri ERi )( R j ERJ )

Replacing Equations 1 and 2 into Equation 4, we have:

i , j Cov( Ri , R j )

(A)
E ( i i I i i i EI )( j j I j j j EI )

After performing multiplications within the brackets, and noting that i and j are
uncorrelated with the index such that the cross product terms drop out, Equation A
simplifies to:

(B) i , j E[ p j ( I EI ) 2 ] i j

We bring the expectations operator inside the brackets to obtain:

(C) i , j i j E ( I EI ) 2 E ( i j )

Since (ij) equals zero by our assumption above that the index captures all sources of
covariance between pairs of securities, Equation C simplifies to:

(5) i , j i j I2

If our covariance calculations were to be based on 60 months of time series returns, we

would compute a single beta value for each of n securities in a portfolio and a variance
for the index itself. Thus, we could compute all (n2-n)/2 covariances from n betas and one

4
In sum, n2 covariances need to be computed for the standard portfolio variance model. However, this
number of covariances can be reduced to (n2-n)/2 non-trivial covariances since n of the covariances will
actually be variances (the covariance between any security i and itself is variance) and each i,k will equal
k,i. By this formula, we can compute that a ninety-security portfolio would require 4005 covariance
calculations.

23
variance. When n is large, the time to complete these computations will be substantially
less than the time to compute (n2-n)/2 covariances from 60 months of raw returns data.
In most cases, the single index model relies on an index representing market
returns. The most frequently used index for academic studies is the S&P 500, but other
indices such as those provided by the exchanges, Value Line and Russell may be used as
well.

Adjusted Betas
Historical betas are most frequently estimated on the basis of covariances and
variance drawn from sixty months of historical security returns. However, historical
returns and their volatility are not necessarily the best indicators of future betas.
Corporate circumstances change over time as do the markets evaluation of those
circumstances. Furthermore, any historical beta estimate would be subject to sampling
and measurement error. Blume [1975] has shown a tendency for betas to drift towards 1
over time. He proposed a correction for this tendency to drift towards one:

(6) i , F 0 1i , H

where i,F is the forecasted beta for a future five year period and i,H is the historical beta
estimated using the procedure described above. The coefficients 0 and 1 are determined
by performing a regression of five-year betas against betas estimated over the
immediately preceding five-year period. For example, the beta estimates i,F for the
period 1955-1961 based on beta estimates for 1948-1954. i,H were obtained from
adjustment coefficients 0 = .343 and 1 = .677. Note that the coefficients will normally
sum to approximately one. Other adjustment procedures exist as well, including that
proposed by Vasicek [1973].

G. Testing CAPM
Here we are concerned with how well the Capital Asset Pricing Model explains
the variation of security returns with respect to one another. First, the Capital Asset
Pricing Model implies the following:

1. The Two-Fund Theorem holds; that is, every investor holds a combination
of the market portfolio and the riskless asset.
2. The zero-beta asset return only should equal the riskless rate.
3. Only security Betas and the riskless return should affect asset returns.
Thus, residual variances, P/E ratios, firm size, dividends, squared betas,
etc. should not explain security returns.
4. Security returns are linearly related to security betas.
The slope of the SML = (rm-rf).
5. Long run market returns should exceed long-term riskless return rates.

Thus, if any of the above results do not hold for securities and markets, then the
CAPM is not a perfect descriptor for that market. A test of the Capital Asset Pricing
Model first requires that we know what returns that investors expect for securities. This is
necessary because the CAPM is an expectations based model while we can test based

24
only on historical returns. Testing the CAPM is much simplified if we may assume over
the testing and control periods the following:

a) Investors have rational expectations. Thus ex-post realizations are just

random drawings from ex-ante expectations. Then, investing in securities
is a fair game - on average realized returns equal expected returns,
b) market model holds in every period,
c) CAPM holds in every period and
d) Beta is stable over time.

To perform the actual test, we first run a first pass time-series regression to
estimate security Betas. Next, a second pass cross-sectional regression is run to test
whether CAPM implications hold. Perhaps, test the following form of CAPM:

R pt R ft 0 p 1t pt

for whether 0 0 and 1t ( Rmt R ft ).

In an early study of the CAPM, Sharpe and Cooper [1972] formed portfolios of
stocks ranked by p. The advantage of portfolio formation was that beta estimation errors
were neutralized. This study found that high portfolios earned higher returns over 36
years than low portfolios. Ninety five percent of return variation was explained by
stock Betas. Thus, the results of this study were consistent with the CAPM. The
following summarizes the Sharpe and Cooper testing procedure:

a) Group all NYSE stocks into deciles based upon beta using the previous 60
months
b) once a year, for years 1931 to 1967, betas are updated
c) an equally weighted portfolio is formed with securities from each decile
d) a strategy would be to hold securities from one decile for entire period
e) this would require trading, because stocks change deciles, and
reinvestment of dividends.

Black, Jensen and Scholes [1973] divided securities into portfolios ranked by
beta. They used the previous year's Beta as a surrogate for risk. They found in their test
that:
~
Ri 0i i 1 ,
01 rf and 1 rm rf

Though, otherwise, the CAPM seemed to hold.

Miller and Scholes [1972] discussed why the empirical SML intercept exceeds the
theoretical intercept and its slope is less than the theoretical slope. Among the
explanations are that most of the non-standard CAPM's have smaller theoretical slopes
and larger theoretical intercepts. These CAPM's may explain the empirical CAPM'S.

25
Secondly, the CAPM test is mis-specified in that the market model excludes the riskless
rate of return in estimating equation. If the riskless rate of return is constant over
estimation period there is no problem. The intercept (alpha) should be equal to (1 - i)rf.
However, if the correlation between riskless and market returns is negative, as Miller and
Scholes show (not surprising given the negative relationship between interest rates and
market returns), we have a missing variable bias and the estimated beta will be biased. In
the second pass regression, this will result in an upward bias in the intercept, the apparent
riskless rate, and a downward bias in apparent market returns, for that period, on an ex-
post basis. Furthermore, any errors in measuring bi in the first pass regression will lead to
a downward bias in the coefficient for bi in the second pass regression. In a sense,
random measurement tends to cause betas to drift towards zero. Finally, Miller and
Scholes show that errors in estimating beta will show up in residuals if they are at all
correlated, which will make it appear that residual risk should have been priced. This
effect was found to be significant.
Most tests of the CAPM, particularly the earlier ones found that the model did a
reasonable job explaining the return structure of securities. The tests did indicate some
empirical deviations from model predictions. First among these is that the empirical
Securities Market Line intercept was higher and the empirical SML slope was lower than
the CAPM predicted:

0t 0 & 1t ( Rmt R ft )

On the other hand, among the empirical findings consistent with the model was that
Squared Betas do not explain returns well. Security returns do seem linear in Beta. Also,
long run market returns exceed long run riskless rates.

References

Black, Fischer, Michael Jensen and Myron Scholes (1973): "The Capital Asset Pricing
Model: Some Empirical Tests," in Jensen (ed.), Studies in the Theory of Capital Markets,
New York: Praeger.

Greene, William H. (2003): Econometric Analysis, 5th ed. Englewood Cliffs, New
Jersey, Prentice-Hall.

Miller, Merton, and Myron Scholes (1972): Rate of Return in Relation to Risk: A
Reexamination of Some Recent Findings, in Studies in the Theory of Capital Markets.
Michael C. Jensen, ed. New York: Praeger, pp. 47-78.

Sharpe, W.F. and G.M. Cooper (1972): Risk-return classes of New York Stock Exchange
common stocks, 1931-1967. Financial Analysts Journal (March-April), 46-54, 81.

Vasicek, O. (1973): A note on using cross-sectional information in Bayesian estimation

of security betas. Journal of Finance 28, pp. 1233-1239.

26
 Exercises

1. Given the following data for GNP and sales, use a simple OLS regression to forecast
2010 sales for Smedley Company:

X Y
GNP ($000) $ Sales Year
18,000 10,000 2004
32,000 12,000 2005
47,000 16,000 2006
72,000 23,000 2007
61,000 19,000 2008
80,000 22,000 2009
90,000 ? 2010

2. The following represents sales levels for company Y for years 2004 to 2010 and GNP
levels for years 2004 to 2011:

Y Sales GNP
Year (000s) ($ billions)
2004 20 400
2005 30 450
2006 50 500
2007 80 600
2008 100 700
2009 150 800
2010 250 1000
2011 ? 1200

Based on this limited data set, forecast a sales level for 2011.

3. Historical returns for Berger Company stock, Warren Company stock and the market
portfolio along with Treasury Bill (T-Bill) rates are summarized in the following chart:

Year Berger Co. Warren Co. Market T-Bill

2006 12% 4% 10% 6%
2007 18% 20% 14% 6%
2008 7% 2% 6% 6%
2009 3% -3% 2% 6%
2010 10% 9% 8% 6%

a. Calculate return standard deviations for each of the stocks and the market
portfolio.
b. Calculate correlation coefficients between returns on each of the stocks and
returns on the market portfolio.

27
 c. Prepare graphs for each of the stocks with axes (Rit - Rft) and (Rmt - Rft), where Rit
is the historical return in year (t) for stock (i) ; Rmt and Rft are historical market
and risk-free returns in time (t). The axes that you will label are for a
Characteristic Line. Plot Characteristic Lines (characteristic line is the term for a
regression line of stock risk premiums on market risk premiums) for each of the
two stocks.
d. Calculate Betas for each of the stocks. How do your Betas compare to the slopes
of the stock Characteristic Lines?

4. Based on the following data, would you conclude that the price of Canseco Company
stock (ct) affected by the number of employees (yt) that the company hires?

Year ct yt
1 325 350
2 335 364
3 355 385
4 375 405
5 401 438
6 433 473
7 466 512
8 492 547
9 537 590
10 576 630

5. Historical returns for the Ripco Fund and the market portfolio along with Treasury
Bill (T-Bill) rates (rf) are summarized in the following table:

Year Ripco Market T-Bill Year Ripco Market T-Bill

1 0.35 0.28 0.05 11 -0.07 -0.15 0.05
2 0.04 0.05 0.05 12 -0.11 -0.21 0.05
3 0.10 0.09 0.05 13 0.42 0.26 0.05
4 -0.01 -0.02 0.05 14 0.01 0.04 0.05
5 0.38 0.32 0.05 15 0.05 0.08 0.05
6 0.31 0.25 0.05 16 0.07 0.11 0.05
7 0.33 0.26 0.05 17 -0.01 -0.03 0.05
8 0.42 0.30 0.05 18 -0.12 -0.37 0.05
9 0.07 0.14 0.05 19 0.33 0.30 0.05
10 -0.02 -0.06 0.05 20 0.05 0.13 0.05

a. Calculate the fund beta over the 20-year period.

b. Calculate the fund alpha over the 20-year period. Did the fund outperform the
market during this period on a risk-adjusted basis?

6. Portfolio risk can be measured by the sensitivity of the portfolios return to returns on
the market portfolio or index, captured by the portfolios beta, p, a measure of

28
undiversifiable risk as per the Capital Asset Pricing Model (CAPM).5 The Jensen alpha
measure was discussed in the chapter. A second CAPM-based risk-adjusted performance
indicator is the Treynor Index, computed as follows:

Rp - rf
Tp = p

The Treynor Index or Ratio calculates the portfolios reward to risk ratio, as a function of
overall portfolio risk and as a function of portfolio sensitivity to the market. The
following table provides historical percentage returns for the Patterson and Liston Funds
along with percentage returns on the market portfolio (index or fund):

Year Patterson Liston Market

2007 4 19 15
2008 7 4 10
2009 11 -4 3
2010 4 21 12
2011 5 13 9

Suppose that the riskless rate of return (or T-Bill rate) were 3% for each year. Calculate
the following based on the preceding table:
a. mean historical returns for the two funds and the market portfolio.
b. variances associated with Patterson Fund returns and Liston Fund returns along
with returns on the market portfolio.
c. the historical covariance and coefficient of correlation between returns of the
Patterson Fund and returns on the market portfolio.
d. the historical covariance and coefficient of correlation between returns of the
Liston Fund and returns on the market portfolio.
e. Historical betas for Patterson and Liston Funds, based on 5 years of returns data.
(Note that the data sets are too limited in size to be considered to be very reliable.)
f. Suppose that a market return of 8% was obtained in the year 2012, and that the
riskless return for 2012 was 3%. Further suppose that the Patterson and Liston
Funds earned 10% and 14%, respectively over 2012. Based on a Treynor Index,
how did each of these funds perform relative to the market?
g. Comment on the 2012 risk-adjusted performance for each of the two funds based
on Treynor Index results.

7. Table 7 in the text of this chapter provides a process for measuring slippage based on
perfect foresight half-spreads and transaction sizes. Table 8 continues to calculate logs of
both of these variables for a log-log regression. Create a new variable, cm, where c is
transaction size and m is an exponent parameter 1.5. Conduct a simple OLS of the perfect
foresight half-spread S on cm, the size of the transaction raised to the 1.5 power where the
regression form is: St = b0 + b1(ct)m + t. Table 7 is reproduced as follows:

5
A security beta, i, or portfolio beta, p, can be computed as the slope term from a simple OLS regression
of risk premiums on the security (ri - rf) or portfolio (rp - rf) against risk premiums of the market (rm - rf).

29
 Perfect
Foresight
Subsequent Half-
Purchase Transaction Spread Transaction
Purchase Price Price (S) Size
1 50.00 50.30 0.30 1000
2 50.26 50.30 0.04 200
3 50.26 50.35 0.09 400
4 50.17 50.33 0.16 600
5 50.32 50.33 0.01 100
6 50.41 50.46 0.05 300
7 50.51 50.55 0.04 200
8 50.44 50.47 0.03 300
9 50.71 51.12 0.41 1200
10 51.03 51.04 0.01 100
11 50.97 50.99 0.02 200
12 50.57 50.75 0.18 700
13 50.71 50.74 0.03 200
14 50.92 50.96 0.04 200
15 50.91 50.93 0.02 200
16 51.18 51.32 0.14 600
17 51.71 51.76 0.05 300
18 52.04 52.12 0.08 400
19 52.03 52.34 0.31 1000
20 52.30 52.34 0.04 300
a. What are the parameters of this regression?
b. How do these parameters compare to the parameters of the log-log regression in the
text of the chapter? (See Table 8)
c. How does the statistical significance of the parameters compare to those of the log-
log regression in Table 8 in the text of the chapter?
d. Conduct a simple linear OLS of the perfect foresight half-spread S on c. Comment
on the statistical significance of your parameters relative to those in parts a and b
of this question and in Table 8.

8. Briefly discuss the strengths and weaknesses of each of the following techniques as a
means to estimate the anticipated covariance between returns of two securities:
a. Forecasted Covariance as a function of potential return outcomes and their
associated probabilities
b. Historical Covariances
c. Single Index Betas

9. In Section 2.B, r-square values for relationships between two variables, y and x, were
calculated two different ways:

30
 n
[ ( xi x )( yi y )]2
( x , y ) 2
x2, y i 1
2 2 n n
x y
( xi x )2 ( yi y )2
i 1 i 1
(4)

n n

(6) ( yi y ) 2 i2
x2, y i 1
n
i 1

(y
i 1
i y) 2

Verify that these two equations are identical. (Hint: While there are many different ways
to complete this verification, consider using the Single Index Model).

31
 Exercise Solutions

1. First, find mean Y and X values:

n Xt 6 GNPt
X 51,667
t 1 n t 1 6
n
Y Yt/n 17,000
t 1

Next, find standard deviations of Y and X:

n 1/2
x [1/n (Xt X)2] 21,777
t 1

n 1/2
y [1/n (Yt Y)2 ] 4,850
t 1

Find the covariance between the dependent and independent variables. If desired,
find the correlation coefficient:
n
xy [1/n (Yt Y)(Xt X)] 103,490,000
t 1

x,y x,y /(x y) .984

Finally, determine regression coefficients and predicted Y values:

x y/x Pxy .218245

Y x X 5723.94

Yn 1 x Xn 1 25,365.986

2. Solve in order for , , x, y, x,y, , and Y2011:

= 635.71 = 97 x = 197.69
y = 74 x,y = 14,602 = .373
= -140.52 Y2011 = 307.08

3.a. First, calculate return standard deviations for each of the stocks (h is Berger, w is
Warren) and the market portfolio:
5
Rh Rhtn
t 1
(.12 .18 .07 .03 .10)/5 .10

32
 5
Rw Rwtn
t 1
(.04 .20 .02 .03 .09)/5 .064
5
Rm Rmtn
t 1
(.10 .14 .06 .02 .08)5 .08
1/2
(.12 .10)2 (.18 .10)2 (.07 .10)2 (.03 .10)2 (.10 .10)2
h .050199
5
1/2
(.04 .064) 2 (.2 .064)2 (.02 .064)2 ( .03 .064)2 (.09 .064)2
w .078128
5
1/2
(.10 .08)2 (.14 .08)2 (.06 .08)2 (.02 .08)2 (.08 .08)2
m .04
5

b. Calculate correlation coefficients between returns on each of the stocks and

returns on the market portfolio.
5
h,m (Rh,t E(Rh,t))(Rm,t E(Rm,t)) n
t 1

[(.12 .10)(.10 .08) (.18 .10)(.14 .08) (.07 .10)(.06 .08)

(.03 .10)(.02 .08) (.10 .10)(.08 .08)]/5
(.02)(.02) (.08)(.06) ( .03)( .02) ( .07)( .06) 0
.002
5
5
h,w (Rh,t E(Rh,t))(Rw,t E(Rw,t)) n
t 1

[(.12 .10)(.04 .064) (.18 .10)(.20 .064) (.07 .10)(.02 .064)

(.03 .10)(.03 .064) (.10 .10)(.09 .064)]/5
(.02)( .024) (.08)(.136) ( .03)( .044) ( .07)( .094) 0
.00366
5
5
w,m (Rh,t E(Rh,t))(Rw,t E(Rw,t)) n
t 1

[(.04 .064)(.10 .08) (.14 .08)(.20 .064) (.06 .08)(.02 .064)

(.02 .08)( .03 .064) (.08 .08)(.09 .064)]/5 .00284

33
 h,m covh,m/hm h,m/hm

.002/(.0502)(.04) .996

w,m covw,m/wm w,m/wm

.00284/(.078)(.04) .909
c. The slopes of the lines are the stock Betas.
d. Calculate Betas for each of the stocks. How do your betas compare to the slopes
of the stock characteristic lines?
2
h hmh,m/m

(.0502)(.04)(.996)/(.04)2
.05/.04 1.25
2
w wmw,m/m

(.078)(.04)(.909)/(.04)2
1.775

m 1

The betas are the slopes of the characteristic lines.

4. The following table summarizes preliminary computations:

Year Ct Yt Ct- Yt- (Ct-)(Yt- ) (Yt- )2

1 325 350 -105 -119 12,495 14,691
2 335 364 -95 -105 9,975 11,025
3 355 385 -75 -84 6,300 7,056
4 375 405 -55 -64 3,520 4,096
5 401 438 -29 -31 899 961
6 433 473 3 4 12 16
7 466 512 36 43 1,548 1,849
8 492 547 62 78 4,836 6,084
9 537 590 107 121 12,947 14,641
10 576 630 146 161 23,506 25,921
Preliminary Computations:
c 4,295 c 430
t

y 4,694 y 469
t

y 2 2,289,172
t

34
 [(Ct C)(Yt Y)] 76,038

(Yt Y)2 85,810

VAR[Y] 8,581
Regression Coefficients:
b1 76,038 85,810 0.89

b0 C b1Y 430 (0.89)(469) 13

Testing for Significance (Regression Diagnostics):

Year Ct E[Ct]=b0+b1Yt et=Ct-E[C] e2
1 325 325 0 0
2 335 337 -2 4
3 355 356 -1 1
4 375 373 2 4
5 401 403 -2 4
6 433 434 -1 1
7 466 469 -3 9
8 492 500 -8 64
9 537 538 -1 1
10 576 574 2 4
Computations:
et2 SSE 92

VAR[e t] et2 (n 2) 11.5

VAR[Y] VAR[et]
2x,y .99
var(Y)

VAR[e t]x 2 11.5 2,289,172

se 2 (b0 ) 31
2 10 85,810
n(xt x)

VAR[et] 11.5
se 2 (b1 ) .0001
(xt x) 2 85,810

b0
t(b0) 13 5.6 2.32
se(b0)

b1
t(b1) 0.89 .01 89
se(b1)

df = (n-2) = 8
Formal test of hypotheses:

35
 H0: b0 = 0 ; HA: b0 0
H0: b1 = 0 ; HA: b1 0
Test for significance at the 95% confidence level (two tailed test)
t(b0) = 2.32 > 2.306 ; therefore, b0 is significantly different from zero; reject the
first null hypothesis
t(b1) = 89 > 2.306 ; therefore, b1 is significantly different from zero; reject the
second null hypothesis
E[ct] = 13 + 0.89*E[yt] x,y = .99
(2.32) (89) df = 8
Thus, we can conclude that the Canseco Stock price is directly related to the
number of employees that the firm employs.

5.a. First, calculate risk premiums for both the fund and the market as follows:
t Rp Rm rf Rp-rf Rm-rf
1 0.35 0.28 0.05 0.3 0.23
2 0.04 0.05 0.05 -0.01 0
3 0.1 0.09 0.05 0.05 0.04
4 -0.01 -0.02 0.05 -0.06 -0.07
5 0.38 0.32 0.05 0.33 0.27
6 0.31 0.25 0.05 0.26 0.2
7 0.33 0.26 0.05 0.28 0.21
8 0.42 0.3 0.05 0.37 0.25
9 0.07 0.14 0.05 0.02 0.09
10 -0.02 -0.06 0.05 -0.07 -0.11
11 -0.07 -0.15 0.05 -0.12 -0.2
12 -0.11 -0.21 0.05 -0.16 -0.26
13 0.42 0.26 0.05 0.37 0.21
14 0.01 0.04 0.05 -0.04 -0.01
15 0.05 0.08 0.05 0 0.03
16 0.07 0.11 0.05 0.02 0.06
17 -0.01 -0.03 0.05 -0.06 -0.08
18 -0.12 -0.37 0.05 -0.17 -0.42
19 0.33 0.3 0.05 0.28 0.25
20 0.05 0.13 0.05 0 0.08

Next, run a simple ordinary least squares regression of (Rp-rf) on (Rm-rf). The funds beta
is the slope term in the regression and the funds alpha is its vertical intercept:

SUMMARY OUTPUT
Regression Statistics
Multiple R 0.910030353
R Square 0.828155243
Adjusted R Square 0.818608312
Standard Error 0.079179681
Observations 20

36
ANOVA
df SS MS F Significance F
Regression 1 0.543845407 0.543845407 86.74570346 2.63657E-08
Residual 18 0.112849593 0.006269422
Total 19 0.656695

Coefficients Standard Error t Stat P-value

Intercept 0.045004834 0.018088349 2.488056431 0.022868369
X Variable 1 0.895978331 0.096199656 9.313737352 2.63657E-08
Thus, the fund beta is .8959.
b. The fund alpha is .04. Since the alpha is statistically significant at the 5% level, we
can conclude that the fund has outperformed the market.

6. a. Average historical returns (in decimal format) are as follows:

b. = .000696
= .008824
= .001576
c. COV[P,M] = [ (.04-.062) (.15-.098) + (.07-.062) (.10-.098) + (.11-.062) (.03-.098)
+ (.04-.062) (.12-.098) + (.05-.062) (.09-.098) ] 5
= -.000956

d. COV[L,M] = [ (.15-.098) (.19-.106) + (.10-.098) (.04-.106)

+(.03-.098) (-.04-.106)+(.12-.098) (.21-.106)+(.09-.098) (.13-.106) ]5 = .003252

e. Stock or fund beta is: COV[S,M]/VAR[M]. The two fund betas are:

f. The Treynor Index for the market is TM = (.08 - .03)/1 = .05.

The Treynor Index for the Patterson Fund was TP = (.10 - .03)/(-0.61) = -.1148.
The Treynor Index for the Liston Fund was TL = (.14 - .03)/(2.06) = .0539
g. The Treynor Index for the market is always simply the markets risk premium.
Thus, there is nothing particularly noteworthy here other than the standard by which we
compare the performance of the two funds. The Treynor Index for the Patterson Fund is
negative, which would normally compare unfavorably to that of the market Treynor
Index. However, since the risk premium for Patterson is positive while its beta is
negative, this negative risk premium actually indicates strong risk-adjusted performance.
Thus, Patterson outperformed the market on a risk-adjusted basis. Liston also
outperformed the market on a risk adjusted basis since its Treynor Index of .0539
exceeded that of the Liston Fund.

37
7.a. Data and regression results are presented as follows:
Subsequent Perfect
Purchase Transaction Foresight Transaction Transaction
Price Price Half-Spread Size Size^1.5
1 50.00 50.30 0.300000 1000 31622.7766 SUMMARY OUTPUT
2 50.26 50.30 0.040000 200 2828.427125
3 50.26 50.35 0.090000 400 8000 Regression Statistics
4 50.17 50.33 0.160000 600 14696.93846 Multiple R 0.996
5 50.32 50.33 0.010000 100 1000 R Square 0.993438752
6 50.41 50.46 0.050000 300 5196.152423 Adjusted R Square 0.993074238
7 50.51 50.55 0.040000 200 2828.427125 Standard Error 0.009553279
8 50.44 50.47 0.030000 300 5196.152423 Observations 20
9 50.71 51.12 0.410000 1200 41569.21938
10 51.03 51.04 0.010000 100 1000 ANOVA
11 50.97 50.99 0.020000 200 2828.427125 df SS MS F
12 50.57 50.75 0.180000 700 18520.25918 Regress. 1 0.24873 0.24873 2725.38
13 50.71 50.74 0.030000 200 2828.427125 Residual 18 0.00164 9.126E-05
14 50.92 50.96 0.040000 200 2828.427125 Total 19 0.250375
15 50.91 50.93 0.020000 200 2828.427125
16 51.18 51.32 0.140000 600 14696.93846 Coefficients Standard Err. t-Stat
17 51.71 51.76 0.050000 300 5196.152423 Intercept 0.000848602 0.002890442 0.29
18 52.04 52.12 0.080000 400 8000 X Var. 1 9.75148E-06 1.86791E-07 52.20
19 52.03 52.34 0.310000 1000 31622.7766
20 52.30 52.34 0.040000 300 5196.152423

The vertical intercept b0 is not statistically different from zero.

b1 = .00001 (rounded)
St = b0 + b1 (ct)m + t
(0.29) (52.20)
b. v is approximately equal to b0 and m is approximately equal to m.
c. Parameter t-statistics, r and r-square values all indicate that the regression and
parameters are highly significant in both regressions.
d. Regression results are presented as follows:
SUMMARY OUTPUT
Regression Statistics
Multiple R 0.988779577
R Square 0.977685053
Adjusted R Square 0.976445333
Standard Error 0.017618023
Observations 20
ANOVA
df SS MS F
Regression 1 0.244787895 0.244787895 788.6342151
Residual 18 0.005587105 0.000310395
Total 19 0.250375
Coefficients Standard Error t Stat P-value
Intercept -0.047029077 0.006623535 -7.100298957 1.28304E-06
X Variable 1 0.000351833 1.25285E-05 28.08263191 2.57066E-16
Regression r-square values and t-statistics are not quite as high as in the earlier
regressions, but they still clearly imply statistical significance.

8.a. Technically, these forecasted values would be preferred, but they depend on
obtaining meaningful probabilities, which are difficult to obtain.

38
 b. Historical values are reasonable estimates when betas are stable over time.
However, historical calculations require many inputs and calculations.
c. The Single Index Model is acceptable if only one index is needed to explain the
covariance structure of security returns. The advantage to the Single Index Model
is that fewer computations may be required to write the covariance matrix for a
large sample of securities

9. We start our verification with Equation 4, then rewrite the covariance term based on
the derivation of the slope coefficient (b1) from covariance (b1 ):
( ) 2
b1 ( x )
2 2 2
b2 2
x2, y 2x , y 2 n n 1 x
x y 1 1 n
1

n i 1
( xi x ) 2 ( yi y ) 2
n i 1

n i 1
( yi y ) 2

From the Single Index Model, we know that , and we rewrite the
above equation as follows:
n n

y2 2 ( yi y ) 2 i2
x2, y i 1 i 1

y2 n

(y
i 1
i y)2

39
 Appendix 2.A: Statistics Tables

Table 2.A.1:
The z-Table

z 0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09
0.0 .0000 .0040 .0080 .0120 .0159 .0199 .0239 .0279 .0319 .0358
0.1 .0398 .0438 .0478 .0517 .0557 .0596 .0636 .0675 .0714 .0753
0.2 .0793 .0832 .0871 .0909 .0948 .0987 .1026 .1064 .1103 .1141
0.3 .1179 .1217 .1255 .1293 .1331 .1368 .1406 .1443 .1480 .1517
0.4 .1554 .1591 .1628 .1664 .1700 .1736 .1772 .1808 .1844 .1879
0.5 .1915 .1950 .1985 .2019 .2054 .2088 .2123 .2157 .2190 .2224
0.6 .2257 .2291 .2324 .2356 .2389 .2421 .2454 .2486 .2517 .2549
0.7 .2580 .2611 .2642 .2673 .2703 .2734 .2764 .2793 .2823 .2852
0.8 .2881 .2910 .2939 .2967 .2995 .3023 .3051 .3078 .3106 .3133
0.9 .3159 .3186 .3212 .3238 .3264 .3289 .3315 .3340 .3365 .3389
1.0 .3413 .3437 .3461 .3485 .3508 .3531 .3554 .3577 .3599 .3621
1.1 .3643 .3665 .3686 .3708 .3729 .3749 .3770 .3790 .3810 .3830
1.2 .3849 .3869 .3888 .3906 .3925 .3943 .3962 .3980 .3997 .4015
1.3 .4032 .4049 .4066 .4082 .4099 .4115 .4131 .4147 .4162 .4177
1.4 .4192 .4207 .4222 .4236 .4251 .4265 .4279 .4292 .4306 .4319
1.5 .4332 .4345 .4357 .4370 .4382 .4394 .4406 .4418 .4429 .4441
1.6 .4452 .4463 .4474 .4484 .4495 .4505 .4515 .4525 .4535 .4545
1.7 .4554 .4564 .4573 .4582 .4591 .4599 .4608 .4616 .4625 .4633
1.8 .4641 .4649 .4656 .4664 .4671 .4678 .4686 .4693 .4699 .4706
1.9 .4713 .4719 .4726 .4732 .4738 .4744 .4750 .4756 .4761 .4767
2.0 .4772 .4778 .4783 .4788 .4793 .4798 .4803 .4808 .4812 .4817
2.1 .4821 .4826 .4830 .4834 .4838 .4842 .4846 .4850 .4854 .4857
2.2 .4861 .4864 .4868 .4871 .4875 .4878 .4881 .4884 .4887 .4890
2.3 .4893 .4896 .4898 .4901 .4904 .4906 .4909 .4911 .4913 .4916
2.4 .4918 .492 .4922 .4925 .4927 .4929 .4931 .4932 .4934 .4936
2.5 .4938 .4940 .4941 .4943 .4945 .4946 .4948 .4949 .4951 .4952
2.6 .4953 .4955 .4956 .4957 .4959 .4960 .4961 .4962 .4963 .4964
2.7 .4965 .4966 .4967 .4968 .4969 .4970 .4971 .4972 .4973 .4974
2.8 .4974 .4975 .4976 .4977 .4977 .4978 .4979 .4979 .4980 .4981
2.9 .4981 .4982 .4982 .4983 .4984 .4984 .4985 .4985 .4986 .4986
3.0 .4986 .4987 .4987 .4988 .4988 .4989 .4989 .4989 .4990 .4990

40
Table 2.A.2:
The t-Distribution
Right-tail area,

df 0.100 0.050 0.025 0.010 0.005

1 3.078 6.314 12.706 31.821 63.657
2 1.886 2.920 4.303 6.695 9.925
3 1.638 2.353 3.182 4.541 5.841
4 1.533 2.132 2.776 3.747 4.604
5 1.476 2.015 2.571 3.365 4.032
6 1.440 1.943 2.447 3.143 3.707
7 1.415 1.895 2.365 2.998 3.499
8 1.397 1.860 2.306 2.896 3.355
9 1.383 1.833 2.262 2.821 3.250
10 1.372 1.812 2.228 2.764 3.169
11 1.363 1.796 2.201 2.718 3.106
12 1.356 1.782 2.179 2.681 3.055
13 1.350 1.771 2.160 2.650 3.012
14 1.345 1.761 2.145 2.624 2.977
15 1.341 1.753 2.131 2.602 2.947
16 1.337 1.746 2.120 2.583 2.921
17 1.333 1.740 2.110 2.567 2.898
18 1.330 1.734 2.101 2.552 2.878
19 1.328 1.729 2.093 2.539 2.861
20 1.325 1.725 2.086 2.528 2.845
21 1.323 1.721 2.080 2.518 2.831
22 1.321 1.717 2.074 2.508 2.819
23 1.319 1.714 2.069 2.500 2.807
24 1.318 1.711 2.064 2.492 2.797
25 1.316 1.708 2.060 2.485 2.787
26 1.315 1.706 2.056 2.479 2.779
27 1.314 1.703 2.052 2.473 2.771
28 1.313 1.701 2.048 2.467 2.763
29 1.311 1.699 2.045 2.462 2.756
30 1.310 1.697 2.042 2.457 2.750
1.282 1.645 1.960 2.326 2.576

Examples: The t value for degrees of freedom that bounds a right-tail area of 0.025 is 1.960. The t value for
degrees of freedom that bound left and right-tail areas (two tails) summing to 0.05 is 1.960.

41
 APPENDIX 2.B: DERIVATIVES OF POLYNOMIALS

The derivative from calculus can be used to determine rates of change or slopes.
They are also useful for finding function maxima and minima. For those functions whose
slopes are constantly changing, the derivative is to find an instantaneous rate of change;
that is, the change in y induced by the tiniest change in x. Assume that y is given as a
function of variable x. If x were to increase by a small (infinitesimal that is,
approaching, though not quite equal to zero) amount h, by how much would y change?
This rate of change is given by the derivative of y with respect to x, which is defined as
follows:

dy f ( x h) f ( x )
(1) f ' ( x) lim
dx h 0 h

One type of function which appears regularly in finance is the polynomial function. This
type of function defines variable y in terms of a coefficient c (or series of coefficients cj),
variable x (or series of variables xj) and an exponent n (or series of exponents nj). Strictly
speaking, the exponents in a polynomial equation must be non-negative integers;
however, the rules that we discuss here still apply when the exponents assume negative or
non-integer values. Where there exists one coefficient, one variable and one exponent,
the polynomial function is written as follows:
(2) y c xn

For example, let c = 7 and n = 4. Thus, our polynomial is written as follows: y = 7x4. The
derivative of y with respect to x is given by the following function:

dy
(3) c n x n 1
dx

Taking the derivative of y with respect to x in our example, we obtain: dy/dx = 7 4 x4-
1
= 28x3. Note that this derivative is always positive when x > 0; thus the slope of this
curve is always positive when x > 0. Consider a second polynomial with more than one
term (m terms total). In this second case, there will be one variable x, m coefficients (cj)
and m exponents (nj):

m
y c j x
nj
(4)
j 1

The derivative of such a function y with respect to x is given by:

dy m
c j n j x j
n 1
(5)
dx j 1

42
That is, simply take the derivative of each term in y with respect to x and sum these
derivatives. Consider a second example, a second order (the largest exponent is 2)
polynomial function given by: y = 5x2 - 3x + 2. The derivative of this function with
respect to x is: dy/dx = 10x - 3. This derivative is positive when x > .3, negative when x <
.3 and zero when x = .3. Thus, when dy/dx > 0, y increases as x increases; when dy/dx <
0, y decreases as x increases, and when dy/dx = 0, y may be either minimized or
maximized. Also notice that y is minimized when x = .3; at this point, dy/dx = 0.
As suggested above, derivatives can often be used to find minimum and
maximum values of functions. To find the minimum value of y in function y = 5x2 - 3x +
2, we set the first derivative of y with respect to x equal to zero and then solve for x. For
our example, the minimum is found as follows:

10 x 3 0
10 x 3
3
x
10

In order to ensure that we have found a minimum (rather than a maximum), we

check the second derivative. The second derivative is found by taking the derivative of
the first derivative. If the second derivative is greater than zero, we have a minimum
value for y (the function is concave up). When the second derivative is less than zero, we
have a maximum (the function is concave down). If the second derivative is zero, we
have neither a minimum nor a maximum. The second derivative in the above example is
given by: d2y/dx2 = 10, also written f"(x) = 10. Since the second derivative 10 is greater
than zero, we have found a minimum value for y. In many cases, more than one local
minimum or maximum value will exist.
Consider a third example where our second order polynomial is given: y= -7x2 +
4x + 5. The first derivative is: dy/dx = -14x + 4. Setting the first derivative equal to zero,
we find our maximum as follows:

14 x 4 0
14 x 4
4
x
14

We check second order conditions (the second derivative) to ensure that this is a
maximum. The second derivative is: d2y/dx2 = -14. Since -14 is less than zero, we have a
maximum at dx = 4/14.

43