Chapter 1 The Constant Expected Return Model

The first model of asset returns we consider is the very simple constant expected return (CER) model. This model assumes that an asset’s return over time is normally distributed with a constant (time invariant) mean and variance The model also assumes that the correlations between asset returns are constant over time. Although this model is very simple, it allows us to discuss and develop several important econometric topics such as estimation, hypothesis testing, forecasting and model evaluation.

1.0.1

Constant Expected Return Model Assumptions

Let Rit denote the continuously compounded return on an asset i at time t. We make the following assumptions regarding the probability distribution of Rit for i = 1, . . . , N assets over the time horizon t = 1, . . . , T. 1. Normality of returns: Rit ∼ N (µi , σ 2 i ) for i = 1, . . . , N and t = 1, . . . , T. 2. Constant variances and covariances: cov (Rit , Rjt ) = σ ij for i = 1, . . . , N and t = 1, . . . , T. 3. No serial correlation across assets over time: cov (Rit , Rjs ) = 0 for t 6= s and i, j = 1, . . . , N. Assumption 1 states that in every time period asset returns are normally distributed and that the mean and the variance of each asset return is constant over time. In particular, we have for each asset i and every time period 1

2CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL t E [Rit ] = µi var(Rit ) = σ 2 i The second assumption states that the contemporaneous covariances between assets are constant over time. Given assumption 1, assumption 2 implies that the contemporaneous correlations between assets are constant over time as well. That is, for all assets and time periods corr(Rit , Rjt ) = ρij The third assumption stipulates that all of the asset returns are uncorrelated over time1 . In particular, for a given asset i the returns on the asset are serially uncorrelated which implies that corr(Rit , Ris ) = cov (Rit , Ris ) = 0 for all t 6= s. Additionally, the returns on all possible pairs of assets i and j are serially uncorrelated which implies that corr(Rit , Rjs ) = cov (Rit , Rjs ) = 0 for all i 6= j and t 6= s. Assumptions 1-3 indicate that all asset returns at a given point in time are jointly (multivariate) normally distributed and that this joint distribution stays constant over time. Clearly these are very strong assumptions. However, they allow us to development a straightforward probabilistic model for asset returns as well as statistical tools for estimating the parameters of the model and testing hypotheses about the parameter values and assumptions.

1.0.2

Regression Model Representation

A convenient mathematical representation or model of asset returns can be given based on assumptions 1-3. This is the constant expected return (CER) regression model. For assets i = 1, . . . , N and time periods t = 1, . . . , T the CER model is represented as Rit = µi + εit εit ∼ iid. N (0, σ 2 i) cov (εit , εjt ) = σ ij
1

(1.1) (1.2)

Since all assets are assumed to be normally distributed (assumption 1), uncorrelatedness implies the stronger condition of independence.

3 where µi is a constant and εit is a normally distributed random variable with mean zero and variance σ 2 i . Notice that the random error term εit is independent of εjs for all time periods t 6= s. The notation εit ∼ iid. N (0, σ 2 i) stipulates that the random variable εit is serially independent and identically distributed as a normal random variable with mean zero and variance σ 2 i. This implies that, E [εit ] = 0, var(εit ) = σ 2 and cov ( ε , ε ) = 0 for i = 6 j and it js i t 6= s. Using the basic properties of expectation, variance and covariance discussed in chapter 2, we can derive the following properties of returns. For expected returns we have E [Rit ] = E [µi + εit ] = µi + E [εit ] = µi , since µi is constant and E [εit ] = 0. Regarding the variance of returns, we have var(Rit ) = var(µi + εit ) = var(εit ) = σ 2 i which uses the fact that the variance of a constant (µi ) is zero. For covariances of returns, we have cov (Rit , Rjt ) = cov (µi + εit , µj + εjt ) = cov(εit , εjt ) = σ ij and cov (Rit , Rjs ) = cov (µi + εit , µj + εjs ) = cov(εit , εjs ) = 0, t 6= s, which use the fact that adding constants to two random variables does not affect the covariance between them. Given that covariances and variances of returns are constant over time gives the result that correlations between returns over time are also constant: σ ij cov (Rit , Rjt ) = corr(Rit , Rjt ) = p = ρij , σiσj var(Rit )var(Rjt ) cov (Rit , Rjs ) 0 corr(Rit , Rjs ) = p = = 0, i 6= j, t 6= s. σiσj var(Rit )var(Rjs ) Rit ∼ i.i.d. N (µi , σ 2 i ).

Finally, since the random variable εit is independent and identically distributed (i.i.d.) normal the asset return Rit will also be i.i.d. normal:

3 Interpretation of the CER Regression Model The CER model has a very simple form and is identical to the measurement error model in the statistics literature. the model states that each asset return is equal to a constant µi (the expected return) plus a normally distributed random variable εit with mean zero and constant variance. Good news should lead to positive values of εit and εjt . Suppose we are interested in the annual continA ˙ ? Since multiperiod continuously uously compounded return Rit = Rit (12) . 1.0. The random news variable affecting asset i. Then one interpretation of news in this context is general news about the computer industry and technology. Consider the default case where Rit is interpreted as the continuously compounded monthly return on asset i. If the news is bad.4CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL Hence.1) for Rit is equivalent to the model implied by assumptions 1-3. then εjt is negative and the observed return is less than expected. If the news between times t − 1 and time t is good.1) we can write εit as εit = Rit − µi = Rit − E [Rit ] so that εit is defined to be the deviation of the random return from its expected value. on average. εjt . Hence these variables will be positively correlated. The assumption that E [εit ] = 0 means that news. to capture the idea that news about one asset may spill over and affect another asset. let asset i be Microsoft and asset j be Apple Computer. For example. is neutral. neither good nor bad. note that using (1. The random variable εit can be interpreted as representing the unexpected news concerning the value of the asset that arrives between times t − 1 and time t. The assumption that var(εit ) = σ2 i can be interpreted as saying that volatility of news arrival is constant over time. is allowed to be contemporaneously correlated with the random news variable affecting asset j. Time Aggregation and the CER Model The CER model with continuously compounded returns has the following nice property with respect to the interpretation of εit as news. In words. To see this. εit . then the realized value of εit is positive and the observed return is above its expected value µi . the CER model (1.

Hence. is simply 12 times the monthly expected return. Using the results from chapter 2 about the variance of a sum of random variables. we will ignore the fact that some assets do not trade over the weekend. the annual expected return. the variance of the annual news component is just 12 time the variance of the monthly new component: var(εA it ) = var = = 11 X k=0 11 X k=0 Ã 11 X k=0 εit−k ) ! var(εit−k ) since εit is uncorrelated over time σ2 i since var (εit ) is constant over time = 12 · σ 2 i A = var(Rit ) For simplicity of exposition. µi . εA it . 2 .5 compounded returns are additive. The annual random news component. Rit (12) is the sum of 12 monthly continuously compounded returns2 : A Rit = Rit (12) = 11 X t=0 Rit−k = Rit + Rit−1 + · · · + Rit−11 Using the CER model representation (1. µA i . is the accumulation of news over the year.1) for the monthly return Rit we may express the annual return Rit (12) as Rit (12) = 11 X (µi + εit ) t=0 11 X t=0 = 12 · µi + = µA i + εA it εit A where µA i = 12 · µi is the annual expected return on asset i and εit = P 11 k=0 εit−k is the annual random news component.

Rjt ) A The above results imply that the correlation between εA it and εjt is the same as the correlation between εit and εjt : A corr(εA it . recall that the continuously compounded return. εjt ) A var(εA it ) · var (εjt ) 2 12σ 2 i · 12σ j = q = 12 · σ ij σ ij = ρij σi σj = corr(εit . using results from chapter 2 about the additivity of covariances A we have that covariance between εA it and εjt is just 12 times the monthly covariance: Ã 11 ! 11 X X A cov(εA εit−k . εjt ) 1.4 The CER Model of Asset Returns and the Random Walk Model of Asset Prices The CER model of asset returns (1. To see this. is defined from asset prices via ¶ µ Pit = Rit . .1) gives rise to the so-called random walk (RW) model of the logarithm of asset prices.0.6CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL Similarly. εjt ) is constant over time = 12 · σ ij A A = cov (Rit . εjt−k ) since εit and εjt are uncorrelated over time σ ij since cov(εit . Rit . εjt ) = cov k=0 k=0 = = 11 X k=0 11 X k=0 cov (εit−k . εjt−k it . εjt ) = q A cov (εA it . ln Pit−1 Since the log of the ratio of prices is equal to the difference in the logs of prices we may rewrite the above as ln(Pit ) − ln(Pit−1 ) = Rit .

7 Letting pit = ln(Pit ) and using the representation of Rit in the CER model (1. εit = pit − pit−1 − E [pit − pit−1 ]. Notice that at time t = 0 the expected price at time t = 1 is E [pi1 ] = pi0 + µi + E [εi1 ] = pi0 + µi which is the initial price plus the expected return between time 0 and 1. E [pit − pit−1 ] = E [Rit ] = µi . we may further rewrite the above as pit − pit−1 = µi + εit . the unexpected changes in asset prices. In the RW model.3) The representation in (1.1). Further. The RW model says that the price at time t = 1 is pi1 = pi0 + µi + εi1 where εi1 is the value of random news that arrives between times 0 and 1. in the RW model. Similarly. εis ) = 0 for t 6= s) so that future changes in asset prices cannot be predicted from past changes in asset prices3 . That is. The RW model gives the following interpretation for the evolution of asset prices. are uncorrelated over time (cov (εit . the price at time t = 2 is pi2 = pi1 + µi + εi2 = pi0 + µi + µi + εi1 + εi2 2 X = pi0 + 2 · µi + εit t=1 The notion that future changes in asset prices cannot be predicted from past changes in asset prices is often referred to as the weak form of the efficient markets hypothesis. Let pi0 denote the initial log price of asset i. µi represents the expected change in the log of asset prices (continuously compounded return) between months t − 1 and t and εit represents the unexpected change in prices. (1.3) is know as the RW model for the log of asset prices. εit . 3 .

4 . The process of creating such pseudo data is often called Monte Carlo simulation4 . pi0 . The plot shows the log price. P pt . The sailor generally moves in the direction described by µ but randomly deviates from this direction after each step t by an amount equal to εt . the expected price E [pt ] = p0 + 0. plus the accumulated random news over the two periods. The sailor starts at an initial position.1 illustrates the random walk model of asset prices based on the CER model with µ = 0.1) for one asset. outside the bar. The steps to create a Monte Carlo simulation from the CER model are: • Fix values for the CER model parameters µ and σ (or σ 2 ) Monte Carlo referrs to the fameous city in Monaco where gambling is legal.1 Monte Carlo Simulation of the CER Model A good way to understand the probabilistic behavior of a model is to use computer simulation methods to create pseudo data from the model. After T steps the sailor ends up at position PT pT = p0 + µ · T + t=1 εt . To illustrate the use of Monte Carlo simulation. the price at time t = T is piT = pi0 + T · µi + T X t=1 εit . plus the two period expected P2 return.05t and the accumulated random news t t=1 εt . 2 · µi . piT . deviates from the expected price by the accumulated random news T X piT − E [piT ] = εit . consider the problem of creating pseudo return data from the CER model (1. t=1 Figure 1.05. p0 . 1. σ = 0. At time t = 0 the expected price at time t = T is E [piT ] = pi0 + T · µi The actual price. By recursive substitution. t=1 εit . The term random walk was originally used to describe the unpredictable movements of a drunken sailor staggering down the street.8CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL which is equal to the initial price.10 and p0 = 1.

. .The simulated returns are then computed as ∗ Rt = 0.Let {ε∗ 1 . . . εT . The key to simulating data from the above model is to simulate T = 100 observations of the random news variable εt ~iid N (0. t = 1. .1. . . Computer algorithms exist which can easily create such ∗ observations. T. . . ∗ = µ + ε∗ • Create simulated return data Rt t for t = 1. T To mimic the monthly return data on Microsoft.05 + ε∗ t . .1 MONTE CARLO SIMULATION OF THE CER MODEL 9 6 p(t) E[p(t)] p(t)-E[p(t)] 0 0 2 4 20 40 60 80 100 Figure 1. • Determine the number of simulated values. the values µ = 0. ε100 } denote the 100 simulated values of εt .05 and σ = 0.. . • Use a computer random number generator to simulate T iid values of εt from N (0. 100 ∗ values are given in figure A time plot and histogram of the simulated Rt .The simulated return data fluctuates randomly about the expected return . (0. Denote these simulated values are ∗ ε∗ 1 .10 are used as the model’s parameters and T = 100 is the number of simulated values (sample size). σ 2 ) distribution. to create. . .1: Simulated random walk model for log prices. . .10)2 ). .

0914.05. Monte Carlo simulation of a model can be used as a first pass reality check of the model.1 frequency 0.0 -0.3 0.1 0.0 0. The typical size of the fluctuation is approximately equal to SD(εt ) = 0.2 0.0522)) = 0. These values are very close to the population values E [Rt ] = 0. respectively. (0. 1. εt ~iid N (0. However. P100 ∗ 1 The sample average of the simulated q P returns is 100 t=1 Rt = 0.0522 and 100 1 ∗ 2 the sample standard deviation is 99 t=1 (Rt − (0.10.3 return Figure 1.2 -0.1 0.10. Notice that the simulated return data looks remarkably like the actual monthly return data for Microsoft.2 0.1 -0. if simulated data looks reasonably close to the data that the model is suppose to describe then confidence is instilled on the model.1.2: Simulated returns from the CER model Rt = 0.05 + εt .05 and SD(Rt ) = 0.10)2 ) value E [Rt ] = µ = 0.10CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL Simulated returns from CER model 30 Histogram of simulated returns 0. If simulated data from the model does not look like the data that the model is supposed to describe then serious doubt is cast on the model.2 0 20 40 months 60 80 100 return 0 5 10 15 20 25 -0.1 Simulating End of Period Wealth To be completed .

In actuality. We call {Ri1 .1 Estimating the Parameters of the CER Model The Random Sampling Environment The CER model of asset returns gives us a rigorous way of interpreting the time series behavior of asset returns. compare computations where end of period wealth is based on the expected return over the period versus computations based on simulating different sample PN paths and then taking the average. N (µi . Let {ri1 . . we do not know these values with certainty. σ 2 i ). . .2. At the beginning of every month t.2 1. . . T. .1.1).d. The CER model states that Rit ∼ i. It is assumed that the observed returns are realizations of the time series of random variables {Ri1 . . . . however. RiT } . .1. The CER model assumes that the economic environment is constant over time so that the normal distribution characterizing monthly returns is the same every month. riT } the realized values . Our best guess for the return at the end of the month is E [Rit ] = µ . . RiT } a random sample from the CER model (1. . where Rit is described by the CER model (1. riT } denote the observed history of T monthly returns on asset i for i = 1. . . N. . Rit is a random variable representing the return to be realized at the end of the month. compute E [W0 exp( t=1 Rt )] where Rt behaves according to the CER model and compare this to W0 exp(Nµ). . . Our life would be very easy if we knew the exact values of µi . σ 2 i and σ ij . A key task in financial econometrics is estimating the values of µi . 1. . . . Suppose we observe monthly returns on N different assets over the horizon t = 1. Rjt ). σ 2 i and σ ij from a history of observed data. our measure of uncertainty i p about our best guess is captured by σ i = var(Rit ) and our measure of the direction of linear association between Rit and Rjt is σ ij = cov (Rit . .1) and we call {ri1 . the parameters of the CER model.2 Simulating Returns on More than One Asset To be completed 1. .2 ESTIMATING THE PARAMETERS OF THE CER MODEL11 • insert example showing how to use Monte Carlo simulation to compute expected end of period wealth. . Essentially. .i.

riT }. The goal is to estimate θ based on the observed data {ri1 . µ. Let θ denote some characteristic of the CER model (1. . For example. Before the sample is observed. RiT ) denote an estimator of θ treated as a function of the random variables {Ri1 . Let θ(ri1 .2. . . −0. . An estimator of the expected return. if we are interested in the variance of returns then θ = σ 2 i . riT ) is simply an number. . . RT denote a random sample of returns. . the sample average is a simple linear function of the random variables {Ri1 . .2 Statistical Estimation Theory Before we describe the estimation of the CER model. we can use the observed returns to estimate the unknown parameters of the CER model 1. . . The context will determine how to interpret ˆ θ. . RiT } and so is itself a random variable. . . Clearly. RT ) = Rt T t=1 . After the sample {ri1 . . For example. . . . riT }. . . . . 0. . . ˆ θ(ri1 . . . if the observed sample is {0. . .12CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL from the random sample. Under these assumptions. . PT 1 Example 3 The sample average T t=1 Rit is an algorithm for computing an estimate of the expected return µi . . riT }.10) = −0. . . RiT } Definition 2 An estimate of θ is simply the value of an estimator based on the realized sample values {ri1 . is the sample average T 1X µ ˆ (R1 . Definition 1 An estimator of θ is a rule or algorithm for forming an estimate for θ based on the random sample {Ri1 . .03. . riT } is observed. . ˆ θ(Ri1 . . if we are interested in the expected return then θ = µi . .03 − 0. it is useful to summarize some concepts in the statistical theory of estimation. erage estimate is 1 3 To discuss the properties of estimators it is necessary to establish some notation. riT ) denote an estimate of θ based on the realized values {ri1 . . We will often use ˆ θ as shorthand notation to represent either an estimator of θ or an estimate of θ. . . .05 + 0. . . .05. . . RiT ) is a random ˆ variable.1) we are interested in estimating.02. . which is just a number. Example 4 Let R1 . . . RiT }. . . PT 1 the sample average can be evaluated giving T r t=1 it .10} then the sample av(0. . . Let ˆ θ(Ri1 .

where “on average” . Bias Bias concerns the location or center of p(ˆ θ).1 + 0. .e. a good estimator of θ is one that will produce an estimate ˆ θ that is close θ all of the time. a good estimator will have small bias and high precision.05) = (0. θ) = 0. If p(θ) is centered at θ then we say that ˆ θ is unbiased. r3 = 0. In general. If p(ˆ θ ) is centered away from θ ˆ ˆ then we say θ is biased. p(ˆ θ). −0. The exact form of p(ˆ θ) may be very complicated. . we often focus on certain characteristics of p(ˆ θ) like its expected value (center). Then the estimate of the expected return using the sample average is 1 µ ˆ (0. .3 Properties of Estimators Consider ˆ θ = ˆ θ (Ri1 . It means that the estimator produces the correct answer “on average”. variance and standard deviation (spread about expected value).2. For analysis purposes. the pdf of ˆ θ. That is.05) = 0. . . RiT ) as a random variable..1. r4 = −0. RiT . . .05. Intuitively.1. The expected value of an estimator is related to the concept of estimator bias and the variance/standard deviation of an estimator is related estimator precision.1.1.005 5 1. .1 + −0. if E [ˆ θ] = θ or E [error] = 0. . Definition 6 The bias of an estimator ˆ θ of θ is given by bias(ˆ θ. Unbiasedness is a desirable property of an estimator.05. r5 = −0. θ) = E [ˆ θ] − θ. Formally we have the following definitions: Definition 5 The estimation error is difference between the estimator and the parameter being estimated error = ˆ θ − θ.025. . .05 + 0. depends on the pdf’s of the random variables Ri1 .2 ESTIMATING THE PARAMETERS OF THE CER MODEL13 Suppose T = 5 and the realized values of the returns are r1 = 0.025 + −0. . r2 = 0. Definition 7 An estimator ˆ θ of θ is unbiased if bias(ˆ θ. i.

6 0. Although the value of ˆ θ2 is not equal to 0 on average we might prefer the estimator ˆ θ2 over ˆ θ1 because it is generally closer to θ = 0 on average than ˆ θ1 .3. For example. var(ˆ θ1 ) is large. means over many hypothetical samples.2 0. hopefully. θ) = 0.1 0 -10 -5 0 estimator value 5 10 pdf 1 pdf 2 Figure 1. E [ˆ θ 1 ] = 0. It is important to keep in mind that an unbiased estimator for θ may not be very close to θ for a particular sample and that a biased estimator may be actually be quite close to θ. The center of the pdf is slightly higher than θ = 0. The center of the distribution is at the true value θ = 0.4 0. Now consider the pdf for ˆ θ2 .3 0. unbiasedness by itself does not guarantee a good estimator of θ .3: Pdf values for competing estimators of θ = 0. consider the pdf of ˆ θ1 in figure 1. Hence. That is. Precision An estimate is.14CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL Pdfs of competing estimators 0. but the spread of the distribution is small. Our guess most certainly will be wrong but we hope it will not be too far .8 0.25. On average (over many hypothetical samples) the value of ˆ θ1 will be close to θ but in any given sample the value of ˆ θ1 can be quite a bit above or below θ. our best guess of the true (but unknown) value of θ.5 pdf 0. bias(ˆ θ2 . but the distribution is very widely spread out about θ = 0.7 0.

Intuitively. θ) = var(ˆ θ) = E [(ˆ θ − θ)2 ] is just the squared deviation of ˆ θ about θ.2. then we know that θ will almost always be close to θ . . loosely speaking. bias(ˆ θ . . r100 }. . If this expected deviation is small. . . if the mean squared is large then it is possible to see samples for which ˆ θ to be quite far from θ.4. θ) is given in the following proposition ³ ´2 2 ˆ ˆ ˆ ˆ θ)+bias(ˆ θ. A precise estimate. θ)2 Proposition 9 mse(θ. riT } denote the realized values from the random sample. . Hence. is one that has a small estimation error.2 ESTIMATING THE PARAMETERS OF THE CER MODEL15 off. θ) can be ˆ split into a variance component. The proposition states that for any estimator ˆ θ of θ. and a bias component. . The typical size of a deviation about 0.Notice that the data seem to fluctuate up and down about some central value near 0.1). θ )2 . . var(θ ).4 Method of Moment Estimators for the Parameters of the CER Model Let {Ri1 . The magnitude of the estimation error is usually captured by the mean squared error: Definition 8 The mean squared error of an estimator ˆ θ of θ is given by mse(ˆ θ . consider the observed monthly continuously compounded returns.10.03 is roughly 0. A useful decomposition ˆ of mse(θ. . for Microsoft stock over the period July 1992 through October 2000. 1.1. the parameter µi = E [Rit ] in the CER model represents this central value and σ i represents the typical size of a deviation about µi . . an unbiased estimator ˆ θ of θ is good if it has a small variance. mse(ˆ θ. Alternatively. . . θ) will be small only if both components are small. If an estimator is unbiased then mse(ˆ θ. RiT } denote a random sample from the CER model and let {ri1 . These data are illustrated in figure 1. As an example. θ) = E [(ˆ θ − θ)2 ] = E [error2 ] ˆ The mean squared error measures the expected squared deviation of θ ˆ from θ. Consider the problem of estimating the parameter µi in the CER model (1. mse(ˆ θ. Clearly. {r1 . θ) = E [(θ−E [θ]) ]+ E [θ] − θ = var(ˆ The proof of this proposition is straightforward and is given in the appendix. .03.

4: Monthly continuously compounded returns on Microsoft stock. the method of moments estimator solves T T 1X 1X ˆ εit = (rit − µ ˆ i ) = E [εit ] = 0 T t=1 T t=1 (1.1 0. This is the estimated news component for month t based on the estimate µ ˆi. T.16CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL returns -0. t = 1. . The sample error or residual at time t associated with this estimate is defined as ˆ εit = rit − µ ˆ i . . That is.4) 5 In this book. .1 0. quantities with a “ˆ” denote an estimate.2 Q1 Q3 1994 Q1 Q3 1995 Q1 Q3 1996 Q1 Q3 1997 Q1 Q3 1998 Q1 Q3 1999 Q1 Q3 2000 Figure 1.0 0. Now the CER model imposes the condition that the expected value of the true error is zero E [εit ] = 0 The method of moments estimator of µi is the value of µ ˆ i that makes the average of the sample errors equal to the expected value of the population errors. . The method of moments estimate of µi Let µ ˆ i denote a prospective estimate of µi 5 . .4 Q3 Q1 Q3 1992 1993 -0.3 -0.2 -0.

Solving (1. N ) in the CER model is simply the sample average of the observed returns for asset i.2 ESTIMATING THE PARAMETERS OF THE CER MODEL17 Returns on Microsoft returns -0.0 1992 1993 1994 1995 1996 1997 1998 1999 2000 Returns on S&P 500 returns -0.05 1992 1993 1994 1995 1996 1997 1998 1999 2000 Figure 1. .1. .5) Hence.5: Monthly continuously compounded returns on Microsoft. . The returns are shown in figure For the T = 100 monthly . Starbucks and the S&P 500 index over the period July 1992 through October 2000. . T t=1 µ ˆi = (1. Example 10 Consider the monthly continuously compounded returns on Microsoft.0 1992 1993 1994 1995 1996 1997 1998 1999 2000 Returns on Starbucks returns -0. Starbucks and the S&P 500 Index.15 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 0. the method of moments estimate of µi (i = 1.4 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 0.4) for µ ˆ i gives the method of moments estimate of µi : T 1X rit = r ¯.4 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4 0.

Without going into the details. σ i .0276 100 t=1 1 X = rsbux. The method of moments estimates of σ 2 i .25% per month. t=1 rit = µ Notice that (1.8) is the sample covariance of the observed returns on assets i and j and (1. T T (1. σ ij and ρij are given by the sample descriptive statistics σ ˆ2 i 1 X = (rit − r ¯i )2 . σ i .t = 0.8) (1.8% per month whereas the mean return for SP500 is smaller at only 1. (1.9) is the sample correlation of returns on assets i and j. the method of moments estimates of σ 2 i .t = 0. T − 1 t=1 q σ ˆi = σ ˆ2 i. σ ij and ρij are defined analogously to the method of moments estimator for µi . σ i . σ ij and ρij The method of moments estimates of σ 2 i . Example 11 Consider again the monthly continuously compounded returns on Microsoft.0278 100 t=1 100 100 100 1 X = rsp500.6) is simply the sample variance of the observed returns for asset i.7) is the sample standard deviation.9) σ ˆ ij = ρ ˆij = σ ˆ ij σ ˆ iσ ˆj 1 X (rit − r ¯i )(rjt − r ¯j ).t = 0.0125 100 t=1 The mean returns for MSFT and SBUX are very similar at about 2. (1.18CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL continuously returns the estimates of E [Rit ] = µi are µ ˆ msf t µ ˆ sbux µ ˆ sp500 1 X = rmsf t. Starbucks and the S&P 500 index over the period July 1992 . T − 1 t=1 PT 1 where r ¯i = T ˆ i is the sample average of the returns on asset.i.7) (1.6) (1.

8) and (1.15 -0.3 -0. .0379 SBUX has the most variable monthly returns and SP500 has the smallest.15 -0.05 -0.3 0.3 -0.The estimates of σ ij and ρij using (1.0114. σ ˆ msf t.1 0. σ i .5 -0.sp500 = 0.05 0. σ ˆ sbux.5 -0.5 msft 0.2 ESTIMATING THE PARAMETERS OF THE CER MODEL19 -0.3 -0.00 sp500 -0.6. ρ ˆsbux.7) are σ ˆ2 ˆ msf t = 0.1 -0.1 0.0022.sbux = 0.1359 σ ˆ2 sbux = 0.SP500) appear to be the most correlated.2777.10 Figure 1. The estimates of the parameters σ 2 i .3 0. All returns appear to be positively related.1 0.sp500 = 0. σ ˆ sbux = 0. ρ These estimates confirm the visual results from the scatterplot matrix.1 0.3 -0.05 0.9) are σ ˆ msf t.6) and (1.20 -0.sbux = 0.0185.10 -0.1.3 -0.SP500) and (SBUX. σ ˆ sp500 = 0.5 0.10 0.6: Scatterplot matrix of monthly returns on Microsoft.0014.10 -0.05 0. using (1. Starbucks and S&P 500 index.sp500 = 0.1 -0.1 sbux -0. The pairs (MSFT.0022 ˆmsf t.1 -0. through October 2000.3 0.sp500 = 0.4197 ρ ˆmsf t. σ 2 σ ˆ sp500 = 0.20 -0.00 0. The scatterplots of the returns are illustrated in figure 1.0040.5551.1068 msf t = 0.

.20CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL 1. . Using results from chapter 2 about the expectation of a linear combination of random variables it is straightforward to show (details are given in the appendix) that E [ˆ µi ] = µi Hence.12) T . RiT ) = µ ˆi = µ T 1X Rit T t=1 (1. . Precision P µi ) = var(T −1 T To determine the variance of µ ˆ i we must compute var(ˆ t=1 Rit ). . . p(ˆ µi ) is a normal density. (1. . . we treat it as a function of the random sample Ri1 . In other words. Bias In the CER model. Using the results from chapter 2 about the variance of a linear combination of uncorrelated random variables it is easy to show (details in the appendix) that σ2 (1. .1 Statistical Properties of Estimates Statistical Properties of µ ˆi To determine the statistical properties of µ ˆ i in the CER model. RiT : ˆ i (Ri1 .10) is an average of these normal random variables it is also normally distributed.3 1. . µ ˆi an unbiased estimator for µi . Since the method of moments estimator (1.11) var(ˆ µi ) = .10) where Rit is assumed to be generated by the CER model (1. the random variables Rit (t = 1. the mean of the distribution of µ ˆ i is equal to µi .1). ToP determine the mean of T −1 this distribution we must compute E [ˆ µi ] = E [T t=1 Rit ]. . That is. . µit ) The standard deviation of µ ˆ i is just the square root of var(ˆ p σi SD(ˆ µi ) = var(ˆ µi ) = √ .3. . T Notice that the variance of µ ˆ i is equal to the variance of Rit divided by the sample size and is therefore much smaller than the variance of Rit . T ) are iid normal with mean µi and variance σ 2 i .

if SE (ˆ µi ) is large relative to µi then µ ˆ i is a relatively imprecise estimate of µi because p(ˆ µi ) will be spread out about µi .7: Pdfs for µ ˆ i with small and large values of SE (ˆ µi ).13) T SE (ˆ µi ) is in the same units as µ ˆ i and measures the precision of µ ˆ i as an ˆ i then µ ˆ i is a relatively precise of estimate. If SE (ˆ µi ) is small relative to µ µi because p(ˆ µi ) will be tightly concentrated around µi .14) T which is just (1.0 0.6 pdf 1 pdf 2 -3 -2 -1 0 estimate value 1 2 3 Figure 1.4 0. The standard deviation of µ ˆ i is most often referred to as the standard error of the estimate µ ˆ i: σi SE (ˆ µi ) = SD(ˆ µi ) = √ .8 pdf 0. . Figure 1. To get a practically useful measure of precision for µ ˆ i we compute the estimated standard error p σ bi c (ˆ SE µi ) = v d ar(ˆ µi ) = √ (1.7 illustrates these relationships Unfortunately.1.13) withq the unknown value of σ i replaced by the method of b2 moments estimate σ bi = σ i . (1. True value of µi = 0.3 STATISTICAL PROPERTIES OF ESTIMATES 21 0. SE (ˆ µi ) is not a practically useful measure of the precision of µ ˆ i because it depends on the unknown value of σ i .2 0.

.The histogram of the estimated means. r50 }. SE (ˆ µi ) represents the standard deviation of these µ ˆ i values.01359 100 0. . . the c (ˆ values of SE µi ) are 0.Notice that there is considerable variation in the simulated samples but that all of the simulated samples fluctuates about the true mean value of µ = 0. computations involved in evaluating E [ˆ µi ] and SE (ˆ To illustrate. 1000 j∗ j∗ .05. p(ˆ µ). .01068 SE µmsf t ) = √ 100 0.15) and simulate N = 1000 samples of size T = 50 values from the above model using the technique of Monte Carlo simulation. Notice that the center of the histogram is 50 very close to the true mean value µ = 0. Strictly speaking.1068 c (ˆ = 0. . (0. consider the CER model Rt = 0. t = 1. .8. Starbucks and S&P 500 return data. That is. Similarly. A histogram of these 1000 mean values is illustrated in figure 1. .05 0. . on average over the .10 with SE (ˆ µi ) = √ = 0. For each of the 1000 simulated samples the estimate µ ˆ is formed giving 1000 mean estimates {µ ˆ 1. This gives j = 1. .9. .003785 100 Clearly. The first 10 of these sample realizations sample realizations {r1 are illustrated in figure 1. . . . of the estimator µ ˆ which we know is a Normal pdf centered at µ = 0.0379 c (ˆ SE µsp500 ) = √ = 0.22CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL Example 12 For the Microsoft. In this way we can approximate the µi ).05.05 + εit . the mean return µi is estimated more precisely for the S&P 500 index than it is for Microsoft and Starbucks. We may think of these hypothetical samples as Monte Carlo simulations of the CER model. 50 εit ~iid N (0. µ ˆ 1000 }. E [ˆ µi ] = µi means that over an infinite number of repeated samples the average of the µ ˆ i values computed over the infinite samples is equal to the true value µi .10)2 ) (1. µ ˆ j . . µi ) using Monte Carlo simulation Interpreting E [ˆ µi ] and SE (ˆ The statistical concepts E [µi ] = µi and SE (µi ) are a bit hard to grasp at first. .1359 c (ˆ SE µsbux ) = √ = 0.01414. can be thought of as an estimate of the underlying pdf. .

. εt ~iid N (0. In some samples.2 0. µ ˆ 1000 } from the 1000 simulated samples is 1 X j µ ˆ = 0. In fact.05 + εt .3 10 20 30 40 50 Figure 1.05. .0 0.The value of SE (ˆ µi ) may be approximated by computing the sample standard deviation of the 1000 1000 . the estimate is too big and in some samples the estimate is too small but on average the estimate is correct.2 0 -0. the average value of {µ ˆ1.(0.05045 1000 j =1 which is very close to the true value.1 0.1.1 0.10)2 ) 1000 Monte Carlo samples the value of µ ˆ is about 0. . If the number of simulated samples is allowed to go to infinity then the sample average of µ ˆ j will be exactly equal to µ : N 1 X j lim µ ˆ =µ N →∞ N j =1 The typical size of the spread about the center of the histogram represents SE (ˆ µi ) and gives an indication of the precision of µ ˆ i . .3 STATISTICAL PROPERTIES OF ESTIMATES 23 returns -0.8: Ten simulated samples of size T = 50 from the CER model Rt = 0.

16) T σ2 .9: Histogram of 1000 values of µ ˆ from Monte Carlo simulation of CER model. µ ˆ j values v u 1000 u 1 X t (ˆ µj − 0.10 = 0.01414. . (1.10 Estimate of mean Figure 1. If the number Notice that this value is very close to SE (ˆ µi ) = √ 50 of simulated sample goes to infinity then v u N N u 1 X 1 X j 2 j t lim (ˆ µ − µ ˆ ) = SE (ˆ µi ) N →∞ N − 1 j =1 N j =1 The Sampling Distribution of µ ˆi Using the results that pdf of µ ˆ i is normal with E [ˆ µi ] = µi and var(ˆ µi ) = Ti we may write µ ¶ σ2 i µ ˆ i ∼ N µi .08 0.24CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL 0 0.05045)2 = 0.01383 999 j =1 0.06 0.02 0.04 0.0 50 100 150 200 250 0.

with µi = 0 and σ 2 i = 1 for various sample sizes is illustrated in figure 1.3 STATISTICAL PROPERTIES OF ESTIMATES 25 2.5 1.5 2.0 -3 0. If the sample size is very large (as T → ∞) then var(ˆ µi ) will be approximately zero and the normal distribution of µ ˆ i given by (1.1. The distribution of µ ˆ i .5 pdf T=1 pdf T=10 pdf T=50 pdf 0. T The distribution for µ ˆ i is centered at the true value µi and the spread about the average depends on the magnitude of σ 2 i . In the statistics language we say that µ ˆ i is a consistent estimator of µi . if the sample size is very large then we essentially know the true value of µi . the uncertainty in µ ˆ i is larger for 2 larger values of σ i .0 -2 -1 0 estimate value 1 2 3 1 Figure 1.10: N (0. T .10. .0 1. and the sample size. For a fixed sample size. 10 and 50. √ ) density for T = 1. var(ˆ µi ) is smaller for larger sample sizes than for smaller sample sizes.16) will be essentially a spike at µi . . Notice how fast the distribution collapses at µi = 0 as T increases. Notice that the variance of µ ˆ i is inversely related to 2 the sample size T. This makes sense since we expect to have a more precise estimator when we have more data. the variability of Rit . In other words. Given σ i .

In this case α = 0. Suppose further that T − 1 = 60 (five years of monthly return data) so that tT −1 (α/2) = t60 (0. c (ˆ SE µi ) which can be rearranged as ³ ´ c (ˆ c (ˆ Pr µ ˆ i − tT −1 (α/2) · SE µi ) ≤ µi ≤ µ ˆ i + tT −1 · SE µi ) = 0. A confidence interval is an interval estimate of µi such that we can put an explicit probability statement about the likelihood that the confidence interval covers µi . µ ˆ i + tT −1 · SE µi )] = µ ˆ i ± tT −1 (α/2) · SE µi ) d(ˆ This resut follows from the fact that µ ˆ i is normally distributed and SE µi ) is equal to the square root of a chi-square random variable divided by its degrees of freedom. . The above result states that the standardized value of µ ˆ i has a Student-t distribution with T − 1 degrees of freedom6 . . RiT denote a random sample from the CER model.26CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL Confidence intervals for µi The precision of µ ˆ i is best communicated by computing a confidence interval for the unknown value of µi .025) = 2 and t60 (0. the interval covers the true unknown value of µi with probability 1 − α. Then µ ˆ i − µi ∼ tT −1 . To compute a (1 − α) · 100% confidence interval for µi we use (??) and the quantile (critical value) tT −1 (α/2) to give à ! µ ˆ i − µi Pr −tT −1 (α/2) ≤ ≤ tT −1 (α/2) = 1 − α. 6 . Result: Let Ri1 .95. For example. Then the 95% confidence for µi is given by c (ˆ µi ). µ ˆ i ± 2 · SE (1. .95. The construction of a confidence interval for µi is based on the following statistical result (see the appendix for details). Hence.005) = .17) c (ˆ c (ˆ c (ˆ [ˆ µi − tT −1 (α/2) · SE µi ).05 and 1 − α = 0. c (ˆ SE µi ) where tT −1 denotes a Student-t random variable with T − 1 degrees of freedom. suppose we want to compute 95% confidence intervals for µi . .

0201] With probability . RiT ) = σ ˆ2 i (Ri1 . .0050. . These confidence intervals are M SF T : 0. .5% and the upper limit is near 2%. σ ij and ρij .01068 = [0. 0. Similarly.0006.1. To determine the statistical properties of σ ˆ2 ˆ2 i and σ i we need to treat them as a functions of the random sample Ri1 . Example 13 Consider computing approximate 95% confidence intervals for µi using (1. This clearly shows that the mean return for SP500 is estimated much more precisely than the mean return for MSFT or SBUX. . Starbucks and S&P 500 data.2 Statistical properties of the method of moments estimators of σ 2 i . .003785 = [0. .01253 ± 2 · 0.02756 ± 2 · 0. to determine the statistical properties of σ ˆ ij and ρ ˆij we need to treat them as a . the above intervals will contain the true mean values assuming the CER model is valid. The widths are almost 5% with lower limits near 0 and upper limits near 5%. 0.17) based on the estimated results for the Microsoft. In contrast. The approximate 95% confidence intervals for MSFT and SBUX are fairly wide.0062. 1.3 STATISTICAL PROPERTIES OF ESTIMATES 27 The above formula for a 95% confidence interval is often used as a rule of thumb for computing an approximate 95% confidence interval for moderate sample sizes. . the 95% confidence interval for SP500 is about half the width of the MSFT or SBUX confidence interval. 0. T − 1 t=1 q σ ˆi = σ ˆ i (Ri1 . .3.95. .01359 = [0. . .0549] SP 500 : 0. The lower limit is near . σ i . RiT : σ ˆ2 i 1 X = = (Rit − µ ˆ i )2 . It is easy to remember and does not require the computation of quantile tT −1 (α/2) from the Student-t distribution. . RiT ). σ ˆ2 i (Ri1 .02777 ± 2 · 0. RiT ) T Note also that µ ˆ i is to be treated as a random variable. .0489] SBU X : 0.

. E [ˆ σ ij ] = σ ij . . σ ˆ i (Ri1 . σ ˆ ij and ρ ˆij are complicated and the The derivations of the variances of σ ˆ2 i. SE (ρij ) ≈ T 7 (1. 1 X (Rit − µ ˆ i )(Rjt − µ ˆ j ). RjT ) = Bias Assuming that returns are generated by the CER model (1. .σ 7 if the sample size. RjT : σ ˆ ij = σ ˆ ij (Ri1 . . T − 1 t=1 T σ ˆ ij (Ri1 . However.28CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL functions of Ri1 . RiT . T. . . . The proofs of these results are beyond the scope of this book. RjT ) = ρ ˆij = σ ˆ ij (Ri1 . RjT ) . . . . there are ˆ i and ˆ ρij that are valid simple approximate formulas for the variances of σ ˆ2 i. . . . .σ exact results are extremely messy and hard to work with. 2T ) (1 − ρ2 √ ij .19) (1. E [ˆ σi] = 6 σi .18) (1. . i T /2 σi SE (ˆ σi) ≈ √ . . RiT ) · σ ˆ j (Rj 1 . These large sample approximate formulas are given by σ2 i p SE (ˆ σ2 ) ≈ . . . . However. . . . . RiT and Ri1 . . . is reasonably large . .20) The large sample approximate formula for the variance of σ ˆ ij is too messy to work with so we omit it here. . . . . Rj 1 . RiT . 2 E [ˆ σ2 i ] = σi . . . . Rj 1 . . RjT ) but the sample standard deviations and correlations are biased estimators. RiT .1). . they may be easily be evaluated using Monte Carlo methods. E [ˆ ρij ] = 6 ρij . . . . . the sample variances and covariances are unbiased estimators. Rj 1 . Precision ˆ i.

however. SE (ˆ σ i ) goes to zero the fastest and SE (ˆ i ) goes to zero the slowest.σ estimated standard errors σ ˆ2 i c (ˆ p SE σ2 ) ≈ . for a fixed sample size.23) Example 14 To be completed Sampling distribution To be completed Confidence Intervals for σ 2 i . SE T (1. Practically useful formulas 2 replace σ i . the formulas for the standard errors above are inversely related to the square root of the σ2 sample size. Interestingly. σ i and ρij . it appears that σ i is generally estimated more precisely than σ 2 i and ρij . σ i and ρij Approximate 95% confidence intervals for σ 2 i . σ i and ρij are give by σ ˆ2 i 2 2 c p σ ˆ2 ± 2 · SE (ˆ σ ) = σ ˆ ± 2 · .1. 2T (1 − ˆ ρ2 ) c (ρij ) ≈ √ ij . i T /2 σ ˆi c (ˆ SE σi) ≈ √ . i i i T /2 σ ˆi c (ˆ σ ˆ i ± 2 · SE σi ) = σ ˆi ± 2 · √ 2T (1 − ˆ ρ2 ij ) c (ˆ ρij ± 2 · SE ˆ ρij ) = ρ ˆij ± 2 · √ T Example 15 To be completed .21) (1.3 STATISTICAL PROPERTIES OF ESTIMATES 29 where “≈” denotes approximately equal. Hence. As with the formula for the standard error of the sample mean. The above formulas are not practically useful. The approximations are such that the approximation error goes to zero as the sample size T gets very large. because they depend on the unknown quantities σ 2 i . and ρij is estimated generally more 2 precisely than σ i .22) (1. σ i and ρij by the estimates σ ˆ2 ˆ i and ρ ˆij and give rise to the i.

We use the simulation i and σ model (1.016 0 50 100 150 200 0. . Evaluating the Statistical Properties of σ ˆ2 ˆi.11: Histograms of σ ˆ 2 and σ ˆ computed from N = 1000 Monte Carlo samples from CER model. σ ˆ ij and ˆ ρij by Monte i.08 0. Consider first the variability estimates σ ˆ2 ˆ i .. The histogram for the σ ˆ values is more symmetric and is centered near 0.30CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL 200 150 100 0 50 0.010 = σ 2 .15) and N = 1000 simulated samples of size T = 50 to compute the ¡ 2 ¢1 ¡ 2 ¢1000 estimates { σ ˆ .006 0. .The histogram for the σ ˆ 2 values is bellshaped and slightly right skewed but is centered very close to 0..10 = σ . σ Carlo simulation ˆi..012 0.12 Estimate of variance Figure 1. .. The histograms of these values are displayed in figure 1. σ ˆ } and {σ ˆ 1.010 0.11.008 0. . deviation 0.10 Estimate of std. σ ˆ ij and ρ ˆij by Monte We may evaluate the statistical properties of σ ˆ2 i. . σ ˆ 1000 }.004 0.014 0. σ Carlo simulation in the same way that we evaluated the statistical properties of µ ˆ i .

002 50/2 0. Using the formulas (1.009952 1000 j =1 1 X σ ˆ = 0.10)2 p SE (ˆ σ ) = = 0.10 SE (ˆ σ) = √ = 0.010 2 · 50 2 1.1.4 FURTHER READING The average values of σ 2 and σ from the 1000 simulations are 31 1 X 2 σ ˆ = 0.5.4 Further Reading To be completed 1.18) and (1.1 Appendix Proofs of Some Technical Results Result: E [ˆ µi ] = µi .09928 1000 j =1 1000 1000 The sample standard deviation values of the Monte Carlo estimates of σ 2 and σ give approximations to SE (ˆ σ 2 ) and SE (ˆ σ ).19) these values are (0.5 1.

T = . Using the fact that µ ˆ i = T −1 # " T 1X E [ˆ µi ] = E Rit T t=1 " # T 1X = E (µ + εit ) T t=1 i PT t=1 Rit and Rit = µi + εit we have T T 1X 1X = µ + E [εit ] (by the linearity of E [·]) T t=1 i T t=1 T 1X = µ (since E [εit ] = 0.32CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL Proof. . P Proof. . . . . t = 1. Using the fact that µ ˆ i = T −1 T t=1 Rit and Rit = µi + εit we have à ! T 1X var(ˆ µi ) = var Rit T t=1 à ! T X 1 = var (µ + εit ) (in the CER model Rit = µi + εit ) T t=1 i à ! T 1X εit = var (since µi is a constant) T t=1 T 1 X = 2 var(εit ) (since εit is independent over time) T t=1 T 1 X 2 = 2 σ T t=1 i (since var(εit ) = σ 2 i . T ) 1 T σ2 i 2 T σ2 = i. . . . t = 1. T ) T t=1 i 1 = T · µi T = µi . σ2 Result: var(µi ) = Ti .

Z ∼ N (0. The pdf of X is illustrated in Figure xxx for various values of T. . X ∼ χ2 T .5. Notice that X is only allowed to take non-negative values.34.05 then χ2 5 (0.07. Let χ2 T (α) denote this critical 8 value . N (0. . Furthermore.2 Some Special Probability Distributions Used in Statistical Inference The Chi-Square distribution with T degrees of freedom Let Z1 . 1. 1).5 APPENDIX 33 1. The chi-square distribution is used often in statistical inference and probabilities associated with chi-square random variables are needed.5. Then X is a chi-square random variable with T degrees of freedom. Then Pr(X > χ2 T (α)) = α. Assume 8 Excel has functions for computing probabilities from the chi-square distribution. i = 1. Such a 2 random variable is often denoted χ2 T and we use the notation X ∼ χT . . it can be shown that E [X ] = T. . which are just quantiles of the chi-square distribution. suppose we wish to find the critical value of the chi-square distribution with T degrees of freedom such that the probability to the right of the critical value is α.05) = 124.3 Student’s t distribution with T degrees of freedom Let Z be a standard normal random variable. if T = 5 and α = 0. and let X be a chi-square random variable with T degrees of freedom. That is. To illustrate. ZT be independent standard normal random variables.1. 1). . Z2 . . are used in typical calculations. Critical values. For example.i. Zi ∼ i. The pdf is highly right skewed for small values of T and becomes symmetric as T gets large. T. Define a new random variable X such that X= 2 Z1 T X i=1 + 2 Z2 + 2 · · · ZT = Zi2 . .05) = 11.d. . . if T = 100 then 2 χ100 (0.

Notice that the pdf is symmetric about zero and has a bell shape like the normal.00.95. Let tT (α) denote the critical value such that Pr(t > tT (α)) = α. we have that Pr(−tT (α) ≤ t ≤ tT (α)) = 1 − 2α. 1. As T gets large the tails shrink and become close to the normal.6 Problems To be completed .025) = 2.025) = 0.34CHAPTER 1 THE CONSTANT EXPECTED RETURN MODEL that Z and X are independent.025 then t10 (0. if T = 60 and α = 2 then t60 (0. the tails are quite spread out and are thicker than the tails of the normal. if T = 10 and α = 0. For example.025) = 2.025) = 2 and Pr(−t60 (0. Define a new random variable t such that t= p Z . X/T Then t is a Student’s t random variable with T degrees of freedom and we use the notation t ∼ tT to indicate that t is distributed Student-t. For small values of T .025) ≤ t ≤ t60 (0.228. Figure xxx shows the pdf of t for various values of the degrees of freedom T. as T → ∞ the pdf of the Student t converges to the pdf of the normal. The tail thickness of the pdf is determined by the degrees of freedom. The Student-t distribution is used heavily in statistical inference and critical values from the distribution are often needed. For example. In fact.025)) = Pr(−2 ≤ t ≤ 2) = 1 − 2(0. Since the Student-t distribution is symmetric about zero. if T = 100 then t60 (0.

35 . Princeton University Press. Princeton.Bibliography [1] Campbell. Lo and MacKinley (1998). NJ. The Econometrics of Financial Markets.

Sign up to vote on this title
UsefulNot useful