Investments - Applications and Tests

INVESTMENTS Applications and Tests Predictions and Applications • Predictions: – CAPM: In market equilibrium, investors are only rewarded for bearing the market risk. – APT: In the absence of arbitrage, investors are only rewarded for bearing the factor risk. • Applications: – professional portfolio managers: evaluating security returns and fund performance. – regulatory commissions: cost of capital for regulated firms. – court rulings: evaluating claims of lost future income. – corporate manager: capital budgeting decisions. Testability of CAPM and APT The wide acceptance of the CAPM and APT makes it all the more im­ portant to test their predictions empirically. Recall from Class 6, both theories build on assumptions that are un­ realistic at times. How does a product of abstract reasoning hold in reality? Unfortunately, the predictions of the CAMP and APT are hard to test empirically: • Neither the market portfolio in CAPM nor the risk factor in APT is observable. • Expected returns are unobservable, and could be time varying. • Volatility is not directly observable, and is time varying. An Ideal Test of the CAPM In an ideal situation, we have the following inputs: 1. Riskfree borrowing/ lending rate rf 2. Expected returns on the market E(rM ) and on the risky asset E(ri ). 3. The exposure to market risk βi = cov (rM , ri ) var (rM ) (1) These inputs allow us to examine the relation between reward (E(rf ) − rf ) and risk βi 1. More risk, more reward? 2. Do they line up? 3. What is the reward for a risk exposure of 1? 4. Zero risk, zero reward? E(r) A Linear Relation between Risk and Reward α E(rM)-rf rf 0 0 βM Figure 1: Beta with and without risk-free interest rate. Beta (βι ) Some Practical Compromises the market portfolio rM is unobservable: use a proxy, e.g., the S&P 500 index. expected returns E(rM ) and E(rf ) are unobservable: use sample av­ erage T 1 � = rM,t N t µM T 1 � µi = ri,t N t (2) unobservable risk exposure βi = cov (rM , ri ) var (ri ) (3) use sample estimates: βi = where cov � (rM , ri ) � (ri ) var (4) T 1 � var � (ri ) = (rt,M − µM ) N t (5) T 1 � cov (rt,M − µM ) (rt,i − µi ) � (rM , ri ) = N t (6) Testing the Linear Relation pick a proxy for the market portfolio rM , and record N monthly returns: rt,M : i = 1, . . . , N (7) for the same sample period, collect a sample of I firms, each with N monthly returns: rt,1 : i = 1, . . . , I and t = 1, . . . , N (8) construct the sample mean for rM . for the i-th firm, construct the sample mean µi , and the sample es­ timate βi . for i = 1,. . . ,I test the linear relation: µi − r f = γ 0 + γ 1 β i (9) Implications of the CAPM γ0 = 0: zero exposure → zero reward. γ1 = µM − rf : one unit of exposure same reward as the market. 0.7 SML 0.4 0.5 0.6 QCOM Historical Return Forecasted Return 0.1 0.2 0.3 VRTS NT AMGN AA AOL HD 0.0 FDC -0.1 PFE CAH BAC UN MO -0.3 WMI 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 Beta Figure 2: Betas for top 100 market weighted stocks from SP 500 against SP 500. 2.4 2.6 Regression: The Basic Setup Two variables x and y, N pairs of outcomes: (xi , yj ), i = 1, 2, . . . , N. (10) We have reasons to believe that y and x are related. In particular, we would like to use x to explain y: yi = a + bxj + ε. y: the dependent variable. x: the independent (explanatory) variable. εi : a random disturbance with zero mean. coefficients: intercept a, slope b. (11) Regression: Motivation Some motivating examples: 1. On day i, xi is the temperature at Orlando, yi is the price of the futures contract on frozen concentrated orange juice. 2. For firm i, xi is its leverage ratio, yi is its probability of default. 3. On day i, xi is the Fed fund target rate, yi is the 3 month T bill rate. 4. At the i-th second, xi is the number of packets sent by rgallati@mit.edu to jcox@mit.edu, yi is the number of packets received by the latter. In each case, the outcome of x might contribute to the outcome of y: yi = a + b · x i + ε i (12) but there might be other random factors, captured by εi , that have noth­ ing to do with x. The slope coefficient b is of particular interest as it measures the sensitivity of y to x. The Regression Coefficients The objective: find a and b that best capture the linear relation between y and x. How: find a and b that minimize the squared differences: min a∈R,b∈R N � (yi − a − b · xi )2 (13) i=1 o E(µi)-rf o b / βi o o o o o 0 0 Figure 3: Regression. X / βi Beta (β) Regression: The Solution Solving for the optimization problem, we get: • an estimate �b for the slope coefficient b: �N (yi − µ �y ) (xi − µ �x ) �b = i=1� N �x ) i=1 (xi − µ (14) a for the intercept a: � a = µy − �bµx • an estimate � Notice the familiar notation for sample means: T 1 � xi µ �x = N t T 1 � µ yi �y = N t (15) Why do we call our solutions estimates? Why put hats on b and a? a hat is close to the Do we always get �b that is close to the real b, � real a? In the large sample (large N), we are pretty confident that they do, why? The Standard Error In Order to guess the real values of b and a, we use the data to help us. Given N pairs of observations (yi , xi ), our regression solutions �b and a are the best guesses. But we can never be 100% sure. � How to quantify our uncertainty about �b and � a. We think of �b and � a as random variables. For any estimate, say �b, we can get an estimate of its standard deviation, which is usually referred to as the Standard Error. The standard error of an estimate is one measure of its precision. Interpreting the Regression Result For the purpose of this class, you will use a ”canned” regression package (e.g., Excel): Input: (yi , xi ), i = 1, . . . , N Output: • the estimates �b and � a and • their standard errors: s�b and s�a • their t statistics: t�b = �b/s�b and t�a = � a/s�a • the R squared The standard errors and the t stat’s provide measures of precision of your estimates. The R squared tells you how much of the randomness in the depen­ dent variable y is explained by the explanatory variable x. More on the Slope Estimate Recall that our estimate for the slope coefficient b: �N (yi − µ �y ) (xi − µ �x ) �b = i=1� N �x ) i=1 (xi − µ (16) Some familiar notation: T 1 � (rt,M − µM ) � (ri ) = var N t (17) T 1 � (rt,M − µM ) (rt,i − µi ) cov � (rM , ri ) = N t (18) We have βi = cov � (rM , ri ) � (ri ) var (19) Intuitively, b is a measure of the covariance cov(x,y) between x and y, scaled by the variance var(x) of x. Back to testing the CAPM The implication of the CAPM: µ �i − rf = γ0 + γ1 βi (20) � on the 43 industrial portfolios tell us that this relation Our data (� µ, β) does not hold exactly. One possibility: our measures of the expected returns are contaminated by noises that are unrelated to the β’s. What we still like to know: • On average, is reward related to risk at all? γ1 = 0 or not ? • On average, does zero risk result in zero reward? γ0 = 0 or not? • On average, does one unit of risk exposure pay the market return? γ1 = µM − rf = 5.9% or not? (21) Regression in Action Set up a regression: • The dependent variable: yi = µ �i − rf • The independent variable: xi = β�i • Add noise εi that is unrelated to βi . Feed the data to the regression package: estimate standard error t-stat γ0 6% 1.8% 3.5 γ1 0.17% 1.7% 0,1 R squared = 2% Recall the implications of the CAPM: 1. The intercept γ0 = 0 2. The slope γ1 = µ �M − rf = rf 0.7 SML Historical Return Forecasted Return 0.1 0.2 0.3 0.4 0.5 0.6 QCOM VRTS ORCL NT AMGN AA AOL CMCSK HD DOW SUNW EMC 0.0 IP -0.1 BK BAC ONE UN SWY WB -0.3 LMT MO WMI 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 Beta Figure 4: Betas (without rf ) for top 100 market weighted stocks from S&P 500 against S&P 500. Historical Return Forecasted Return 70% 60% 50% 40% 30% 20% 10% 0% -10% -20% -30% 0.00 0.50 1.00 1.50 2.00 2.50 3.00 Beta Figure 5: Betas (with and without rf ) for top 100 market weighted stocks from S&P 500 against S&P 500. A Rule of Thumb with t-Stat To get a sense of how an estimate, say �b, differs significantly from zero, its t-stat � tb is the most telling statistics. tb as standard normal (not a bad assumption A Rule of Thumb: think of � for a large sample). The larger the magnitude (absolute value) of b, the more likely it is significantly different from zero. � 0. Hypothesis Testing: The null: �b = 0, the alternative: �b = 1. A t-stat of 1.960 rejects the null with significance level 5%; 2. A t-stat of 2.576 rejects the null with significance level 1%; tγ1 = 0.1, what can we say about γ �1 ? What about γ �1 ? For example,� A Summary of the CAPM Tests In general, the test results depend on the sample data, sample peri­ ods, statistical approaches, proxy for the market portfolio, etc. But the following findings remain robust: • The relation between risk and reward is much flatter than that pre­ �1 = µ �i − rf . dicted by the CAPM γ • The risk measure β cannot even begin to explain the cross sectional variation in the expected returns. (γ �1 is statistically insignificant, R squared is close to zero.) • Contrary to the prediction of the CAPM, the intercept γ �0 is signifi­ cantly different from zero. Some Possible Explanations 1. Is the stock market index a good proxy for the market portfolio? • only 1/3 non governmental tangible assets are owned by the corpo­ rate sector. • among the corporate assets, only 1/3 is financed by equity • what about intangible assets, like human capital? • what about international markets? 2. Measurement error in β: • Except for the market portfolio, we never observe the true β. • To test the CAPM, we use estimates for β, which are measured with errors. • The measurement error in β will cause a downward biased estimate for the slope coefficient, and an upward biased estimate for the in­ tercept. 3. Measurement error in expected returns �M and µ �i as proxies for the real, unob-servable • we use sample means µ expected returns • it is known that means are hard to estimate, and there are noises in �M and µ �i and our estimates µ �M and µ �i are correlated, then we have a statistical • if the noises in µ problem (errors in variables) 4. Borrowing restrictions • This class covers only one version of the CAPM, assuming unre­ stricted borrowing. • In practice, borrowing restrictions are realistic. It includes margin rules, bankruptcy laws that limit lender access to a borrower’s future income, etc. • Fisher Black showed that borrowing restrictions might cause low-β stocks to have higher expected returns than the CAPM predicts.