Sei sulla pagina 1di 6

11/10/11

Predicting Betas: Correcting for Mean Reversion


In order to correct for mean reversion
Because the average beta is 1

Estimating beta for illiquid stocks


Assume that the stock is not traded in dates t1, t2, What would be the prices in the non-traded days?
Same as in the last trading day, thus zero return
Large return in the first trading day after the break

We calculate the following:

Predicted by the market model: Rt = *RM,t

In the first trading day after the break, the stock s return will accumulate all idiosyncratic noise over the non-traded period

Adjusted Beta =

2 1 HistoricBeta + 1 3 3

Betas will be biased! Dimson (1979): regression including lead and lag values of the market index Rj,t = j + l=-l1:l2j,lRM,t+l + j,t
True beta is a sum of all lead-lag betas: l=-l1:l2j,l

2008 Patrick J. Kelly

97

Estimating beta for illiquid stocks


The trade-to-trade approach:
Compute stock returns from the last traded day to the next traded day (if necessary - over 2, 3 or nt days) Compute market returns for the same periods Run a regression with matched multi-period returns:

What if the stock has a large weight in the index?

Endogeneity problem
In the extreme, when the index is dominated by one stock, this stock will have beta of 1 by construction In Russia, this is a problem for Gazprom, Lukoil,

Rj,nt = j,nt + jRM,nt + t=0:nt-1j,t


How to control for heteroscedasticity?
WLS: the variances are proportional to nt OLS regression with data divided by nt

How to solve it?


Usual way: exclude this stock from the index Theoretical solution: use IV approach with other blue chips (or industry indices) as instruments

(Rj,nt / nt) = j*nt + j(RM,nt /nt) + (t=0:nt-1j,t /nt)

CAPM predictions
should be zero.
If not, there may be missing factors.

CAPM
The evidence

is the only relevant factor Relation between and returns is linear Over long periods the return on the market is greater than the risk free return
In general, riskier stocks should earn higher return on average

Market portfolio is mean-variance efficient If you cross-sectionally regress on risk premia, estimates should equal the average market risk premium ()
101
2011 Patrick J. Kelly 102

2011 Patrick J. Kelly

11/10/11

We observe the series of returns of the asset and of the market index We estimate beta (in-sample) and check whether other variables (const, D/P, P/E, B/M, ) explain remaining time variation in the asset s return

Cochrane (2001, Chapters 2 and 5) shows that any expected return can be related to any mean-variance efficient portfolio lying on the efficient frontier:

Data: US, 1926-1982, monthly returns of 11 industry portfolios and VW-CRSP market index For each individual portfolio, standard CAPM is not rejected But the joint test rejects CAPM

Data: US, 1965-1994, monthly returns of 10 size portfolios, VWCRSP market index Joint test rejects CAPM, esp. in the earlier part of the sample period

We check whether observed betas are (linearly) related to stock returns (as in the Security Market Line)

This structure suggests a natural test for efficiency:

This is the basic intuition of most tests of portfolio efficiency

Intuition: Tests of Mean Variance Efficiency

Approach 2: cross-sectional regression

Gibbons, Ross, and Shanken (1989)

Approach 1: time series regression

Did not reject CAPM But: betas are unstable over time

Testing CAPM

Results

Test whether is non-zero.

CLM (1997, Table 5.3)

Early tests, up to 80s

Differences across tests are mostly econometric refinements

E ( ri ) ! rf = ! + "i,mv " E ( r mv ) ! rf $ # %

E ( ri ) = rf + !i,mv " E ( r mv ) ! rf $ # %

434 The Journal of Finance

103

0.56 0.94 1.02 0.98 1.11 1.14 0.59 0.76 1.08 0.76 1.42

of the the a June year, The preME-7 ME-4 ME-8 ME-5 ME-2 All ME-9 ME-6 ME-3 decile The The The returns of average is and Portfolios Small-ME Large-ME Alllargest with current calculated (as average portfolio year 7 forrequirementsare ME t average equal-weighted is and to Size number (t size July subdivided the Average column of return post-ranking = 0.95 1.07 1.29 1.24 1.52 All 0.89 1.10 1.17 1.25 1.29 1.25 post-ranking CRSP-COMPUSTAT is prior of and available) formed fls number into deciles portfolios the Os stocks establish shows of year 10 data of is for denominateduse t yearly. then month's the f (here in time-seriestheto 3 ending 1963-1990) Returns, the ,: 1.01 1.09 1.08 1.27 1.25 1.34 stocks 0.98 0.95 1.34 1.12 1.71 Low-f the in The full and between statistics per time-seriesJune stocks size-fmillionsreturns of June using averagein 11 portfolios requirements for on all (July all Stocks fl-2 in of of of 0.93 1.05 1.53 1.13 1.42 1.29Panel 0.88 1.21 1.42 1.31 1.57 and month breakpoints. are the average year using breakpoints t A: each 22. for 1963 other year NYSE + : portfolios of forJuly dollars. t. monthly to 1. The Post-Ranking fl-3 the in the Sorted 1.02 1.26 1.39 1.17 1.79 1.10 1.37 1.27 1.54 1.36 1.36 the tables) We allocated onO3s Average group. size 1963 stocks equal-weighted size-: to pre-ranking on size to are use averages December monthly value-weighted ,Bsthe and fl-4 ME 1.14 1.09 1.48 1.70 1.61 0.94 1.20 1.15 1.06 1.39 1.31 the Table deciles of (ME, Monthly of only I equal-weighted 10 CRSP. 2 1990) portfolios sum size-decile portfolio All and size price f-5 ln(ME) of 1.07 1.18 1.42 1.29 1.50 0.93 1.27 1.20 1.34 1.65 1.33 3 in (Down) NYSE of Returns Average December for monthly the sample individual (ME)is the times (in equal-weighted of NYSE, then stocks f-6 NYSE, 1990 Size portfolios 0.89 0.98 1.21 1.06 1.61 1.28 1.23 1.11 1.18 1.10 1.50 stocks slopes returns stocks, shares betweenin smallest that AMEX, For Percent) portfolios. portfolio 15 from on using the AMEX, size a fl-7 the 1.03 1.18 1.18 1.41 1.37 1.24 0.94 1.24 1.13 1.31 1.37 meet post-ranking theand The and and estimated 41,decile returns, the All portfolio outstanding) in Pre-Ranking NYSE fl-8 and at (afterreturns with 0.71 1.02 1.04 1.17 1.31 1.21 0.82 0.62 1.27 1.36 1.63 ( Portfolios resulting regression 2 NASDAQ row varies the of for 100 to the deciles 1972) 5 percent. end fl-9 are stocks each shows from 0.74 1.01 1.07 1.35 1.34 1.25 0.88 1.32 1.18 1.26 1.50 70 averageof The monthly years breakpoints. Formed (Across): to portfolios that of June on NASDAQ CRSP-COMPUSTAT are portfolio. Each 177. of statistics meet number average returns determined the data size in for forThe each stocks. then monthly size on The

There are no systematic deviations from the return predicted by the CAPM

Related findings- Fama French 1992 - Size and Beta

Run CAPM regressions using 342 months of data for 25 size and book-to-market portfolios

Run time series regression for N assets: Ri,t-RF = i + i(RM,t-RF) + i,t (+ iXi,t-1) Usually, N portfolios

Grouped by industry / size / To reduce idiosyncratic variation present in individual stocks

Assuming that Ri,t ~ IID Normal, estimate by OLS

F-test for the joint significance of and coefficients

2011 Patrick J. Kelly

Fama-French (1993)

Time series test

Strongly reject CAPM. F=1.91 with p-value 0.004

H0: i=0 (i=0) for any i=1,,N

High-fl

2010 Patrick J. Kelly

107

2011 Patrick J. Kelly

108

11/10/11

Test of Mean Variance Efficiency


Gibbons, Ross, Shenken (1989)
2+ ( T ! N !1 * " ET ( f ) % $ ' 1+ ! ./!1! ~ FN,T !N !1 * $ !(f) ' N & , ) # !1

Cross-sectional tests
Main idea: Ri,t = 0 + 1i + i,t (+2Xi,t) H0: asset returns lie on the SML

Where f is a return based factor or portfolio return on the meanvariance efficient frontier, ET(f) is the sample mean of the factor, and (f) the sample standard deviation, T is the number of observations, N is the number of test assets, 1 is the number of factors, is a vector of the intercepts from the N test-asset regressions, and is the cross test asset residual covariance matrix, such that E !t!t! = "

We don t know true betas!

The intercept is the risk-free rate: 0 = RF The slope is the market risk premium: 1 = mean(RM-RF) > 0 There are no additional effects: 2 = 0 We first need to estimate them measurement error Usually, use predicted betas (estimated over the previous period)

The realized returns may be very different from the expected returns (in SML)
We need to estimate this cross-sectional regression over many periods (to measure average effect)

Mean of
2011 Patrick J. Kelly 109

Two-stage procedure
Recursive procedure: for each t (e.g., every month), repeat
Time-series regression over the previous K periods to estimate beta

Results
Fama-MacBeth (1973): estimate CS regression Ri,t-RF = 0 + 1i + 22i + 32,i + i,t
1 > 0: positive effect of beta 2 = 0: no non-linear effects of beta 3 = 0: no impact of idiosyncratic risk

Ri,t-k = i + iRM,t-k + i,t-k, for k=1:K


Cross-sectional (CS) regression Ri,t-RF = 0 + 1i + i,t (+ 22i + 3i)

Fama-MacBeth approach (first in 1973)


Running CS regression for each month t, we get the time series of coefficients 0,t, 1,t, Compute mean and st. deviation of s from these time series:
No need for s.e. of coefficients in the cross-sectional regressions! Shanken s correction for the estimation error in betas

Cochrane (2001)
Sort all NYSE stocks into 10 size deciles, add government and corporate bonds Run separate time-series regressions to estimate 12 portfolio betas Regress sample average portfolio returns against estimated betas Compare fitted SML with the one predicted by the CAPM

Assuming normal IID returns, t-statistic = T avg / ()

CAPM does well in explaining stock vs. bond returns, but poorly in explaining large vs. small-cap stocks

Cross-Sectional Approach: Fama-McBeth (1973)


Tests if high beta is associated with high returns and vice versa. General format:
Time-series regressions to get betas for test portfolios
Usually beta sorted porttflios

Monthly cross-sectional regressions to test if high beta is associated with high return

Cochrane (1999) shows that GMM panel regressions are identical under some assumptions

2011 Patrick J. Kelly

114

11/10/11

General FM73 Algorithm


Use monthly data from years 1 through 7 to estimate CAPM Beta1-7 for each company. Use to rank stocks (rank1-7) into 20 equally sized groups (call portfolios formed on rank1-7 portfolio1-7) Use monthly data from years 8 through 12 to estimate CAPM Beta8-12 (= ) for each company. Find the average Beta8-12 and average (equally weighted) return for each Portfolio1-7 in each month of year 13 (i.e. letting delisted companies to drop out each month). Roll one year forward and repeat steps 1-4 till done. Example in step 1 use data from years 2 through 8 For each month run a cross-sectional regression using the average Betas and returns calculated in step 4. Collect the time series of the coefficients (s)

Example from FF92


They look at whether CAPM works.
It doesnt.

Here is what they do:


The sort stocks by size and by the stocks pre-ranking beta.
Pre-ranking betas: betas are calculated for each stock over the 5 years of monthly data preceding (not including) the current year (from t-5 to t-1). Stocks are sorted by size and then each size portfolio is sorted by the preranking beta. For each month in year t, calculate equally weighted portfolio returns For each size/pre-ranking-beta portfolio calculate Beta for the entire sample, including one lag of the value weighted market to correct for non-synchronous trading. Assign the betas to each stock in the portfolio and run monthly cross-sectional regressions.
115 2011 Patrick J. Kelly 116

2011 Patrick J. Kelly

Average Slopes (t-Statistics) from Month-by-Month Regressions of Stock Returns on ,B,Size, Book-to-Market Equity, Leverage, and E/P: July 1963 to December 1990

A word of caution about tests: Rolls Critique (1977)


If a portfolio against which returns are measured is ex-post efficient, then no security will have abnormal returns If the portfolio is inefficient then any abnormal return is possible
If alpha is non-zero is it because the market portfolio is inefficient or is it because there is mispricing?

If returns are linear in Beta, all that proves is that the market is ex-post efficient
If they are not linear in beta you do not know if it is because the market is
ex-post inefficient or because there is a missing factor

2011 Patrick J. Kelly

117

2011 Patrick J. Kelly

118

Rolls Critique (1977) continued


Only true test if the true market portfolio is ex-post efficient
True market portfolios is unobservable. So, good luck.

This critique implies


Time-series tests of efficiency are the better tests to show a model might work Cross-sectional tests are better to show that the model might not work

ICAPM and APT


The evidence

2011 Patrick J. Kelly


2011 Patrick J. Kelly 119

120

11/10/11

What are the factors


Neither APT nor ICAPM are too clear about possible factors
ICAPM: factors forecast changes in the investment opportunity set APT: factors are tradable portfolios that covary with undiversifiable sources of risk.

Chen, Roll, Ross (1986): Testing APT

or is it ICAPM?

In the US: Unexpected Inflation Term structure (Long government bonds short) Default premium (High risk corporate bonds long government bonds) Growth in Industrial production Consumption Growth Oil prices

2011 Patrick J. Kelly

121

2011 Patrick J. Kelly

122

APT only test: Conner and Korajczyk (1993)


Factor Analysis of stock returns
Find evidence of 1 to as many as 6 factors

Data Driven Risk Factors


Research has identified portfolios of stock that appear to get extra high returns:
Small firms out perform large firms
at least they did 1926 through 1980

More recent work provides evidence of both market-wide and exchange-specific factors [Goyal, Perignon, and Villa (2008)]
In US 2 factors common to all stock, plus unique (separate) factors of NASDAQ and NYSE.

Value stocks out perform Glamour/Growth stocks

Fama and French developed a three factor model that includes:


Market Size factor (SMB) Value/Growth factor (HML)

446
2011 Patrick J. Kelly

The Journal of Finance


Table V
123 2008 Patrick J. Kelly 124

Average Monthly Returns on Portfolios Formed on Size and Book-to-Market Equity; Stocks Sorted by ME (Down) and then BE/ME (Across): July 1963 to December 1990
In June of each year t, the NYSE, AMEX, and NASDAQ stocks that meet the CRSPCOMPUSTAT data requirements are allocated to 10 size portfolios using the NYSE size (ME) breakpoints. The NYSE, AMEX, and NASDAQ stocks in each size decile are then sorted into 10 BE/ME portfolios using the book-to-market ratios for year t - 1. BE/ME is the book value of common equity plus balance-sheet deferred taxes for fiscal year t - 1, over market equity for December of year t - 1. The equal-weighted monthly portfolio returns are then calculated for July of year t to June of year t + 1. Average monthly return is the time-series average of the monthly equal-weighted portfolio returns (in percent). The All column shows average returns for equal-weighted size decile portfolios. The All row shows average returns for equal-weighted portfolios of the stocks in each BE/ME group.

Size and Book to Market

Fama-French Three Factor Model


E[ri ] rt = M ,i (E[rM ] rt ) + SMB ,i SMB + HML ,i HML
Market Risk Premium SMB is a portfolio long Small stocks and short large or Big stocks and is called SMB for Small Minus Big HML is a portfolio long Value stocks and short Growth stocks
(value stocks are called High book to market stocks, by finance professor types and Growth stocks are called Low book to market stocks. So, HML is High Minus Low.
2008 Patrick J. Kelly 126

Book-to-Market Portfolios All All Small-ME ME-2 ME-3 ME-4 ME-5 ME-6 ME-7 ME-8 ME-9 Large-ME 1.23 1.47 1.22 1.22 1.19 1.24 1.15 1.07 1.08 0.95 0.89 Low 0.64 0.70 0.43 0.56 0.39 0.88 0.70 0.95 0.66 0.44 0.93 2 0.98 1.14 1.05 0.88 0.72 0.65 0.98 1.00 1.13 0.89 0.88 3 1.06 1.20 0.96 1.23 1.06 1.08 1.14 0.99 0.91 0.92 0.84 4 1.17 1.43 1.19 0.95 1.36 1.47 1.23 0.83 0.95 1.00 0.71 5 1.24 1.56 1.33 1.36 1.13 1.13 0.94 0.99 0.99 1.05 0.79 6 1.26 1.51 1.19 1.30 1.21 1.43 1.27 1.13 1.01 0.93 0.83 7 1.39 1.70 1.58 1.30 1.34 1.44 1.19 0.99 1.15 0.82 0.81 8 1.40 1.71 1.28 1.40 1.59 1.26 1.19 1.16 1.05 1.11 0.96 9 1.50 1.82 1.43 1.54 1.51 1.52 1.24 1.10 1.29 1.04 0.97 High 1.63 1.92 1.79 1.60 1.47 1.49 1.50 1.47 1.55 1.22 1.18

controlling for size, book-to-market equity captures strong variation in average returns, and controlling for book-to-market equity leaves a size effect in returns. averageKelly 2011 Patrick J. 125 B. The Interaction between Size and Book-to-Market Equity The average of the monthly correlations between the cross-sections of ln(ME) and ln(BE/ME) for individual stocks is - 0.26. The negative correlation is also apparent in the average values of ln(ME) and ln(BE/ME) for the portfolios sorted on ME or BE/ME in Tables II and IV. Thus, firms with low market equity are more likely to have poor prospects, resulting in low stock prices and high book-to-market equity. Conversely, large stocks are more likely to be firms with stronger prospects, higher stock prices, lower book-tomarket equity, and lower average stock returns. The correlation between size and book-to-market equity affects the regressions in Table III. Including ln(BE/ME) moves the average slope on ln(ME) from -0.15 (t = -2.58) in the univariate regressions to -0.11 (t = -1.99) in the bivariate regressions. Similarly, including ln(ME) in the regressions

11/10/11

Do the Fama-French Factors Work in the US?


Table Ya lntrrceprs from excess stuck return regressions for 25 stuck purtfulios formed on size and book-to-market 342 months. equity: July 1963 IO December IYYI,

Example with the Fama-French 3 Factor Model


2.52 2.61 2.66 2.51 4 2.20 2.49 2.56 2.61 2.85 High 1.70

(iii)
S I l l; L l l

X(r) -

Size 4 quinlile Big

2 3

0.24 0.52 0.52 0.69 Low 0.76

0.46 0.58 0.61 0.39 2 0.52 (iv)

0.49 0.64 t, 0.52 0.50 3 0.43

RF(I) = u + sSMtqr) + hlfML&) + t(r) Book-to-market equity (HE/ME) quintiles I.92 0.53 0.55 0.97 0.58 0.64 2.00 2.40 0.60 0.66 2.00 2.58 0.62 0.79 2.78 1.55 Low 2 4 High 0.51 0.44 3.41 2.23 - RF(r)] + OI~ERM(I) 0.00 0.92 0.02 0.93 0.05 0.89 0.13 0.96 0.16 0.53

2.24 2.76 I(4 2.25 2.07 3 1.84

~~

Smnll Small 2 2 3 3 4 4 Big


Uig

- 0.34 0.31 -0.11 0.35 - 0.1 I 0.34 0.09 0.4 1 0.21 0.34

- 0.12 0.62 - 0.01 0.63 0.04 0.58 - 0.22 0.27 - 0.05 0.30 (v) H(l) - 0. I3 0.15 - 0.02 0.17 0.04 - 0.1s 0.22 -- 0.14 0.05 - 0.07

- 0.05 0.71 0.08 0.77 - 0.04 0.60 - 0.08 0.4) -0.13 0.25

K(r) (i) W(r) = 0 + h[RM(r) + K(r) - W(r) = (I 0.01 0.80 0.03 0.75 0.05 0.73 0.03 0.69 - 0.05 0.50

sSM&I) h/IML(f) + V(I) + dDEF(f)+ + t$r) - 1.47 - 0.73 - 3.16 2.20 0.75 1.73 1.04 - 1.24 - 0.20 2.60 0.93 I.91 - 1.42 0.47 - 0.47 2.28 I.00 I .9Y I .07 - 2.65 - 0.99 1.96 1.34 1.01 3.27 - 1.46 - 0.67 1.17 1.35 1.27 + dDEF(r) + &) - 0.79 1.54 1.10 - 2.35 0.42 1.n2 - 0.94 - 1.20 1.46 - 0.70

0.22 2.61 0.51 2.85 0.7 I 3.01 0.33 2.X8 0.69 2.36

0.14 2.87 0.34 3.03 0.56 3.1 I I .24 3.35 - I.41 2.14

Sm;ill sm;III 2 2 3 3 4 4Big


Big

--

~~_.___ - 0.04

0.35 0.22 0.1 I 0. IX 0.12 0. I6 0.0X 0.05 0.21

RF(f) = u + h[RM(r) - RF(r)] + sSMB(r) + hHML(r) + ntTERM(r) (ii) R(r) - RF(r) = u + ~[RM(I) - RF(r)] + ($0 0.0 I 0.00 - 1.58 - 0.05 - 3.24 - O.YO 0.42 0.54 0.30 0.73 0.0x 0.04 0.02 - 0.24 ~ 1.29 - 1.00 0.3) I .05 0.53 - 0.36 0.03 0.06 0.05 - 1.45 0.48 - 1.12 0.39 0.23 1.25 0.50I 3 - 0.08 0.04 0. 1.04 - 2.67 - 0.50 - 1.50 0.35 0.57 0.12 3.29 - 0.72 - 0.13 - 0.06 - 0.17 - 0.4Y - 0.95 0.20 0.21 _ ~~~~__.~~~.~. - 0.07 ~~~

0.20 2.19 0.67 2.7Y 0.79 3.20 0.47 2.91 0.73 1.89

0.09 2.53 0.29 3.01 0.56 3.19 1.23 - 3.71 I.51 1.41

See footnote

under table 9c.

Though, the GRS Test Marginally rejects. Ooops!

Factor data are from: http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html


2011 Patrick J. Kelly 127 2008 Patrick J. Kelly 128

Potrebbero piacerti anche