How to determine the:

stability o f beta values

A critique of earlier methods and demonstration ofa better way.

Arthur A. Eubank, Jr. and 1. Kenton Zumwalt

The Journal of Portfolio Management 1979.5.2:22-26. Downloaded from www.iijournals.com by ambreen saadat on 04/01/13. It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

22

8
5cl
'0

; ;
u

lthough the subject of beta stability has been examined in several previous studies, the focus of these studies has been more upon the stability of beta rankings rather than the stability of actual beta values. The purpose of this paper is to examine the stability of securitybetas in different risk classes, using an equal class interval beta classification scheme based upon numerical beta values instead of beta ranks. We use numerical beta class intervals because the numerical beta value is more important than the beta rank for portfolio management purposes. Although the stability of beta ranks is of interest, beta rankings of individual securities may change, with no change occurring in the actual beta values. Conversely, a change in the numerical values can occur with no change in the beta ranks. Previous studies using decile and pentile rankings imply that betas in both high and low beta groups are more stable than betas in the middle groups. This study examines the relative stability of betas in extreme beta groups compared to middle beta groups by comparing the ,stability of individual risk classes using, first, a decile classification method and, second, an equal class interval classification method.
TESTS OF BETA STABILITY

The two primary methods used in earlier studies of the stability of betas have been correlation analyses and transition matrices. Correlation analysis was used by Blume and Levy to examine the stability of security and portfolio betas in successive nonoverlapping time periods.' Both studies found that
1. Footnotes appear at the end of the article.

increases in the size of portfolios were associated with increases in the correlation of portfolio betas between successive non-overlapping time periods. 'These studies also found that security and portfolio betas tended to regress toward a mean of 1.0 over time. The principal difference between the two studies WAS the use of different estimation intervals in the calculation of betas. Blume's study utilized seven-year beta estimation intervals, whereas Levy used 13, 26, arid 52week estimationintervals. Levy found th.at the use of a longer estimationinterval tended to increase the correlation of beta groups between successive nonoverlapping time periods. While the use of correlation analysis indicates the degree to which betas are in thesame group in successive time periods, it does not indicate into which groups betas tend to change over time. In order i o determine how betas change over time, a study by Sharpe and Cooper and a later study by Baesel utilized transition matrices to examine the temporal stability of individual security betas.2 In the study by Sharpe and Cooper, individual security betas were ranked by magnitude and placed into decile groups. The stability of individual security betas was examined by noting the change in decile groupings over time. Sharpe and Cooper concluded that individual securitie, c were resonably stable over time. In addition, they found that betas in the extreme decile groups tended to be more stable than betas in the middle decile groups. Baesei, using different length estimation intervals in the beta calculation, concluded that individual betas were less stable than portfolios, but he rioted that their stability improved as longer estimation intervals were used. He also observed that extreme group betas

~~

appeared to be more stable than those in the middle groups. The differences in beta stability among extreme versus middle decilelpentile groups are somewhat surprising, since extreme betas tend to regress toward the mean and, intuitively, it would seem that securities with more extreme betas would exhibit less stability, i.e., a faster rate of regression toward the mean, than those nearer the mean of 1.0. A possible explanation of empirical results showing betas in more extreme groups to have greater stability over time may be the grouping methodology used in the studies.
The Journal of Portfolio Management 1979.5.2:22-26. Downloaded from www.iijournals.com by ambreen saadat on 04/01/13. It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

DATA AND METHODOLGY

ALTERNATIVE BETA GROUPING METHODS

The use of a decile grouping method with equal numbers of securities in each group on the basis of security beta ranks instead of beta values results in wide numerical beta ranges for extreme groups and narrow beta ranges for middle decile groups. For example, if a security had a beta of 2.5 in one period and a beta of 2.0 in the following period, in all probability the security would still be ranked in the highest risk class group because of the wider range of betas in the more extreme groups as a result of the limited number of betas greater than 2.0. On the other hand, a security in the middle range with a beta of 1.1in the first period and a beta slightly lower, say, 1.0, in the next period could actually shift from the first period beta group to a lower beta group in the folIowing period. It might appear that the security with a middle group beta in the first period had substantially decreased its risk by moving from one risk group to another in the next period, while the larger reduction in beta of the more extreme security would not result in a risk group change from one period to the next. Indeed, it is not the ranking (or ranking change) that investors are concerned with, but the change in risk because of the change in the numerical value of beta. An alternative to the decile grouping method, whereby equal numbers of securities are placed into each group, is the equal class interval grouping method, where securities are placed into beta groups based upon equal beta ranges to define the groups on the basis of numerical beta values. This grouping scheme provides more information concerning the actual numerical value of beta and the stability of its numerical value over time than does a decile grouping method with equal numbers of securities in each group without regard to the actual numerical beta values of the securities. Thus, an equal class interval grouping method provides the basis for a more meaningful examination of the stability of security betas over time than do decile or pentile groups, where the primary concern is only with changes in beta ranks and not with changes in numerical beta values.

The securitybetas used in this study were calculated using the monthly return relatives from the CRSP return file. The market return used for the regressions was the Fisher Index. Sixty months of data were used in the calculation of the betas for each fiveyear period, and the fifty-year time period from January, 1926 through December, 1975resulted in ten nonoverlapping five-year time period^.^ The pairs of companies having valid data for consecutive nonoverlapping time periods resulted in 319 pairs of fiveyear betas for the first comparison period (1926-30and 1931-35), and as successive time periods were used, the number of observations for each pair of nonoverlapping periods increased to 876 pairs of five-year betas for the last comparison period (1966-70 and 1971-75).4 After computing the security betas for companies with available data for a particular comparison period, the betas for each nonoverlapping time period were ranked in ascending order. Two grouping methods were then used to classify the betas. First, following the procedure reported earlier by Sharpe and Cooper using deciles, the betas were divided into ten groups with equal numbers of securities (deciles), and transition matrices were used to determine the proportion of the securities either remaining in the same risk class in periods t and t + 1, or changing classes from periods t to t + 1.j A second grouping method using an equal class interval grouping scheme was then applied to the beta pairs in each time period. By this method, betas were grouped into risk classes on the basis of their numerical values rather than their ranks. Ten classes were established for grouping the security betas, whereby securities with beta values less than or equal to .2 were assigned to group one, those greater than .2 and less than or equal to .4 were assigned to group two, etc. This procedure of grouping securities according to their betas was continued using increments of .2 until all securities were placed into groups, with group ten containing those securities with betas greater than 1.8. This grouping scheme resulted in eight groups with equal class intervals (groups one and ten were openended), but with unequal numbers of securities in each group. Transition matrices were again used to examine how the security betas changed over time. In order to examine further the stability of security betas in successive time periods, the mean square error and the mean absolute deviation were calculated for each of the nine comparison intervals for the complete set of securities, and subsequently for each beta group using the decile grouping method and

The Journal of Portfolio Management 1979.5.2:22-26. Downloaded from www.iijournals.com by ambreen saadat on 04/01/13. It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

the equal class interval grouping method." Finally, the bias, inefficiency, and random error components of the mean square error were calculated in order to shed additional light on the source of any instability that may occur over time.7 The components of the mean square error are reported for the complete set of securities over all time periods as well as for each set of the different risk classes. The mean square error components for the different risk classes were calculated in order to examine the difderential effects of the bias, inefficiency, and random error associated with betas in middle groups compared to betas in the more extreme groups.
EMPIRICAL RESULTS

treme betas is associated with stability in a relative sense in that stocks with the highest or lowest betas in one period also tend to have high or low betas in a subsequcmt period. Table 2 shows the distribution of the number of betas falliing into equal class intervals. As can be seen, the extreme beta intervals have far fewer observations than those near the mean. In order for a secuiity to change from one equal class interval group to another, the numerical value of the beta must change, rather n the than its beta rank. As a result, a given change L numerical value of a security beta will have the same impact upon determining its subsequent group regardless of the group it is in, except for open-ended groups m e and ten.
TABLE 2

24

2
\o

?
s

The transition matrix results for the grouping procedure utilizing deciles are reported in Table 1. Risk class one represents securities with the lowest betas while group ten represents those securities with the highest betas. The values in the transition matrix represent the proportion of securities remaining in a particular risk class in periods t and t + 1.For example, indicates that 39.1% of the value .391 in element (1,l) the betas in the decile with the lowest betas were also in the lowest beta decile group in period t + 1. Similarly, element (1,2) indicates 24.6% of those betas in the lowest beta group in time period t were one beta decile group higher in period t + 1.The values on the main diagonal of the matrix are the proportions of betas which did not change risk classes from time period t to period t + 1. These results are similar to those reported earlier by Sharpe and Cooper and Baesel in that the proportions remaining in the same groups are higher for the more extreme risk classes compared to the middle classes. This finding tends to indicate that extreme betas are more stable than those near the mean of 1.0. However, the apparent greater stability of extreme betas results primarily from the decile grouping method used to place the securities into their respective risk classes. That is, the apparent stability of exTABLE 1
R i s k Class i n P e r i o d T vs. R i s k Class i n P e r i o d 1

The transition matrix results for the equal class interval grouping scheme are reported in Table 3. In comparing the main diagonals of Tables 1and 3, it can be seen that there is a substantial difference between the relative stability of extreme beta groups compared to middle groups indicated by the two grouping methods. In fact, Table 3 shows that the middle groups are more stable than those exhibited in Table 1, whereas the extreme groups are much less stable. Also, the distribution around the main diagonal is more compact in Table 3 than under the decile grouping scheme reported in Table 1.
TABLE 3

+ 1

R i s k Class i n P e r i o d T v s . R i s k Class i n P e r i o d T + 1

Betas Grouped by D e c i l e s
R i s k Class i n Period t
I

Betas Grouped by Equal Class I n t e r v a l s

R i s k Class i n P e r i o d t

+ 1
7
8
!)

R i s k Class i r Period t

R i s k Class i n P e r i o d t + 1

1 ,391 .241 ,141

.083

2 ,246 .222 ,173 .138 .058

,063

2 ,121 ,166 ,162 .146 ,131 ,104

.068 ,056

4 .083 ,136 ,146 .155 ,119 ,111 ,108 .071 .053 .019

5 .053 ,078 ,115 .120 ,170 ,151 ,127 ,100 .056 .031

10

2 3 4 5 6 7 8 9 10

.058 ,024
,032

,048 .015 .029 .010 .011 .006- ,015

.031
,016

,046 ,056 .lo6 .099 .136 .141 ,132 .116 .098

,070

,019 .046
,058 ,098

,125 .135 ,161 .126 ,147

,084

.021 .018 ,053 ,072 ,091 ,128 .135 .168 .156 ,155

.010 ,013 ,014 ,050 ,038 ,079 ,034 ,083 ,059 .114 ,074 ,182 ,137 ,215 ,223 .209 ,395 .010 ,024 ,034

1 2
3 4 5 6 7 8 9 10

1
.063

2 ,250 ,251 ,139 .061 ,024 ,009 ,005 .004 ,004

,000

4 .250 .187
,287

5 ,104 .073 ,160 ,230 ,249 .251 ,151 ,112 ,119 .038

I
,000 .M10 ,001 .004 ,009 ,029 ,054 ,092 .189 ,195

,066 .010 ,012

.003 .001

.004

.OOO

,208 ,372 .306 ,194 .lo4 .058 ,029 .013

.008 ,017

,104
,030

.OOO
,004

,267 ,230 .147 .114 .052 .025 .025

.055 .145
.218

,234 .226 ,226 .169 ,169

,000 ,000 .021 ,009 ,003 .009 .025 ,012 .004 .060 ,022 ,005 ,096 ,050 ,018 .164 ,075 ,032 .219 ,125 ,073 ,244 ,177 ,078 ,202 .156 .128 ,186 ,169 ,195

The Journal of Portfolio Management 1979.5.2:22-26. Downloaded from www.iijournals.com by ambreen saadat on 04/01/13. It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

A summary of the beta stability results for the two grouping alternatives is given in Table 4. Under the decile grouping method, the securities appear to become more stable as the beta moves farther away from the middle risk classes, while the middle classes are more stable under the equal interval grouping scheme. Again, this apparent conflict results from the difference in the ranges of the numerical values of the betas falling into each of the respective risk classes for each of the two grouping methods. The decile method causes the ranges of betas in the middle risk classes to be narrow while the ranges of beta values in the more extreme classes are wider. Hence, a givenbeta change in a more extreme decile group may have less impact in terms of causing a security to change from one class to another than a similar beta change in a middle decile class. On the other hand, the equal class interval grouping method does not have this characteristic, since a given change in the numerical value of a beta will cause the same change in beta risk class regardless of the securitysbeta group (except for the highest and lowest open-ended risk classes).
TABLE 4 Beta S t a b i l i t y Using A l t e r n a t i v e Grouping Methods P r o p o r t i o n i n Same R i s k Class R i s k Class 1 Deciles .391 .222 ,162 .155 ,170 .141 ,161 ,168 ,215 ,395 Equal Intervals P r o p o r t i o n W i t h i n One R i s k Class
Deci 1es

group mean is typically less for the decile grouping method than for the equal interval method. This is attributable to the smaller difference among the mean betas for each decile group compared to the larger differences between the mean betas for the equal class interval groups.
TABLE 5 Beta Group Means for Periods T and T
.
t

Risk Class

Betas Grouped by Oeciler Period T Period T + 1 Chanqe .361 ,545 ,667 .778
,886

__

Betas Grouped by Equal Intervals Chanqe Period T Period T + 1 ,115 ,325

,510

9 10

.988 1.100 1.230 1.409 1.792

,575 ,677 ,770 ,857 ,939 ,998 1.050 1.146 1.256 1.419

,214 .I32 ,103 ,079 ,053 ,010

-.050
-.084

-.153 -.373

,701 .900 1.098 1.293 1.490 1.690 2.052

.595 ,533 ,663 i9

,480 ,208 ,153

.me
,024

.9y1

1.064 1.183 1.a2 1.421 1.468

-.034
-.110 -.I88

-.E9
-.584

Equal I n t e w a 1s

,063
.251

,637 ,629 ,481 ,421 ,425 ,427 ,428 ,476 ,594 ,604

,313
,689

3
4

,306
,267 .249 ,234 ,219 ,177 .128 ,195

,132
,691 ,691 ,649 ,570 ,499 ,473 ,390

The forecast errors were calculated for each beta group between beta estimates for periods t and t 1 for both grouping methods. The mean square error (MSE) and the mean absolute deviation (MAD) are presented in Table 6. The forecasterrors for the middle groups for both grouping methods tend to be similar, although the more extreme groups appear to be less stable (less predictable) when actual numerical value beta changes are used instead of changes in beta ranks. Contrary to the results of earlier studies showing more stability for extreme groups, this study shows that the more extreme groups based upon equal class intervals tend to contain betas that are substantially less stable.

5
6
7

TABLE 6 Forecast E r r o r s Between Beta Estimates f o r Periods T and T + 1 Betas Grouped by Oeciles MSE MA0 ,128 ,095 ,098 .lo5 .096 ,107 ,132 ,128 ,154 ,359

8
9 10

1
I

Risk Class

Betas Grouped by Equal I n t e r v a l s MSE MAD

REGRESSION TENDENCIES OF BETA GROUPS

1 2

.263
,226

,330
,113 .097 ,099 ,097 .128 ,144 ,176 ,257

,486 .249
,227 ,239 ,242

The mean betas of each risk class associated with each of the two grouping methods were calculated for periods t and t 1and are presented in Table 5. As expected, the mean betas of the different risk classes regress toward a mean of 1.0 from time periods t to t + 1. There is a noticeable difference in the magnitude of the regression tendency between the two grouping schemes, however. For example, when the securities are grouped by deciles, the mean of the lowest decile group increased from .361 to .575, an increase of .214. However, when the equal interval scheme is used, the mean of the lowest beta class increases from a much lower .115 to .595, a change of .480. Typically, the farther a group mean is from the middle, the greater is the numerical shift of the group mean over time. Also, the numerical shift of each

,235 ,252
,240

5
6 7 8
9

,245

,268
,283 ,316 .480

,266 ,303 ,335

,415 ,614

10

,550

Finally, the components of the forecasting error as measured by the MSE are presented in Table 7. As can be seen, the bias components, indicating a tendency to incorrectly estimate both high betas and low betas, is much higher for the extreme beta groups and is relatively small for the middle beta groups. This indicates a greater amount of stability for betas falling in either middle decile groups or middle equal class interval groups. The greatest amount of estimation inefficiency is associated with group ten of both group-

lable 7 Components o f Mean Square E r r o r Betas Grouped by Deciles Risk Class 1 2 3 4 5 6 7 8 9 10

On the Assessment of Risk Some Further Considerations, Jor,iraZ of Finance (December, 1974). Individual:security betas, pi's, were calculated using the following regression equation
MSE
,12773 .09544 .09820 .lo543 .09602 .lo707 .13170 .12806 .15376 .35918

Bias
.04580 .01766 .01067 .00616 .00279 .00011 ,00251 .00696 ,02329 .13913

Inefficiency ,00579 ,00131 ,00079 .00140 .00026 .00009 .00028 .00076 .00150 .04949

__ Error
,07614 .07648 .OB674 .09788 ,09297 .lo687 .12890 .12033 .12897 ,17056

Random

Rit = a,+ PIRm,.t E ,

,where

R,t is the natural logarithm of the security return rdative in period t (including cash dividends and adjusted for stock splits and stock dividends), and R,, is the natural logarithm of the Fisher Index return relative in period t.
Each comparison period includes two consecutive, nonoverlapping five-year periods. The first five years of each successive comparison period is equivalent to the last five years of the preceding period. This procedure of selecting comparison periods for examining the stability of security betas between the first five-year period of the cornpoxison period and second five-year period resulted in nine tenyear comparison periods. Period t refers to the first five years and period t + 1refers to the second five years of each respective comparison period. The mean square error (MSE) was calculated using the following equation:
N

The Journal of Portfolio Management 1979.5.2:22-26. Downloaded from www.iijournals.com by ambreen saadat on 04/01/13. It is illegal to make unauthorized copies of this article, forward to an unauthorized user or to post electronically without Publisher permission.

Betas Grouped by Equal I n t e r v a l s Risk Class 1 2 3 4

5

Bias .23045 .04331 .02338 .00952 ,00119 .00113 .01214 .03521 .07275 .34026

Inefficiency .00018 ,00038 .00158

___ Error
.09965 .06912 .07214 .08930 ,09585 .12645 .13175 .13981 .le381 .15027

Random
~

MSE

,33028 .11281 .09709 .09890 .09739 .12782 ,14413 .17566 ,25668 .55014

.00008
.00036 .00024 .00024 .00064 .00012 .05961

6 7 8 9 10

i=l

MSE = -N, where N refers to the number of securities with available data for both five-year periods of each of the nine comparison periods and Pliand PZi refer to the ithsemTitiys betas for the first five years and second five years of each comparison period, respectively. The mean absolute deviation (MAD) was calculated using the following equation:
N

ing methods. The random error component is larger for the open-ended equal class interval groups, a s expected, and smaller for the decile groups.
SUMMARY AND CONCLUSIONS

MAD = ______

is1

Numerical beta changes are of more interest for investment analysis than changes in beta rankings. In contrast to the findings of other studies, this study shows that betas associated with extreme beta groups are less predictable than those closer to a mean beta of 1.0. The apparent greater stability of betas associated with extreme groups reported in earlier studies is a re-

A description of the MSE components is given by :R. C. Klemkosky and J. D. Martin, The Adjustment of Beta Forecasts,JournuZofFinance (September, 1975).The co:mpo:nents of the MSE: are determined as follows: Total MSE where

MSE =:

B2 -

+ (1 -- bJ2 Sg, + w Bias Inefficiency

+ (1 - rt,

9 ,

L - 2

Random Error

sult of the grouping schemes used in those studies and

is not because of inherently greater stability of extreme beta securities or portfolios. Transition matrices, changes in group means, a n d the magnitude of forecast errors provide evidence of greater instability of
betas falling into more extreme beta groups based upon a n equal class interval grouping method versus a decile grouping method.
I

PI

= the mean of all period t betas across all comparison periods,

= the mean of all period t

PP
b,
S&

ison periods,
= the sample variance of

+ 5 betas across all comparpli,

= the slope coefficient of PZi on

PI,

M. E. Blume, On the Assessment of Risk, Journal of Finance (March, 1971); R. A. Levy, On the Short-Term Stationarity of Beta Coefficients, Financial Analysts Journal (NovemberDecember, 1971).

Si2 = the sample variance of P2, and r&,81= the coefficient of determination for the regressi,onof

p2 on p,.

W. F. Sharpe and G . M. Cooper, Risk-Return Classes of New York Stock Exchange Common Stocks, 1931-1967, Financial Analysts journal (MarchlApril, 1972); J. B. Baesel,

The bias term indicates the over- 01 under-prediction of P2 by PI. The inefficiency term indicates the tendency of prediction errors to be positive at low predicted values and negative at high predicted values. The random error term represents i i random disturbance element which is iinrelated to the value of the predictor, p,, or the predicted,.P2.