A HIGH RISK - LOW BETA BUSINESS?
BETA ESTIMATION IN THE INTERNATIONAL TRANSPORT INDUSTRY
a ； b a aStephen X.H. Gong, Michael Firth, Kevin Cullinane, Teng-fei Wang
a Department of Shipping and Transport Logistics, The Hong Kong Polytechnic University, Hong Kong b Department of Accountancy, The Hong Kong Polytechnic University, Hong Kong
The water and air transport industries are perceived as risk-laden, cyclical businesses suffering from high financial and operating leverage. Contradicting this common perception, however, existing studies have documented relatively low beta risks for the securities of these industry sectors. On the basis of the theoretical determinants of beta, we argue that the reported low beta risk levels may not be representative of the industries‟
true beta risk, and are not robust to different estimation procedures. Our analyses using alternative estimation procedures reveal that the industry beta estimates are indeed sensitive to different estimation designs. In addition, we find that the beta estimates reported in previous studies are confounded by problems in the sample selection. By obtaining a more representative sample of industry firms and by implementing a number of different estimation procedures, we produce a range of industry beta estimates. Overall, our results lead us to conclude that the qualitative conclusion about beta risk for water and air transportation securities varies with the estimation procedure and that the industries‟ beta risk may not be stable over time.
Key words: Systematic risk, asset pricing, cost of capital, international transport
* To whom correspondence should be addressed. Phone: 852-2766 7411; fax: 852-2330 2704; e-mail: STLXHG@POLYU.EDU.HK.
The securities market has an extremely important role to play in a market economy because it helps to direct scarce economic resources to their most productive uses. Investors decide whether to buy or sell a security on the basis of their perception of the risk-return trade-off; they require high return for a security that has high investment risk. In modern finance, the risk (i.e. systematic risk, or beta) of an individual security for a reasonably well-diversified investor is measured in terms of its marginal contribution to the portfolio risk of all the securities the investor holds, since most of the non-systematic risks can be diversified away by holding just a few randomly selected securities. Because the business prospects of companies tend to be affected by the same macroeconomic factors, securities generally exhibit a certain degree of co-movement. Changes in the economic, political and sociological environment are, therefore, sources of systematic risk. At an individual firm level, these macroeconomic factors exert their influence on security risk and return via such factors as business cycles (and the resultant earnings variability), sales, profits, etc. Other microeconomic factors affecting a security's risk include operating and financial leverage, dividend payout ratio, standard of management in the industry, and the like (Brealey and Myers, 2000; Bodie and Merton, 2000).
The ocean and air transport sectors are perceived as highly risky, cyclical businesses that are international in nature and are susceptible to changes in macroeconomic factors (Cullinane, 1991). This is further aggravated by their typically high operating leverage and financial leverage. As a result of these economic and operating characteristics, it is generally expected that securities of these sectors should have a relatively high systematic risk, as industry practitioners often claim (see Stokes, 1996). Indeed, Cullinane and Gong (2002) find support for this conventional wisdom by presenting evidence that new issues of water transportation securities incurred higher underpricing costs than other transport industry sectors because of their perceived higher investment risks.
Despite the commonly held belief that the water and air transport sectors are risky, a number of recent studies have documented rather low beta levels for both. Kavussanos and Marcoulis (2001)—hereafter, K&M—for example, report that the average beta for
U.S. listed water transportation shares during the period July 1984 to June 1995 is only 0.9411, which is significantly lower than the market average of unity. They also report an average beta of 0.9748 for air transportation shares in the same period, which is
1. The magnitude of these beta values computed from insignificantly different from unity
monthly returns appears to be much lower than those estimated by Rosenberg and Guy (1976), which are 1.8 for the air transportation stocks and 1.19 for railroads/shipping. Cavarra, Stover and Allen (1981), using daily returns, report an average beta of 1.45 and 2.435, respectively, for U.S. airlines before and after the Airline Deregulation Act of 1978. Banker, Das and Ou (1997), also using daily returns, report an average beta of 1.492 and 1.675, respectively, for US airlines before and after the Airline Deregulation Act. The discrepancies among these studies may be a result of the different sampling periods and/or
2the use of different samples, and/or due to different estimation methods.
The risk perception of an industry in the capital markets has important implications for its cost of public equity capital, pricing policies, as well as investment decisions. In this study we set out to investigate whether (and which of) the afore-said results are representative of the „true‟ systematic risk of the two industry sectors, as well as to suggest possible reasons for such discrepancies. Section 2 reviews the literature relating to the theory and estimation of systematic risk. Section 3 discusses the testing methodology and data selection. Section 4 presents sensitivity tests of the industry beta estimates to different estimation procedures, using both the sample firms originally used in K&M and a re-specified sample that is considered to be more representative of the respective industries.
1 In their footnote 17, Kavussanos and Marcoulis (1997) report a mean beta of 0.933 (0.891) when a multifactor model (the methodology of Fama and French, 1992) is applied to their sample of 28 water transportation stocks over the period January 1985-December 1994. Neither of these is statistically different from unity at a conventional level. The situation is, however, complicated by the fact that these 28 stocks include companies (e.g. energy) that are not pure water transportation companies. In another study, Kavussanos and Marcoulis (2000), using a multifactor model and the Seemingly Unrelated Regression Method, estimated a mean beta of 0.9438 (statistically lower than unity) and 0.9471 (insignificantly different from unity) for a subset of 14 relatively pure water transportation stocks and 13 air transportation stocks, respectively, over the period July 1984 to June 1995. Our discussion of K&M‟s findings for the risk level of the two transport industries focuses on this latter sample. Our main source of reference is Kavussanos and Marcoulis (2001), the latest publication that covers several of their previous studies including Kavussanos and Marcoulis (1997, 1998, 2000). 2 The chosen methods in each study naturally depended on the purpose of the investigation.
Section 5 interprets and discusses the implications of the results, while section 6 summarizes and concludes.
2. REVIEW OF THE RELEVANT LITERATURE
The relationship between security risk and expected return is a focal point in modern finance. The Capital Asset Pricing Model (CAPM) due to Sharpe (1964) and Lintner (1965) posits that the expected return to a security is a simple linear function of the security's systematic risk and the expected market return. This ex ante relationship is
captured in the following equation:
whereis the expected return to stock i, is the (one-month treasury bill) risk-free E(R)Rfi
rate of return, is the expected return to the market portfolio, andis stock i‟s E(R)，mi
systematic risk or beta, a measure of the stock‟s sensitivity to movements in the market portfolio.
In most practical work, the market model of Sharpe (1964) and Fama (1976) is usually used to estimate beta from historical data. One justification for preferring the market model to CAPM is that even if CAPM is not descriptive of the security return-generating process (for such evidence, see Fama and French, 1992, 1993, 1996), there are still compelling reasons to retain the concept of beta (Black, 1993; Chan and Lakonishok, 1993). The market model is represented by the following equation:
~~~ (2) R，！？，R？(,itiimtit
~~where is the (raw) return to stock i at time t, is the return to the market portfolio at RRitmt
~time t, and is a zero-mean disturbance term. The parameter;is often estimated via (，iti
OLS over a sufficiently long historical period. For monthly returns, the optimal estimation period is from five to seven years, during which the assumption that beta is constant is a reasonable one (Fama, 1976). Over longer periods, the beta values of individual stocks may change. Indeed, change of beta over time is likely the norm rather than the exception. Stock beta reflects the degree of co-movement between the security and the market
portfolio, i.e. the responsiveness of the security price to stock-market-wide movement, itself a function of the general economic environment. This can be better seen in the following equation:
~~RRcov(,)im， (3) ，,~i2：R()m
~2If over time the market becomes more volatile (higher variance, ) and the stock's ：(R)m
~~price responsiveness to the market (i.e. ) remains the same as before, then the cov(R,R)im
security's beta becomes smaller in magnitude. If, on the other hand, the market is stable but the security's own idiosyncrasies (e.g. the covariance of its return with the market) change over time, then future beta will change. In such a case, the historical beta (either the average beta or the beta closest to the current period) is a biased predictor of future beta because fundamentals of the firm have changed within the historical interval (Rosenberg and Guy, 1976). A rapidly accumulating body of empirical studies on the statistical properties of beta has cast doubt on the validity of the assumption of beta stability. Fabozzi and Francis (1978), for example, find evidence to suggest that beta coefficients that are estimated from historical returns via OLS regression may be a random coefficient, i.e. it may vary considerably from period to period. This is corroborated by evidence provided in other subsequent studies (see Brailsford, Faff and Oliver, 1997 for a summary). In reality, the choice of the estimation period involves a trade-off between possibly unstable beta and a statistically meaningful number of observations.
On the other hand, Scholes and Williams (1977), Dimson (1979) and Cohen et al. (1983)
show that, when stocks do not trade at the same level of frequency as the market index, OLS may produce biased beta estimates. For infrequently or thinly traded stocks, this problem is exacerbated as the sampling interval is reduced (e.g. from monthly to daily), since the frequency of trading declines with the reduction in the sampling interval. Hawawini (1983) presents a model that explains the direction and size of changes in beta resulting from changes in the sampling interval. The basic idea is that beta is sensitive to the return interval because the covariance of security returns with market returns does not change proportionately as the return interval is varied. As the return interval is
lengthened, betas of thinly traded stocks (a problem that is particularly significant with those securities of smaller firms) increase, while betas of frequently traded stocks (these are likely to be large firms) decrease. Brailsford and Josev (1997) provide empirical evidence from the Australian stock market that is consistent with this prediction.
Roll (1992) presents evidence to suggest that the correlation of an industry‟s security returns with a market index is related to how well the industry is represented in the relevant market index (as a proxy for the theoretical but unobservable market portfolio of all risky assets). If industry effects/movements are not adequately reflected in the market index, an industry stock‟s beta may be under-estimated, as compared to the case where the
market index fully captures the industry‟s performance.
Past studies of beta risk for shipping and other transport sectors have not considered the robustness of beta to these research design issues. The main focus of this study is to test the sensitivity of transport industry beta estimates to different estimation designs, including choice of the sample and sampling period, the market index, the sampling interval, and the estimation techniques (c.f. Unal and Kane, 1988; Breen and Lerner, 1972). Applying alternative estimation procedures provides a range of beta estimates, the reasonableness of which may be evaluated in the light of the industry‟s economic and operating characteristics. Our results indicate that at least in part, the results in previous studies are due to the specifications of the estimation method used, and due to confounding effects in the classification of sample firms.
3. TESTING METHODOLOGY AND DATA SELECTION
In this study, we use two estimation techniques that are designed to deal with the problem of thin trading, a problem that is believed to be particularly significant in generally small-capitalization transportation shares. The first is the Scholes-Williams (1977) method, which suggests the following adjustment to account for thin trading:
？1？1ˆˆˆ，，，？？()iSWii，， (4) ,iˆ？，(12)m
？11？ˆˆˆ, and are obtained in an OLS regression of stock return on one-period where ，，，iii
lagged market return, on contemporaneous market return, and on one-period lead market
ˆreturn, respectively; is the first-order serial correlation coefficient of market returns. ，m
The second procedure is that of Dimson (1979), which is an aggregated coefficient estimate that can be estimated by the following:
mDimˆ (5) ，，，,i?？ik，？km
ˆwhere the beta estimates are obtained from a multiple OLS regression of individual ，i？k
3stock returns against lagged, contemporaneous and leading market returns. The number
of leads and lags is indicated by the m in equation (5). Since an increase in m achieves a
reduction in the bias but at the cost of a reduction of estimation efficiency, in accordance with most existing studies, we limit the number of leads and lags to 1 only.
To test the sensitivity of transport industry beta estimates to the use of different estimation procedures, and to be comparable with existing studies in the transport industry, we use two sets of samples. The first set consists of 14 water transportation (i.e. shipping) shares and 13 air transportation shares listed on the U.S. stock exchanges during the period July 1984 to June 1995; this is the sample used in Kavussanos and Marcoulis (2001). Our use of the same sample but different estimation methods facilitates comparability of the results and the identification of estimation problems that might have led to the conclusion of relatively low betas for the highly risky shipping and air transport industries.
Our second sample aims to be as representative as possible of the population of international shipping and air transportation shares in the world‟s major stock exchanges
4(see Appendex A for a list of the countries represented). The sampling period is
3 Fowler and Rorke (1983) argue that Dimson‟s beta is incorrectly specified when the market returns are serially correlated. 4 Gong, Firth and Cullinane (2002) examine the sensitivity of beta estimates to different research designs for U.S. listed transportation securities, using both the original K&M sample and a re-specified sample which corrects the sample selection problems identified in K&M and which is considered more
determined with the intention of maximizing the number of listed firms available for the period and should be long enough to cover at least one business cycle. The period January 1980-December 2001 is a period during which there exists a relatively large number of listed transportation securities worldwide and which may be considered long enough. This is the sampling period used here.
Returns data for the U.S. listed transportation companies (those used in K&M) are retrieved from the University of Chicago's Centre for Research in Security Prices (CRSP) database as at December 2001. International transportation shares are those classified by Bloomberg as air transport (126 shares) and water transport (219 shares) and must have at least 60 consecutive monthly returns during the 1980-2001 period. The time series of simple returns for these shares (adjusted for stock splits and stock dividends) are obtained from Datastream International.
4. EMPIRICAL ESTIMATION
4.1. Shipping Industry Beta Estimates
Exhibits 1 to 3 present a summary of the parameter estimates using different estimation
5) and during different time methods (OLS market model and Scholes-Williams model
periods (whole period versus sub-periods). It is noteworthy that some of the sample firms K&M classified as water transportation companies (SIC Code 44) were in fact classified in CRSP as other industries during the period under examination. For example, among the 14 water transportation stocks used in K&M, GATX Corp was listed under railroads equipment (SIC Code 3743), Amerada Hess Corp under petroleum refining (SIC 2911), Hawaiian Electric under electric services (SIC Code 4911), Nicor Inc under natural gas distribution (4924), and Lowe‟s Cos under lumber, plywood and millwork wholesale (SIC
6Code 5031). This (misclassification) may have the effect of introducing „noise‟ into the
representative of the population of U.S. listed transportation securities. Their conclusions are similar to those reached here. 5 All results using the Dimson (1979) method are similar to those using the Scholes-Williams method and thus are not separately discussed. 6 The SIC codes reported here are the primary SIC codes for the last reporting date on or before the starting date of data retrieval from CRSP. An inspection of the financial reports for these firms revealed
beta estimate for the water transport industry, a problem that is also noted in the sample of air transportation stocks.
4.2. Air Transport Industry Beta Estimates
Exhibit 4 reports summary results of the parameter estimates for the whole period July 1984 to June 1995 for 13 air transport companies used in K&M. Out of these companies classified as air transportation companies (SIC Code 45), two observations, American West Airline and Ronson Corp., had either missing or insufficient data points from the CRSP database during this period and are thus excluded from the sample, leaving a total of 11 companies. Three other remaining companies, Kimberly Clark (SIC Code 2621, paper products), Airborne Freight (SIC Code 4712, transportation services), and Rowan Companies (SIC Code 1381, drilling oil and gas wells) were actually classified as non-air transport companies in CRSP. To enable a comparison with K&M, in Exhibit 4 we report beta estimates for the remaining 11 air transportation stocks for the whole period and two sub-periods. Since misclassification of stocks would confound industry beta estimates, in subsequent estimations we replace the aforesaid non-qualifying observations with randomly selected observations that were categorized as air transportation stocks and that had sufficient data points during the sampling period. Since these random replacements help to make the sample a more homogeneous group of air transportation stocks, the
7. results are believed to be more representative of the „true‟ beta risk of the sectorSummary results of parameter estimates for the sample in which disqualifying observations are replaced are reported in Exhibits 5, 6 and 7.
4.3. Sensitivity of Beta Estimates to Different Estimation Procedures: Some
that shipping activities represented only a minor proportion of their operating revenues. Two probable exceptions are CSX Corp and American President which, despite their coding as holding companies (SIC 6711), are widely known as shipping companies or as having significant shipping operations. On the other hand, Alexander and Balwin (SIC 4421), although often quoted as a water transport company, is heavily involved in properties development. Its shipping operations are also rather restricted to the Hawaiian area rather than being international. Inclusion of such observations tends to confound (and downward bias) the beta estimates. The SIC code of some companies changed during the period examined. However, this does not alter the conclusion here. 7 A more thorough way of achieving representativeness would be to include all eligible industry stocks. This is feasible given the small population and is the approach taken in Gong, Firth and Cullinane (2002).
The main purpose of this section is to investigate the sensitivity of beta estimates to different estimation procedures for a large sample of internationally traded transportation stocks. In particular, we are interested in whether beta estimated via the OLS market model differs significantly from that estimated via the Scholes-Williams methodology or the Dimson methodology. We are also interested in the temporal variations in beta (which is indicative of beta stability) when it is estimated over overlapping as well as non-overlapping periods.
4.3.1. Beta stability (beta shift over time)
Beta stability is important both as a theoretical construct and as a practical investment selection technique. While CAPM provides important insights into the nature of actual capital markets, it is of limited value for the selection of an investment strategy unless additional specifications are made concerning the stability and/or predictability of key measures within the theory (e.g. beta)--see Sharpe and Cooper (1972). In both theory and practical work, it is usually assumed that beta is stable over time, but changes in risk-class membership are not unimportant since they give rise to transactions costs for investors who adopt a strategy of basing their prediction of future beta on measures of beta in the past (Sharpe and Cooper, 1972). One way to investigate whether such an assumption is tenable is to see whether securities within a certain risk class might change their membership in another period. This is achieved by adopting the Sharpe and Cooper (1972) methodology as follows. For every year during the period 1980 through 2001, beta for each stock that has monthly returns data available in the preceding 60 months (minimum 48 months) is estimated using the OLS market model. The beta estimates are then classified into one of ten deciles of beta values, with 10% of the highest-beta stocks being assigned into decile 10 and 10% of the lowest-beta stocks into decile 1 and so on. This procedure is repeated in each of the following years, using returns data from the immediately preceding 60 months. Then, a stock‟s beta risk class in each year is first compared with the class in the succeeding year and then the class five years hence. While the first comparison uses 48 months of common data, the second involves no overlapping