This action might not be possible to undo. Are you sure you want to continue?

BooksAudiobooksComicsSheet Music### Categories

### Categories

### Categories

Editors' Picks Books

Hand-picked favorites from

our editors

our editors

Editors' Picks Audiobooks

Hand-picked favorites from

our editors

our editors

Editors' Picks Comics

Hand-picked favorites from

our editors

our editors

Editors' Picks Sheet Music

Hand-picked favorites from

our editors

our editors

Top Books

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Audiobooks

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Comics

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Sheet Music

What's trending, bestsellers,

award-winners & more

award-winners & more

Welcome to Scribd! Start your free trial and access books, documents and more.Find out more

Michael C. Jensen

Harvard Business School mjensen@hbs.edu A considerable amount of statistical investigation of security price movements by economists and statisticians indicates that successive changes in security prices are (for all practical purposes) independent random variables.1 That is, all the statistical evidence indicates that future security price changes cannot be predicted by using the past price series. This has become known as the theory of random walks. It implies that the trading rules and security selection procedures long advocated by “technical” analysts or “chartists”, which are based solely on past price movements, will not be useful in aiding the investor to increase his returns.2 The technical analysts have responded to the evidence presented by the academicians by claiming their models and theories are not really captured by these statistical tests. Alexander (1961; 1964) and Fama and Blume (1966) have examined the returns earned by various “filter” rules for selecting securities which purportedly capture the essence of many technical theories. The evidence indicates these trading rules are not able to consistently earn returns superior to those of a simple buy and hold policy. Thus, the results of these studies also support the random walk hypothesis. Robert Levy (1966; 1967) has recently tested a number of additional trading rules based on technical theories. Some of his results seem to be inconsistent with the theory of random walks. In particular Levy’s article, “Random Walks: Reality or Myth” (1967) contains interesting results regarding the returns earned by several mechanical stock market trading rules in the five-year period October 1960 to October 1965. Levy calculates the returns earned by a number of variations of his trading rules and finds these returns generally higher than the returns earned on a “random selection policy.” Commenting on these results Levy states: “The evidence above conclusively proves that technical stock analysis could have produced greater-than-random profitability at less-than-random risk for the 1960-1965 period.” On this basis he concludes that “. . . the theory of random walks has been refuted.”3 Unfortunately, his results do not completely support these strongly worded statements. His results have not refuted the random walk hypothesis and indeed they are subject to criticism on a number of points. It is the purpose of this “Comment” to point out and clarify some of the issues not adequately treated by Levy.

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

© Copyright 1967. Michael C. Jensen. All rights reserved Financial Analysts Journal, November-December 1967. You may redistribute this document freely, but please do not post the electronic file on the web. I welcome web links to this document at http://ssrn.com/abstract=350420. I revise my papers regularly, and providing a link to the original ensures that readers will receive the most recent version. Thank you, Michael C. Jensen

Random Walks: Reality Or Myth—Comment

Michael C. Jensen

Financial Analysts Journal, November-December 1967

A considerable amount of statistical investigation of security price movements by economists and statisticians indicates that successive changes in security prices are (for all practical purposes) independent random variables.1 That is, all the statistical evidence indicates that future security price changes cannot be predicted by using the past price series. This has become known as the theory of random walks. It implies

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

that the trading rules and security selection procedures long advocated by “technical” analysts or “chartists”, which are based solely on past price movements, will not be useful in aiding the investor to increase his returns.2 The technical analysts have responded to the evidence presented by the academicians by claiming their models and theories are not really captured by these statistical tests. Alexander (1961; 1964) and Fama and Blume (1966) have examined the returns earned by various “filter” rules for selecting securities which purportedly capture the essence of many technical theories. The evidence indicates these trading rules are not able to consistently earn

1 2

C.f. Cootner (1964) and Fama (1965).

It is highly unlikely that the random walk theory is an exact description of the behavior of security prices. From the economist’s point of view it is only important that any dependencies in the price series be so small as to preclude their use in earning higher returns given the transactions costs which exist. Michael C. Jensen will receive the Ph.D. degree shortly from the University of Chicago Graduate School of Business and is currently Assistant Professor of Business Administration at The University of Rochester. The author expresses his appreciation for the helpful comments and criticisms received from Eugene Fama, Myron Gordon, and Myron Scholes.

M. C. Jensen

2

1967

returns superior to those of a simple buy and hold policy. Thus, the results of these studies also support the random walk hypothesis. However, Robert Levy (1966; 1967) has recently tested a number of additional trading rules based on technical theories. Some of his results seem to be inconsistent with the theory of random walks. In particular Levy’s article, “Random Walks: Reality or Myth” (1967) contains some very interesting results regarding the returns earned by several mechanical stock market trading rules in the five-year period October 1960 to October 1965. In essence his rules consist of variations of the following: (1) At the end of each week calculate the ratio of a security’s current price to its average price over the previous 27 weeks. (In his terminology this is the C/A26 ratio). (2) Rank the C/A26 ratios for all securities under consideration from high to low and invest an equal dollar amount in the X% of the securities having the highest values (or lowest ranks).

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

(3) Recalculate the C/A26 ratios and rerank all securities each week thereafter. (4) At each following week sell all securities currently held whose current rank exceeds a certain “cast out rank” and (5) immediately reinvest all proceeds from such sales equally in the X% of the securities currently having the highest values of the C/A26 ratio. Levy calculates the returns earned by a number of variations of this policy (varying both the “cast out ranks” and the value of X) and finds these returns higher in general than the returns which could have been earned on a “random selection policy.” Commenting on these results Levy states: “The evidence above conclusively proves that technical stock analysis could have produced greater-than-random profitability at less-than-random risk for the 1960-1965 period.” On this basis he concludes that “. . . the theory of random walks has been refuted.”3 Unfortunately,

3

C.f. Levy (1967).

M. C. Jensen

3

1967

his results do not completely support these strongly worded statements. His results have not refuted the random walk hypothesis and indeed they are subject to criticism on a number of points. It is the purpose of this “Comment” to point out and clarify some of the issues not adequately treated by Levy. While there are numerous small points on which Levy’s results can be questioned there are four basic issues not clearly treated in his work which tend to mitigate his results. These involve: (1) the definition of the naive standard of comparison (the returns on a “random selection policy”) as the “geometric average”, (2) the definition and treatment of the “risks” of the random selection policy and the trading rules, and (3) the lack of treatment of the problems associated with sampling error or selection bias,

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

(4) the implicit assumption that all trades can be executed at the same prices used in arriving at the buy and sell decisions. We shall discuss these issues in the order listed, and then we shall offer an interpretation of Levy’s trading rules in the context of two basic (and complementary) economic models of the co-movement of security prices and the effects of risk on the prices of capital assets. (1) —The Definition of the Returns on a “Random Selection. Policy” Levy (1967) argues that the appropriate definition of the returns earned by a “random selection policy” (to be used as the standard of comparison) is the “geometric average of the 200 stock prices,” claiming that “Geometric averages, as related to security prices, have the virtue of continually equalizing the dollar investment in each component stocks.” It is impossible from this definition to tell exactly how the returns of 10.6% for the “geometric average” given in his Table 1 (Levy,

M. C. Jensen

4

1967

1967) and reproduced in Table 1 here were calculated. However, personal communications with Levy have cleared this up, and before discussing the economic issues involved we shall clarify the definition. Let us define the wealth relative of any security as the ratio of the terminal value of an investment in that security (assuming all proceeds from dividends and rights are reinvested in the security when received) to the initial value of the investment. Thus over some holding interval the wealth relative for any security is simply the percentage gain (or loss) from an investment in that security plus unity. Now the 10.6% annual rate of return for the “geometric average” given by Levy is simply the percentage return obtained by taking the geometric average of these 200 wealth relatives adjusted to an annual compounding base.4 Now it can be rigorously demonstrated that the returns calculated in this manner are equivalent to the returns which would have been earned on a portfolio

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

which consists of an equal initial investment in each of the 200 securities with continuous redistribution of the assets of the portfolio at each instant of time so as to maintain an equal dollar investment in each security. Of course, this interpretation implicitly involves the assumption that security prices change continuously through time (rather than in discrete steps with a minimum change of 1/8 as is actually the case). As an alternative to this standard of comparison one could calculate the returns would have been earned by a naive investor with no information who simply buys all the securities under consideration and holds them to the end of the period. (The period in this case being approximately five years.) Specifically this buy and hold

4

For the technically minded the exact definition v assuming a holding interval of 5 years is

)# 200 , & +% ! w j5 ( 1 200. 1 5 " 1 .106 = +% . ( *$ j " 1 w j0 ' -

M. C. Jensen

5

1967

policy says invest equal dollar amount in each of the 200 securities and reinvest all cash proceeds from each security in that security when received, otherwise simply hold the portfolio until the end of the period with no redistribution. Levy claims the continuous reinvestment assumption eliminates the “nonrandomness” with the changing weights of securities in a buy and hold portfolio. What he does not tell the reader is that except for the case in which all security returns are identical, the returns calculated under the assumption of continuous redistribution will be less than the returns from a simple buy and hold policy. Thus, while one is free to choose any standard of comparison which is appealing, the author owes it to his audience to discuss the issues involved. Since a continuous redistribution policy is actually impossible to achieve in the real world, and since a simple buy and hold policy is a policy which could be implemented by a naive investor with no information, I for one prefer its use as the naive standard of

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

comparison. The returns on the buy and hold policy are calculated as the simple arithmetic average of the individual wealth relatives over the 5 year interval and then transformed to an annual compounding base. Fortunately, Levy has calculated the returns from the buy and hold policy for these 200 securities and reports the results elsewhere (Levy, 1966, p. 380) as 13.4% per year for the 5 year interval under consideration.5 The reader will note that the use of 13.4% per year as the returns from a no information naive investment strategy instead of the 10.6% obtained by the

5

The exact definition is:

) , 200 +! 1 $ ' w j5 . 1 5 ( 1 .134 = +# & . " % * 200 j ( 1 w j0 -

M. C. Jensen

6

1967

“geometric average” reduces the margin of superior profitability of the trading rules.6 We shall now consider Levy’s results in somewhat more detail. Levy’s main results are summarized in Table 1. First of all, it should be pointed out that the arithmetic averages of the 4-week returns presented by Levy are meaningless numbers. That is, knowing only the arithmetic average of a time series of sequential returns on a given security will not allow one to infer terminal value of an initial investment in that security. The terminal value will also be a function of the dispersion of the returns through time.7 The appropriate measure of the average 4week return would in this instance be the geometric average of these returns (which in this case can be calculated as the 13th root of unity plus the annual returns). But this yields no information not already contained in the annual returns so we may simply ignore the average 4-week returns presented by Levy. An examination of Table 1 indicates that the net annual returns from some of the

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

trading models appear to be substantially greater than the 13.4% returns of the buy and hold policy. In particular the variant of Model A with a “cast out rank” of 160 yielded 20% per year and Models B and C respectively yielded 26.1% and 29.1%. However, it also is apparent that the riskiness of all three of these models is substantially greater than the risk of the buy and hold portfolio (if we are willing to accept the standard deviation of four week returns as a measure of risk). We turn now to a discussion of some of the issues associated with these risk differentials and a suggested straightforward method of adjusting the returns to a comparable base.

6

It should be mentioned that the calculation (if the returns of 13.4% ignores all taxes on dividend? and commissions on their reinvestment, but the returns far the trading rules are calculated under the same assumptions. That is, the trading rules also assume all dividends and rights proceeds are reinvested in the securities from which received and ignore taxes and commissions on these transactions.

7

For a proof of this statement for the particular case in which returns are lognormally distributed cf. Aitchison (1957, pp. 7f).

M. C. Jensen

7

1967

Table 1 Net Returns Earned by Levy’s “X% Relative Strength” Portfolio Policies With Various “Cast Out Ranks”, the Returns on Levy’s “Geometric Average”, and the Returns on a Buy and Hold Policy1 Cast Out Ranks 020 050 100 150 160 180 195 Geometric Average Buy and Hold2

“10% Relative Strength”–Model A Net Annual Return Arithmetic Average 4-Week Returns Standard Deviation of 4-Week Returns Number of Securities in Ending Portfolio Percent of Ending Portfolio in 5 Largest Holdings -3.2 -0.19 4.87 NA NA 11.1 0.86 4.90 NA NA 16.3 1.22 4.39 NA NA 19.1 1.41 4.41 NA NA 20.0 NA3 4.60 72. 33. 17.8 1.35 4.55 NA NA 13.2 1.05 4.61 NA NA 10.6 0.83 3.52 200. 2.5 13.4 1.02 3.5 200. NA4

Other Models Model B5 For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420 Net Annual Return Standard Deviation of 4-Week Returns Number of Securities in Ending Portfolio Percent of Ending Portfolio in 5 Largest Holdings 26.1 5.40 45 52 Model C6 29.1 5.95 30 63

1. Except where noted all figures obtained from Levy (1967). 2. Obtained from Levy (1966), p. 380. 3. NA = not available. 4. Results reported by Levy (1966, p. 114) indicate that the 20 largest securities in the buy and hold portfolio accounted for 22% of the ending value. 5. Model B—”cast out rank” = 140, “5% relative strength” criteria. That is, the initial portfolio consists of an equal investment in the 5% of the securities having the highest C/A26 ratios and all proceeds from sales are reinvested equally in the 5% of the securities having the currently highest C/A26 ratios. (cf. Levy (1967)). 6. Model C—same as Model B except that to be eligible for selection those securities fulfilling the “5% relative strength” criterion must also rank among the 25 securities with the highest coefficient of variation in weekly closing prices over the immediately preceding 6 months.(cf. Levy (1967)).

(2)—The Definition and Treatment of “Risk” We have seen that a plausible argument can be made for measuring the returns on the standard of comparison as the returns on a naive buy and hold policy. For Levy’s data the buy and hold policy yields 13.4% per year. An examination of the net returns for the trading rules given in Table 1 indicates that this reduces the gap between the naive policy

M. C. Jensen

8

1967

and the trading rules but by no means eliminates it.8 These results are extremely interesting and it is in interpreting them that we must be very careful with regard to the treatment of “risk”. Postponing for the moment the exact definition of the term “risk”, let us consider the reasons for its great importance in problems of this sort. It is generally felt that the observed behavior of most people indicates they dislike risk. That is, the fact that most investors hold diversified rather than single asset portfolios, and the fact that most people purchase insurance (and thereby shift the risk of many types of losses to insurance companies) rather than bear the risks themselves indicate that people will pay money (or sacrifice returns) in order to avoid “risk.” Alternatively, this means that investors will not hold “risky” stocks unless they are compensated in the form of higher expected future returns than could be obtained from less risky securities. Hence, a highly risky asset (or portfolio of assets) must be expected to earn higher returns than a less risky asset or it would not be held. Thus, in evaluating

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

trading rules such as those proposed and tested by Levy we must be extremely careful to allow for all differences in risk between the standard of comparison (which is likely to be only moderately risky if as in this case it consists of a portfolio of 200 Securities) and the portfolio yielded by the trading rule. (Also it will be argued below that there are good reasons for expecting Levy’s trading rules to select predominately high risk portfolios during periods when the market is rising—which incidentally was the case in the 5-year period.) In the context of portfolio analysis the term “risk” is meant to refer to the uncertainty regarding the future returns to be earned from a particular investment or a particular investment policy. As such, of course, the risk of a particular investment or

8

Incidentally, in order that there be no confusion it should be made clear that the returns on the trading rules were calculated by taking the wealth ratio for each trading rule (the ratio of the ending value of the trading portfolio to the beginning value) and adjusting the return to an annual basis. (This definition was confirmed by Levy in personal communication with the author.)

M. C. Jensen

9

1967

investment strategy is strictly a subjective phenomenon and is not objectively measurable. However, a case can be made for using some measure of the dispersion of the distribution of returns through time as a proxy variable for the “true risk of an investment—the dispersion of the subjective probability distribution of returns over the holding interval under consideration. This has often been done in financial research, and Levy likewise provides us with measures of the standard deviation of 4-week returns for each of the trading rules and for the “geometric average.” Levy also reports elsewhere (1966, p. 380) the standard deviation of 4-week returns on the buy and hold policy. These results are also summarized in Table 1. Levy attempts to adjust for the differences in risk between the trading rules and the “geometric average” by a method which involves varying the fraction of the portfolio invested in securities (selected by the trading rule) and the fraction invested in a hypothetical bond yielding 4.75% according to whether or not the market did “well” or

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

“poorly” in the preceding 4-week period. If the market did “well” in the preceding 4week period the fraction invested in bonds is reduced and the fraction invested in the current trading rule portfolio is increased and vice versa if the market did “poorly”. The exact procedure is extremely arbitrary and the reader is referred to Levy (1967) for a detailed description of the procedure. These procedures for adjusting for risk are themselves subject to evaluation since they are in essence additional trading rules. In addition, in applying these variable ratio trading rules Levy essentially used future data regarding the performance of the market in arriving at the fractions invested in bonds and stocks. For this reason all the results on these models are extremely suspect. Hence we ignore them unless and until they are confirmed by a correct analysis which uses only past data available at the time of trading. We shall now suggest a straightforward manner of adjusting the returns of the valid trading models to a risk base comparable to the standard of comparison. Let us assume (as did Levy), the existence of a completely risk free asset (such as a bond or

M. C. Jensen

10

1967

insured savings account) yielding 4.75% during the 1960-65 test period.9 Now it is easily shown that if at each instant in time we keep a fraction X of our funds in a risky portfolio and the remaining fraction (1 – X) of our funds in the riskless asset, the standard deviation of the combined investment will be equal to the fraction X times the standard deviation of the risky portfolio. In addition, the returns on the combined investment will be equal to the weighted average of the returns on the portfolio and the riskless asset— the weights, of course, are X and (1 – X) respectively.10 Thus, let us take the most profitable of Levy’s “10% relative strength” trading rules (the one using a “cast out rank” of 160) which yields a net return of 20% unadjusted for the risk differential. The standard deviation of the 4-week returns for this rule as reported by Levy was 4.60. Recall now that the comparable standard deviation of the

9

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

Actually, such a riskless rate during the period was probably much closer to 4 % since a 5-year government bond (which is indeed a completely riskless investment for an investor intending to remain invested for 5 years) offered a yield to maturity in March, 1960 of 4.1%. (Obtained from yield curve in May, 1960, Treasury Bulletin.)

10

That is, let R! and R f be the returns on the risky portfolio and risk free asset respectively with standard deviations ! ( R" ) and ! ( R f ) . By assumption ! ( R f ) = 0 . The returns, R p , and standard deviation, in f are:

**! ( R p) , on the combined portfolio of 100 X% of the funds in r and 100 (1 - X) % of the funds
**

R p = X • Rr + (1 ! X ) R f 1 2 2 ! R p = x 2 • ! 2 ( Rr) + (1" X ) • ! 2 R f

( )

[

( )]

= X • ! ( Rr) since

! ( R f ) = 0.

It is important to note that the equations hold only when the returns R p Rr and R f are stated in terms of continuous compounding. If one desires R p stated on a basis of annual compounding the transformation is: e Rp N

[

]

where N is the number of years under consideration N is approximately equal to 5 in our case. The standard deviations in the equation for ! ( R p) are the standard deviations of the returns stated on a basis of continuous compounding. But since these will be virtually identical to the standard deviations of the 4week relatives we shall ignore the slight inconsistency involved and use Levy’s standard deviations as proxies for ! ( Rr ) .

M. C. Jensen

11

1967

4-week buy and hold returns was 3.50. Thus, at each instant of time if we keep a fraction X = 3.50/4.60 = .761 of our funds in the portfolio yielded by the trading rule and (1 - X) = .239 in the riskless asset yielding 4.75%, the returns on an annual basis on the combined investment would be 16.2%.11 The standard deviation of 4-week returns would of course be 3.50 and equal to the standard deviation of the buy and hold policy. The adjusted returns for models B and C are 18.1% and 18.5% respectively. The adjusted returns for the three models calculated under the somewhat more realistic assumption of a risk free rate of 4.1% (see footnote 9) are: Model A (160 “cast out rank”) = 16.0%, Model B = 17.9% and Model C = 18.2%. Thus it appears that even after adjustment to an equivalent risk base the returns on Levy’s most profitable trading rules are substantially higher than the 13.4% yielded by the buy and hold policy. It is this result which is extremely

**interesting. It is clear that if it is not due to random chance it is inconsistent with the
**

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

hypothesis that security prices behave according to a random walk and implies a significant amount of positive serial dependence in price changes. But, this brings us immediately to the third main point mentioned earlier. (3)—The Problems of Sampling Error and Selection Bias As noted above several of Levy’s trading models would have yielded substantially higher returns than a naive buy and hold policy even when adjusted to an equivalent risk level. This result is somewhat surprising since to the best of my knowledge no valid statistical study of the price movements of individual securities

11

Assuming for simplicity a holding interval of exactly 5 years the returns are calculated by: R p = (.761)(1ne 2.488) + (.239)(1ne 1.261) = .748

Putting this on a basis of annual compounding we have: .162 = e 748 5 ! 1

M. C. Jensen

12

1967

has ever revealed a substantial degree of dependence in price changes through time.12 Moreover, to the best of my knowledge no properly executed evaluation of mechanical trading rules based upon past security price movements has found them significantly profitable after allowance for transactions costs.13 Now let us accept the risk adjusted results of 16.0%, 17.9%, and 18.2% given above as valid returns for this particular sample. One must be extremely careful in making inferences from these results regarding the validity of the models that generated them. It is especially important that we pause at this time to consider some of the issues involved in making inferences of this type. That is, the wide usage of large computers and the availability of large bodies of data, coupled with the intense interest of many people investigating such trading rules, undoubtedly means there is and will be a great many studies of this type performed. I ask the reader to assume for the moment that the random walk hypothesis is

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

valid (and of course there is no way this can be proven in the strict sense of the word). Let us also assume that we have access to a large computer and a body of security price data. Now if we begin to test various mechanical trading rules on the data we can be virtually certain that if we try enough rules with enough variants we will eventually find one or more which would have yielded profits (even adjusted for any risk differentials) superior to a buy and hold policy. But, and this is the crucial question, does this mean the same trading rule will yield superior profits when actually put into practice? Of course not, since we assumed the random walk model was true in the first place. Thus, we cannot be certain that Levy’s results did not arise from mere chance alone. Indeed, a cursory examination of Levy’s dissertation (1966), (of which the

12 13

The reader is referred to the large number of studies reprinted in Cootner (1964) and to Fama (1965). See especially Fama (1966).

M. C. Jensen

13

1967

present article is an apparent outgrowth) indicates that he tested the performance of roughly 13 major types of trading models with a total of about 68 variants, and most of these models yielded lower returns than the buy and hold policy. Furthermore, all the models were apparently tested on the same body of data used in the present article. All this, coupled with the following facts, lead us to suspect the results may be due to chance alone: (1) The profitable models all seem to have surprisingly specific parameters (and are sensitive to them)14, and (2) The author apparently cannot supply any underlying economic rationale or explanation of the economic forces which cause the models to work in this particular period or any reasons why we should expect them to work in the future. In addition it should be pointed out that while Levy presents results for a number of models yielding superior profits, the reader must be careful not to interpret these models as independent observations since they certainly are very closely

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

related. Furthermore, it should also be noted that the success of all the superior models seems to be highly dependent on the selection of a very small number of very successful securities. That is, the data given by Levy and reproduced in Table 1, indicate that the percent of the ending trading portfolio in the 5 largest holdings is substantial in all cases ranging from 33% to 63%. These suspicions are confirmed by the additional data presented in Table 2 which were provided by Levy in personal communication with the author. Table 2 indicates that the five largest terminal holdings of Models B and C were identical and four of these securities were also represented in the five largest holdings of Model A. Thus, Levy’s results appear to be significantly influenced by the inclusion of the same few highly successful securities

14

For instance the 10% rule yields maximum profits with a “cast out rank” of 160 and the 5% rule yields maximum profits with a “cast out rank” of 140. Levy provides no reason for why this should be so.

M. C. Jensen

14

1967

in all the portfolios.15 It appears that the more complicated and more restrictive models are more successful not because they pick out additional superior securities but because they include fewer securities with “normal” returns.

Table 2 Five Largest Holdings In Ending Portfolio For Three Of Levy’s Trading Models* Size Rank 1 2 3 4 5 Model A Kresge Gen. Cable Bucyrus-Erie Burlington Joy Mfg. Model B Kresge Bucyrus-Erie Gen. Cable Joy Mfg. Admiral Model C Kresge Bucyrus-Erie Admiral Joy Mfg. Gen. Cable

*Data obtained from Levy in personal communication with the author.

**Since it is extremely difficult to perform the standard types of statistical tests of
**

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

significance on results of models like Levy’s (and indeed they would be invalid in the presence of possible selection bias anyway), we shall have to rely on the results of replications of the models on additional bodies of data and for other time periods. Incidentally, we all face problems like this when testing models for which a strong theoretical framework, specifying the exact form of the variables and the tests, is lacking. Thus, in these cases it is always a good idea to set aside a validation or replication sample from original data before starting. If possible half the original sample should be allocated for this purpose. Then we can feel quite free to be pure empiricists with the first sample and do all the “fishing” we please. Then, after settling on the final form of the model or models which seem to work best on the “fishing” sample, we test the models on the

15

We might also note that data provided by Levy to the author regarding the five largest holdings in the terminal portfolios of the three models presented in Levy’s (1967) Table 3 indicate only one common security—Admiral. However, since Admiral was the largest holding in each of these portfolios it appears their superior returns may be due only to this single security.

M. C. Jensen

15

1967

validation sample. If the results of these tests on the validation sample agree with the results of the first sample we can feel much more secure in the belief that the results were not mere random chance. On the other hand, if they do not agree we will not be misled by the results of one particular sample. (4)—The Treatment of Transaction Prices Levy’s method of calculating the returns on the trading models involves the implicit assumption that all trades can be executed at the same prices used in arriving at the buy and sell decisions. That is, Levy uses the Friday closing prices to generate the trading decisions and in calculating the returns he also assumes all trades were executed at the Friday closing prices. Of course, in practice this cannot be done. A better procedure would be to assume all trades were executed at the opening or closing price of the following Monday. Unfortunately, this complicates the data collection considerably. It

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

should be noted that if the random walk hypothesis were strictly correct, Levy’s procedure would not bias his results. However, a number of investigators have found evidence of a small amount of positive serial dependence in security price changes.16 The evidence indicates this dependence is so small that it cannot be used to make abnormal profits after allowance for normal transactions costs. This evidence does indicate, however, that Levy’s procedure will cause his returns to be somewhat higher than they would be if prices subsequent to the decision point were used for execution of the transactions. There is one other small point regarding the costs associated with implementing Levy’s rule which ought to be mentioned. That is, the buy and hold policy can be implemented with very little costs incurred by the investor during the holding period. However, implementation of Levy’s rules would require incurring expenses for the

16

See for example Alexander (1961; 1964) and Cootner (1964).

M. C. Jensen

16

1967

periodical collection of large amounts of security price data and their analysis. These expenses might very well be non-trivial. Allowing for them would also tend to decrease the superiority of the trading rules. We shall now consider a fundamental model of security price behavior and some issues associated with risk aversion and the pricing of capital assets which suggest an explanation for the fact that Levy’s rules systematically selected high risk portfolios in his sample period. The “Market Model” And The Capital Asset Pricing Model And Their Relationship To Levy’s Results It was mentioned above that the lack of a theoretical explanation of the success of the trading models is one of the reasons for seriously questioning whether Levy’s results are due only to chance (and therefore apply only to the sample he considered). On the

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

other hand there has recently been a considerable amount of investigation by academicians (at both the theoretical and empirical levels) of the behavior of stock market prices. These studies have yielded important insights which allow us to interpret Levy’s results in the context of some fundamental facts regarding the behavior of security prices and the effects of “risk” on the returns of any security or portfolio. The first of these models is the “diagonal” model originally suggested by Markowitz (1959, p. 100) and analyzed in considerable detail at the theoretical and empirical levels by Sharpe (1963; 1964; 1967), Blume (1967), Fama (1967a, b), Fama, Fisher, Jensen and Roll (1967). Jensen (1967), King (1966) and Scholes (1967). The more descriptive term “market” model has been suggested by Fama (1967a, b) and is used here. The second of these models is the capital asset pricing model apparently derived independently by Sharpe (1964), Treynor (1961) and Lintner (1965), recently extended by Fama (1967a, b) and Jensen (1967), and tested and utilized in the evaluation of mutual fund portfolios by Jensen (1967).

M. C. Jensen

17

1967

In its simplest form the market model simply says that, R, the returns on any security for any time period can be expressed as a linear function of the returns on a “market index” or a “market factor,” M. The relationship is (1) R = a + bM + u

where R is the return on any security and M is the return on a “market factor”. (In practice the returns on the “market factor” for any period are estimated by using some average of the returns on all securities for that period.) The coefficients a and b are constants which may vary from security to security and u is a random error term which has a mean of zero. In words equation (1) simply says that the returns on any security for any particular time period are equal to a constant “a” plus some fraction “b” (which may be greater than l) of the returns on the market factor for that time period plus a random

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

component u which is specific to each security and on average equals zero. The reader familiar with regression analysis will immediately recognize eq. (1) as a simple linear regression model. Detailed empirical tests of the model on securities on the New York Stock Exchange by Blume (1967), Fama, et. al. (1967), and Scholes (1967) indicate that it is indeed a valid regression model fulfilling most of the usual statistical assumptions. In addition, the results of these tests indicate that the average constant, a , close to 0 and the average slope coefficients, b , is close to unity, and all securities have “b” coefficients which are greater than zero—that is, in general the prices of all securities tend to move in the same direction. Now within the context of the market model, what can we expect to happen if we select securities for a portfolio on the basis of Levy’s C/A26 ratio—the ratio of the security’s current price to its average price over the immediately preceding 26 weeks? If the market return has been substantially positive in the previous six months, we expect, on average, that those securities with the highest b coefficients will have increased in

M. C. Jensen

18

1967

value the most, thus in general yielding the highest C/A26 ratios. Indeed this would seem to be the case for Levy’s particular sample, since data presented by Levy elsewhere (1966, pp. 378 ff) indicates that for 48 of the 58 non-overlapping 4-week periods covered by his analysis, the S&P 500 Stocks yielded positive returns. Thus given the fact that the market factor M was generally positive in this particular period, it is not surprising that a method which tends to select high b stock yields returns which are above the average for all stocks in the sample. But again the question of risk becomes crucial and while it will now be argued that the high b stocks are precisely the high risk stocks, there is no reason for them to systematically earn returns higher than those implied by their level of risk. It is at this point that the results of the theory of capital asset pricing become extremely important, and we shall attempt to give a non-technical summary of the important aspects of the model. The reader interested in a detailed (and necessarily mathematical) treatment of the model is urged to consult the references mentioned

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

earlier.17 Essentially the model is derived under the assumption that all investors are averse to risk and that they obey the normative theory of portfolio selection put forth by Markowitz (1959). Then under the assumption that the capital markets are in equilibrium it can be shown that the returns on each security will be a linear function of the “systematic risk” of that security. In its simplest form the function is (2) R = R F + (M ! RF )b + e

where as before R and M are respectively the returns on any security and the returns on the market factor. The variable RF is the riskfree rate of interest for the time period under consideration (the returns on a government bond or an insured savings account) and the variable e is a random error term which has a mean value of zero. The coefficient b is the

17

Especially Fama (1967a, b), Jensen (1967), Sharpe (1964).

M. C. Jensen

19

1967

measure of the “systematic risk” of the security and surprisingly enough, for all practical purposes, turns out to be equivalent to the slope coefficient b in the market model of eq. (1).18 It should be emphasized that the relationship between the coefficient b in the market model and the concept of systematic risk which arises out of the theory of capital asset prices is not assumed; it simply arises out of the capital asset pricing model as the result of the basic assumptions regarding the behavior of investors and the existence of equilibrium in the capital markets. Hence, if the capital asset pricing model is a valid description of the manner in which the returns on assets are related to their risk (and the empirical evidence in Jensen (1967) suggests it is), the portfolios selected by Levy’s trading rules will be high risk portfolios. This implication is supported by the high standard deviations of returns for the trading rule portfolios. However, it should be emphasized again that while these considerations explain why Levy’s trading rules select high risk (and therefore high

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

return) portfolios they do not explain the superior returns earned by Levy’s rules after adjusting for risk differences. As before the question regarding the persistence of these results must be left to the results obtained by replication of the tests on other data and for other time periods. At this point the reader may wonder about the validity of the method used earlier for adjusting the returns on the trading models for the risk differential. That is, if we believe the asset pricing model, it would seem that we would be led to the use of the coefficient b (which can be estimated by usual least squares procedures) as the measure of risk. The answer to this seeming inconsistency involves a relatively subtle point regarding the definition of “efficiency” and the difference between an asset’s “systematic risk” and its total risk.

18

Certain technical but empirically unimportant issues associated with equation (2) have been omitted here. The reader interested in the complete development and discussion of these technicalities is urged to consult Fama (1967a, b; 1966).

M. C. Jensen

20

1967

The capital asset pricing model is derived under the assumption that all investors hold only “efficient” portfolios in the Markowitz (1959) sense. That is, an efficient portfolio is defined to be one which has maximum expected return for a given level of risk and minimum risk for given level of expected return. It is easily shown that the total risk of an asset is equal to its systematic risk plus its non-systematic or insurable risk. Thus, since by properly diversifying, an investor can avoid bearing the non-systematic risk of any particular asset in his portfolio, it turns out that the systematic component is the only portion of the total risk which should influence the price of the asset. But, if we are considering the risk and return of a particular portfolio or portfolio policy we must consider the total risk of that portfolio since it may not be efficiently diversified (and therefore its total risk will be greater than its systematic risk). It is for this reason that we earlier adjusted the returns on the trading rules for the differences in the total risk of the portfolios. This is an admittedly subtle and confusing point and, since a complete

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

discussion is not warranted here, the technically minded reader is urged to consult Jensen (1967) for a complete and rigorous discussion of the issues and proofs of the statements. Summary And Conclusions The main purpose of this comment has been to discuss some of the problems involved in evaluating trading models of the type proposed by Levy. It has been shown that a good case can be made which implies the magnitude of Levy’s results are somewhat overstated as a result of his definition of the returns on a “random selection policy,” his treatment of risk differentials, his implicit assumption regarding the identity of the decision and trading prices and his neglect of the costs of operating the rules. In addition there are strong indications that the remaining superiority of his trading rules is the result of a subtle type of selection bias. It was also shown that two basic economic models of the behavior of security prices (which are consistent with the random walk

M. C. Jensen

21

1967

hypothesis) aid us in interpreting why Levy’s trading rules selected high risk portfolios during the time period he considered. However, it should be emphasized that these qualifications do not entirely explain all of Levy’s results and we are left in the position of having to wait and see if they are validated by additional tests on other data and for other time periods. References

Aitchison, J. and J. A. C. Brown. 1957. The Lognormal Distribution. Cambridge: Cambridge University Press. Alexander, Sidney S. 1961. “Price Movements in Speculative Markets: Trends or Random Walks.” Industrial Management Review II: May, pp 7-26. Alexander, Sidney S. 1964. “Price Movements in Speculative Markets: Trends or Random Walks, Number 2.” Industrial Management Review V: pp 25-46. Blume, Marshall. 1967. “The Assessment of Portfolio Performance”. Chicago, IL: University of Chicago. Unpublished Ph. D. dissertation.

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

Cootner, Paul H., ed. 1964. The Random Character of Stock Market Prices. Cambridge, MA: MIT press. Fama, Eugene F. 1965. “The Behavior of Stock Market Prices.” Journal of Business 37: January 1965, pp 34-105. Fama, Eugene F. 1967a. “Capital Asset Prices: Some Clarifying Comments”. Chicago, IL: University of Chicago. Unpublished manuscript. Fama, Eugene F. 1967b. “Risk, Return, and General Equilibrium in a Stable Paretian Market”. Chicago, IL: University of Chicago. Unpublished manuscript. Fama, Eugene F. , Lawrence Fisher, Michael C. Jensen, and Richard Roll. 1967. “The Adjustment of Stock Prices to New Information”. Report No. 6715, University of Chicago: Center for Mathematical Studies in Business and Economics. Forthcoming in the International Economic Review. Fama, Eugene F. and Marshall Blume. 1966. “Filter Rules and Stock Market Trading.” Journal of Business 39: pp 226-41. Jensen, Michael C. 1967. “Risk, the Pricing of Capital Assets, and Evaluation of Investment Portfolios”. Chicago, IL: University of Chicago. Unpublished preliminary draft of Ph. D. thesis. King, Benjamin F. 1966. “Market and Industry Factors in Stock Price Behavior.” Journal of Business 39, Part 2: pp 139-190.

M. C. Jensen

22

1967

Levy, Robert A. 1966. An Evaluation of Selected Applications of Stock Market Timing Techniques, Trading Tactics and Trend Analysis. Unpublished Ph. D. dissertation, the American University. Levy, Robert A. 1967. “Random Walks: Reality or Myth.” Financial Analysts Journal November/December. Lintner, John. 1965. “Security Prices, Risk, and Maximal Gains from Diversification.” Journal of Finance 20: December, pp 587-616. Markowitz, Harry. 1959. Portfolio Selection: Efficient Diversification of Investments. Cowles Foundation Monograph No. 16. New York: Wiley & Sons, Inc. Scholes, Myron. 1967. A Test of the Competitive Market Hypothesis: An Examination of the Market for New Issues and Secondary Offerings. Unpublished Ph. D. dissertation, University of Chicago. Sharpe, William F. 1963. “A Simplified Model for Portfolio Analysis.” Management Science 19: September, pp 425-442. Sharpe, William F. 1964. “Capital Asset Prices: A Theory of Market Equilibrium under Conditions of Risk.” Journal of Finance 19: September, pp 425-442. Sharpe, William F. 1967. “Linear Programming Algorithm for Mutual Fund Portfolio Selection.” Management Science 13: pp 499-510.

For an electronic copy of this paper, please visit: http://ssrn.com/abstract=350420

Treynor, Jack L. 1961. “Toward a Theory of Market Value of Risky Assets”. Unpublished manuscript, undated.

random walk theory

random walk theory

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue listening from where you left off, or restart the preview.

scribd