Professional Documents
Culture Documents
North-Holland
This study of 862 press recommendations demonstrates that the size effect can distort longer-term
performance measures, and hence event studies. Relative to similar sized companies, post-publica-
tion performance is neutral. However, market adjustments, the CAPM and Market Model, with
equally or capitalization weighted indexes, all produce biased results. Event studies are most
exposed to such bias when the measurement interval is long, event securities differ systematically
in size or weighting from the index constituents, the size effect is large and/or volatile, and when
CAPM-type methodologies are used. These distortions are avoided by explicitly controlling for
size.
1. Introduction
*We are grateful for the helpful comments of Dick Brealey, Stewart Hodges, Clifford Smith.
Marc Reinganum (the referee), and finance workshop participants at the London Business School.
Stockholm School of Economics, European Institute for Advanced Studies in Management
(Brussels), European Finance Association (Manchester), and the Western Finance Association
(Vancouver). We would also like to thank Ramin Hakimneiad, Tamara Harrison and staff of the
Consumers Association for data collection and verification. Financial support was received from
the Consumers Association.
For example, if you had invested fl,OOO in the 1957 Nap Selections, and reinvested the
proceeds at the end of each year in the new annual selections, your initial f1.000 would now be
worth f251.211 (before gains tax and expenses) against a mere f1.841 if you had invested in the
FT Index. (IC Newsletter advertisement, Financial Times, November 1979)
E. Dimon and P. Marsh, Ecent srudies and rhe sr:e effect 115
Brown (1984)], and a wide variety of other factors. The most important
empirical regularity so far observed is the size effect (the tendency for small
capitalisation stocks to outperform their larger counterparts) and the tum-of-
the-year effect (the tendency for outperformance to occur at the beginning of
the tax year). Consequently, as Beaver (1981) and others point out, an event
study which focuses on smaller (larger) firms is likely to witness positive
(negative) abnormal returns relative to the market index; this is potentially
exacerbated by event date clustering at the turn of the year.
At first sight, Beavers view seems at variance with the evidence from the
Brown and Warner (1980,1985) simulation experiments. However, Brown and
Warner are careful to point out that their studies are based on randomly
selected samples which are designed not to have marked exposure to extra-
market factors. Methodologies which are reliable for such representative
groups of securities may nevertheless perform poorly for samples which differ
substantially from the market index, and which, in the absence of an event,
exhibit abnormal performance. Where these problems are likely to be severe,
Schwert (1983) points out that event studies should give explicit consideration
to the size effect. Yet very few studies make any attempt to do so.
Schwert (1983) also argues that when the market model is used the size
effect is not a problem. The assumption here is that market model alpha
estimates encapsulate any size effect. However, the use of historical alphas
introduces considerable noise, and both Brenner (1979) and Brown and
Warner (1980) find the market model the least efficient of the standard event
study methodologies. More seriously, we know that the small-firm effect varies
considerably over tune and, for US data, exhibits a turn-of-the-year seasonal
[see Keim (1983) and Brown, Kleidon and Marsh (1983)]. Thus, if events are
clustered in a particular calendar time period, the mean size effect may differ
between the estimation and predictions periods, introducing bias into the
market model alphas. A similar problem will arise from seasonality in the
event dates, if this coincides with a seasonal in the size effect. For these
reasons we cannot assume that the market model fully resolves the problems
introduced by the size effect.
Like its American counterpart, the UK size effect is very marked. The
difference between the compound annual rates of return on the Dimson-Marsh
(1985) equal weighted and capitalisation weighted indexes was some 7 percent
over the 30 years (1955-84) covered by the London Share Price Database
(LSPD). The magnitude of the small-firm premium varies considerably from
year to year: Levis (1985) estimates that it is nearly as volatile as the return on
the underlying small stocks, though he tinds no evidence of US-type seasonal-
ity. While the UK size effect remains an empirical regularity for which only
partial explanations have been offered [see Reinganum and Shapiro (1984) and
Dimson and Marsh (1984b)], it is nevertheless too large to be dismissed, for it
may well dominate the results of empirical studies of stock returns.
116 E. Dimon and P. Marsh, Eoenr studies and the si:e eflect
Sunday Express tips were given as shares for the decade. Our sample is
distributed fairly evenly over time, with most years containing around 100
tips. [See Dimson and Marsh (1948b) for more details.]
Three index series are used in the study. Our initial proxy for the market, in
estimating both betas and abnormal returns, is the capitalisation weighted
FTA Index. Index returns are calculated by incorporating the published
dividend yield [see Marsh (1979)]. As an alternative market proxy, we use the
Dimson-Marsh (1985) equally weighted (EW) Index of monthly returns
(including di vi d en d s) on all UK listed shares. As a benchmark for measuring
size-adjusted performance, we also employ the Dimson-Marsh (1985) size
decile indexes. These measure the returns on the ten percent of companies
with the highest market value, the next decile by market value, and so on to
the ten percent of companies with the lowest capitalisation. The capitalisation
deciles are constructed by ranking all stocks in the EW Index by their
end-November capitalisations. Ranked stocks are assigned to deciles, and
equally weighted returns are computed for each decile over the twelve-month
period commencing with the end-December quotation, employing the same
methodology used to calculate the EW Index. By construction, therefore, the
average of the decile returns equals the return on the EW Index.
Returns on the event securities are calculated using monthly data from the
LSPD, utilising middle-market quotations prices, exactly as in construction of
the EW and size decile indexes. The initial performance of the recommended
stocks is computed relative to the (specially collected) selection-date price,
while subsequent performance is evaluated using month-end prices.3
The betas of recommended stocks are estimated relative to both the capi-
talisation weighted FTA and the EW Index, using two different estimation
methods. First, betas are computed relative to the FTA using the trade-to-trade
method to avoid thin trading bias. This involves calculating returns from dated
transaction prices, and regressing these on index returns measured over the
same calendar period. This requires daily observations on a market index
which suffers from negligible thin trading [see Marsh (1979)]. Because the FTA
Index meets this criterion [see Dimson (1979)], this method is expected to
produce efficient, unbiased beta estimates [see Dimson and Marsh (1983)j.
Unfortunately, this approach cannot be used to estimate betas relative to the
EW Index, since directly observed daily EW Index values are not available.
For the purpose of computing abnormal returns from the intra-month recommendation date to
the month-end, we estimate a series of daily values for the EW and size decile indexes. We regress
the monthly index returns on the FTA, apportioning each months residual return equally to each
trading day within the month, and then add this to the expected index return, conditional on the
daily FTA index return [see Dimson and Marsh (1984a. sect. 3.2)].
3 When returns are computed from the selection date to the end of the recommendation month,
the period in fact runs to the end-month trading day immediately following publication. Thus, the
initial performance of a tip selected on. say, Friday, 29 December and published on 30 December
would be evaluated to the last trading day of January; and similarly, post-publication returns
would be measured from the end of January until a subsequent month-end.
E. Dimson and P. Marsh, Ecent studiesand the size effect 119
(Even if they were. they would suffer from significant thin trading bias, since
the quotations for smaller companies are not continuously updated.) The
aggregated coefficients method, which provides unbiased, but less efficient,
estimates is therefore selected as the next best alternative. Following Dimson
(1979) we use one leading and five lagged market return terms.
Stocks may attract the attention of tipsters after a period of unusual
performance. To avoid bias in parameter estimates of beta, and of course
alpha, we must exclude all prior months in which abnormal returns occur.
Since returns might reasonably be expected to trigger a recommendation
within a year, we initially exclude data for the twelve-month pre-tip period,
estimating alphas and betas over the 60 months prior to this exclusion period.
We also examine the impact of both lengthening and shortening the exclusion
period (see section 4.3.4 below).
3. Research design
same way as the securities in the market index. u,, will also be unbiased if the
benchmark expected return adequately controls for the size effect. Finally,
even when bias is present, its magnitude will depend on the length of the
observation period. Over very short periods, such as the days or even the
month immediately following an event, bias from benchmark misspecification
is likely to be small, relative to event-related returns and noise. As the
observation period is extended, however, the bias will be magnified, so that
over longer periods, it becomes dominant and assumes (apparently) economi-
cally significant proportions.
If the size effect is in evidence, then serious bias in longer-term performance
measures can potentially arise either from an improper choice of methodology
or from an unsuitable choice of index or weighting scheme. For example, the
choice between the CAPM or market model, or between an equally or market
value weighted index, is no longer simply a question of balancing theoretical
considerations against statistical efficiency. Rather, this decision becomes a
key component of the experimental design.
Two types of benchmark are likely to suffer from the biases described
above. The first is a market index which fails to match the size exposure of the
event securities. The second is a risk-adjusted version of this, whereby the
index is adjusted to have the same beta as the event securities. The latter is
illustrated by the Sharpe (1964) - Lintner (1965) CAPM, by Blacks (1972)
two-factor model and by the Watts (1978) - Banz (1981) arbitrage portfolio
techniques. AI1 these beta-adjusted methods suffer in the same way from the
size effect, and we proxy them by the CAPM.
Given the choice between the capitalisation weighted FTA and its equally
weighted counterpart, our first four approaches are thus as follows:
In these equations, R,s, is the return on Treasury bills, .R,,, and R,,, are
the returns on the PTA and EW Indexes respectively, and /3,,,, and &. are the
security betas estimated relative to these two indexes (see section 2.2 above).
These equations represent widely used, standard event study methodologies
which assume a zero size effect. They differ simply in whether they use the
capitalisation weighted PTA [eqs. (2) and (3)] or the equally weighted EW
Index [eqs. (4) and (5)], and in whether beta is assumed equal to unity [eqs. (2)
and (4)] or is estimated using OLS [eqs. (3) and (5)].
(sizeandbeta).
In eq. (6), the abnormal return on stock j is measured as the difference
between the stocks return and the return on its control portfolio, RiCjjsr,
represented by the size decile which includes constituents with approximately
the same capitalisation as the event security. In eq. (7), this method is modified
by blending the control portfolio with the market portfolio plus Treasury bills,
to replicate the beta of the event securities. The abnormal return therefore
equals the difference between the CAPM-adjusted performances of stock j
and of portfolio i(j), thus purging the CAPM method [eq. (3)] of any size bias
in the u~,~.~
The size control portfolio approach ensures that E(u~,,]@~_,) = 0, where
QX_, incorporates information on the stocks capitalisation. The estimated
abnormal returns thus reflect the performance of similar sized companies
during the period (s, t).
4This is consistent with the genera&d asset pricing model suggested by Banz (1981), in which
expected excess returns are postulated as a linear function of both beta and capitalisation. Other
forms of size control may also be devised. For example, the u,,, may be specified as the difference
between R,, and the return on the control portfolio, where the weights given to the control
portfolio constituents are set so as to equate the portfolio beta to that of the event security [see
Banz (1981)]. These variants produce similar results to eq. (7) and are not analysed here.
122 E. Dmson and P. Marsh. Event studies and the si:e effect
If the size effect were constant over time, it would be unnecessary and
unduly restrictive to devise control portfolios. An alternative would be to use
Sharpes (1963) single index market model (SIMM). With a constant size
effect, the superior performance of small stocks would be reflected in their
estimation period returns and hence in their alpha estimates. These estimates
could then be used to compute prediction period expected returns which
encapsulate the size effect. The expected value of the mean abnormal return
(u,,) should be zero, since each set of stocks acts as its own control. This will
still hold true with a stochastic size effect, as long as its average magnitude is
the same over the (set of) estimation and prediction periods. This should tend
to be the case for events spread randomly over a very long calendar period.
We implement this approach using both the capitalisation weighted FTA
and the equally weighted EW Index. By regressing event security returns on
the chosen index and assuming the parameters to be stationary, we estimate
the expected return conditional on index movements. Denoting the alpha of
stock i relative to, the FTA by &+,, the predicted return on stock i in period T
is simply I?~,,,+ ,B/,,,Rq7,. The difference between the actual and predicted
return is R,,, - B,, - /3,,R,,,. If this prediction error is cumulated over time,
we obtain the prediction error for stock j over the period (s, t):
uj,, = n (1 + Rj,
T-S
- ;yie - bj,R...) - 1 (SIMM using EW) . (9)
Approaches other than the market model [i.e., eqs. (2) to (7)) measure the
economic gain from holding the event securities in preference to (or in
conjunction with a short position in) an alternative benchmark portfolio. In
each case, the benchmark represents a viable investment strategy, comprising
pre-specified holdings in one or more published indexes, blended, if necessary,
with borrowing or lending. By contrast, the market model benchmark does not
conform to any possible, pre-specifiable investment strategy, and market
model abnormal returns can be interpreted only by invoking notional agree-
ments between hypothetical investors [see Scholes (1972)]. The SIMM is thus
primarily a statistical device, and is more suited to testing for the existence of
abnormal returns than to establishing their economic magnitude.
Other variations on the market model approach involve forcing the betas to
take specific values. For example, if betas are forced to be zero, this is
equivalent to Masulis (1980) mean adjusted returns method. This approach
E. Dmson and P. Marsh, Ecenr studies and rhe sr:e effect 123
has strengths and weaknesses which are similar to those of the SIMM and
consequently it is not evaluated here.
The annual observation periods were also chosen to mitigate any seasonal small-firm or
turn-of-the-year effect [see Keim (1983) and Lakonishok and Smidt (1984)]. In practice, however.
the UK size premium displayed no such seasonality over the period spanned by this study [see
Dimson and Marsh (1984b)].
Reweighting involves reducing the holding in stocks which have apparently appreciated and
increasing the holding in stocks which have apparently depreciated. When price measurement
errors are later rectified, the portfolio return is enhanced as a result of the previously reduced
holding in (spuriously) overpriced stocks, and the previously increased holding in (spuriously)
underpriced stocks.
124 E. Dimon and P. Marsh. Event studies and the six effect
It is possible that the abnormal performance measures, u,,. exhibit serial dependence, either
because of thin trading bias, or because New Year tips occur in both December and January, so
that event month t for one stock can correspond to event month t - 1 for another. However, we
ftnd no evidence of serial dependence and hence make no adjustment to eq. (11) [see Ruback
(1983)]. Similarly, because the portfolio constituents scarcely change in event time, and because
the pre-recommendation history for some stocks is short, the benefit from estimating JatTe
(1974) - Mandelker (1974) standard errors is small or non-existent.
E. Devon and P. Marsh, EIXW studies and the see effect 125
Z *I = 0,*//m. (12)
4. Results
Table 1 summarises the performance of recommended stocks evaluated
using all eight methodologies described in section 3.2. The monthly pre- and
post-recommendation performance figures are listed in the appendix, while
post-publication performance is plotted in fig. 1 below.
We begin by discussing the prior year and tip month performance (sections
4.1 and 4.2). Since our qualitative conclusions here are insensitive to the choice
of methodology, we focus, for ease of exposition, on the first two methodolo-
gies, namely simple market adjustment [eq. (2)] and the CAPM [eq. (3)] using
the capitalisation weighted FTA. We defer our discussion of the differences
between methodologies to section 4.3, when we evaluate post-publication
performance.
J.FE.- E
126 E. Damon and P. Marsh. Ecenr studies and the x1-e effect
Table 1
Abnormal performance of recommended stocks using alternative benchmarks and methods of
averaging.
Performance
Cumulative percentage returns (and t-statistics)
Row benchmark and Method of
IlO. equation number averaging Prior year Tip month One year Two years
1. FTA Index only (2) Equal 14.8 (7.0) 4.6 (10.1) 2.5 (1.4) 6.9 (2.2)
2. CAPM using FTA (3) Weighting 15.1 (7.2) 4.7 (10.5) 3.2 (1.8) 8.5 (2.8)
3. FTA Index only (2) Capitahsation - 1.2 (0.3) - 0.7 (0.8) - 2.1 (0.6) - 5.4 (0.9)
4. CAPM using FTA (3) Weighting - 0.1 (0.0) - 0.5 (0.6) - 1.7 (0.5) - 3.9 (0.7)
5. EW Index only (4) Equal 1.9 (3.5) 3.7 (7.7) - 5.6 (2.9) - 14.1 (4.0)
6. CAPM using EW (5) Weighting 3.5 (1.6) 3.3 (7.2) - 7.2 (3.8) - 18.9 (5.6)
7. Size control only (6) Equal 12.4 (7.4) 3.9 (10.8) -0.5 (0.4) - 1.3 (0.5)
8. Size and beta (7) Weighting 11.1 (6.5) 3.9 (10.6) -0.4 (0.3) - 1.3 (0.5)
9. SIMM using FTA (8) Equal 4.7 (2.4) 4.4 (7.7) - 2.6 (1.4) -4.8 (1.8)
10. SIMM using EW (9) Weighting 5.3 (3.1) 3.3 (6.8) - 3.2 (2.0) - 6.4 (2.8)
The prior year return is an average of 862 abnormal returns measured over the period from 12
months before the tip month to the date of the tip. The tip month return is an average of 792
abnormal returns on buy recommendations plus 70 abnormal returns from taking a short position
in stocks recommended for sale. measured from the date of the tip to the end of the tip month.
The one-year and two-year returns are also based on 792 long positions plus 70 short positions.
and are measured from the end of the tip month until 12 or 24 months later. The abbreviations
used here are as follows: FTA denotes the capitalisation weighted Financial Times-Actuaries All
Share Index; EW denotes the Dimson-Marsh (1985) Equally Weighted Index of all UK equities;
CAPM denotes the Capital Asset Pricing Model; SIMM denotes the Single Index Market Model.
The second column of table 1 shows that, between the recommendation date
and the end of the publication month, the tips achieve an abnormal return of
some 4 percent. Indeed, only three of the twelve sources of tips record perverse
performance. If financial journalists do not cheat (for example by withdraw-
ing a tip from publication after the selection date) and if deals are executed at
the middle market prices used in computing returns, then this could be
interpreted as a violation of market efficiency.
The magnitude of the abnormal performance is, however, relatively small,
and any potential short-term profits from following tipsters recommendations
would be more than consumed by UK transaction costs. Furthermore,
At the time of this study, most deals incurred round trip commissions of at least 1.5 to 2
percent, plus a 2 percent tax (stamp duty) on all purchases. In addition, the bid-ask spread is
estimated in work in progress by the authors to range from 1 percent for larger stocks to 10
percent or more for very small, infrequently traded securities. These trading costs are higher than
those found for large US stocks by Phillips and Smith (1980). but are comparable with the figures
for smaller NYSE and AMEX stocks documented by Schultz (1983). Interestingly, the highest tip
month abnormal returns, even after adjusting for size, are recorded for the tipsters who tend to
recommend smaller companies. These are precisely the companies for which the bid-ask spread.
and hence transaction costs. are the greatest.
E. Dmson and P. Marsh. Event studies and the sr:e effect 127
investors cannot expect to deal at the price quoted by the journalist. Jobbers
(floor traders) read the recommendations, and adjust prices before any post-
publication trading takes place. It is therefore possible that the subsequent
quotation represents a switch from an average of the bid-ask prices to an
effective ask-only price. If so, we expect to witness a swift reversal of the initial
abnormal returns once the impact of publication is abated. In fact, we observe
neutral performance in the month after publication, both relative to the FTA
and using alternative methodologies (see appendix). Thus the 4 percent
abnormal return in the tip month appears to be attributable to the journalists
selection skills.
Several days may elapse between selection by the journalist and publication.
During this period, the underlying information content may find its way into
market prices. This is best illustrated by the portfolio tips made by Mr.
Bearbull. This weekly magazine becomes available to the public on Friday.
The publishers take delivery from the printers the previous day, and the
recommendation (press) date is three days before this. A study of daily returns
over the seven days following recommendation reveals a 4.1 percent abnormal
return [see Dimson and Marsh (1984b)). Of this, one third is realised prior to
the publishers taking delivery, and almost two thirds before the journal is
distributed to the public.
Either tipsters tend to recommend stocks which are expected to experience a
swift flow of favourable information, or they acquire information which leaks
to the market. Alternatively, there may be investment opportunities for the
tipsters themselves, although the editors of the periodicals represented here
indicated that pre-publication dealing would constitute grounds for dismissal.
Whatever the explanation, it is clear that this short-term performance is not
available to the public. If investors follow tipsters advice, they must buy at the
post-publication price.
The third column of table 1 shows that over the one-year period from the
end of the recommendation month, the mean abnormal return using either of
the FTA based methodologies is around 3 percent [2.5 percent with simple
market adjustments [eq. (2)] and 3.2 percent using the CAPM [eq. (3)]. This
rate of outperformance is maintained over the following twelve months, and
the two-year abnormal gain is 6.9 percent with simple market adjustments and
8.5 percent using the CAPM. The two-year outperformance is significant at the
5 percent level, and over this horizon, all but three publications achieved
positive abnormal returns.
128 &. Dimson and P. Marsh. Event srudies and the xe effect
5 -
5 -
-10 -
-15 -
-~.I,,,
0 2 4 6 8 10 12 I4 16 18 W 22 24
Monlh alter recommendalion
Fig. 1. Cumulative performance of 862 stocks recommended during 1975-82, evaluated using
eight alternative benchmarks. The abbreviations used here are as follows: FTA denotes the
capitalisation weighted Financial Times-Actuaries All Share Index: EW denotes the
Dimson-Marsh (1985) Equally Weighted Index of all UK equities; CAPM denotes the Capital
Asset Pricing Model; SIMM denotes the Single Index Market Model.
It seems unlikely that British tipsters can predict such substantial abnormal
returns, with as much accuracy for months 13 to 24 as for months 1 to 12. In
addition, the pattern of month-by-month cumulative abnormal returns dis-
plays a persistent upward drift rather than sudden jumps (see fig. 1). These
observations would, if methodologically correct, represent a violation of
market efficiency. One possible explanation is that these results occur by
chance, and our significance estimates are misleading, For example, serial
dependencies in mean abnormal returns might cause us to underestimate their
variance, so that significance levels are overstated. This explanation may be
tested by restricting calculations to recommendations evaluated over non-over-
lapping periods. Over the post-publication year, the New Year tips are
cross-serially independent, except for a possible one-month overlap. The
initial, one-year and (for completeness) two-year abnormal gains of the
restricted sample are 5.9, 4.1 and 14.0 percent, respectively, and the r-statistics
are virtually unchanged (7.6, 1.3 and 2.5, respectively) in spite of the reduced
sample size. Hence this explanation is rejected. Similarly, there is no evidence
E. Dimon and P. Marsh. Ecent smdies and Ihe size effect 129
in our sample have over 8,000 times the weight of the ten smallest. Clearly,
giving capitalisation weighting to event security returns is statistically ineffi-
cient, in spite of its apparent consistency attractions. An obvious alternative.
which meets the same consistent weighting criterion, but provides greater
statistical efficiency, is to revert to equal weighting for event security returns
and re-evaluate results against an equally weighted index. Although previous
British research has employed the capitalisation weighted FTA as the best
available index, equally weighted indexes have frequently been employed in
American studies.
Rows 5 and 6 of table 1 present the abnormal returns computed using the
EW Index as the benchmark [eqs. (4) and (5)]. Betas as well as returns were
estimated using this index. The results reinforce the above conclusions:
recommended stocks significantly underperform the EW Index by over 14
percent over the two years after publication, and by 19 percent on a risk
adjusted basis. This dramatic downward shift when abnormal returns are
measured against the EW Index rather than the FTA is reflected in the results
for all twelve sources of tips, with nine of them now exhibiting inferior
performance. Far from displaying selection skills, the analysts now appear
perverse.
We thus have evidence of both statistically significant overperformance by
recommended stocks relative to the FTA, and statistically significant under-
performance relative to the EW Index. In both cases, the abnormal returns are
larger in absolute magnitude the longer the post-publication period examined
(see the appendix). As Ball (1978) points out, when trading strategies offer
apparent abnormal returns which increase steadily over time, it suggests a
deficiency in the model of expected returns. Our results suggest that the
CAPM is an inappropriate benchmark for this event study. This is true
whether the model uses estimated betas, or assumes betas of unity; an equally
or capitalisation weighted index; or equally or capitalisation weighted event
security returns.
The major misspecification is the failure of the benchmark to reflect the size
composition of the sample. While tipsters tend to recommend larger compa-
nies, the probability of a stock being tipped is much less than proportional to
its capitalisation (see section 2.1). Hence a capitalisation weighted index is just
as misleading as an equally weighted index. Similarly, capitalisation weighting
event security returns places even greater weight on larger stocks than the
Note that risk adjustment has a greater impact when using the EW Index than with the FTA.
This is due to the higher excess return on the EW Index over the sample period (43 percent over
the two-year post-publication period, compared with 22 percent for the FTA), coupled with the
fact that the mean beta of stocks against the EW Index is 1.16. This high mean beta reflects the
above average capitalisations of recommended stocks, and the positive correlation between size
and beta for UK securities [see Dimson and Marsh (1983)].
E. Dmson and P. Mursh. Ewnr srudes and rhe see efecr 131
Row 7 of table 1 presents the abnormal returns computed using the simple
size control portfolio approach. Over all periods, the switch to size adjustment
pulls the abnormal returns to an intermediate figure between that obtained
using the FTA (rows 1 and 2) and the EW Index (rows 5 and 6). The largest
changes occur for tipsters who tend to recommend smaller firms.
Our general conclusions on pre-recommendation and tip month perfor-
mance are relatively unaffected by size adjustment. The outperformance in the
pre-tip year remains substantial at 12 percent. Similarly, size adjustment
reduces the tip month abnormal return only slightly from 4.7 percent (CAPM
with FTA) to 3.9 percent, and this figure remains statistically significant at the
0.1 percent level. The size adjusted post-publication results, however, cast an
entirely new light on our earlier conclusions. Performance is virtually neutral,
disclosing evidence neither of longer-term forecasting skills nor perversity.
Over the one- and two-year post-publication periods, recommendations under-
perform similar sized companies by only 0.5 and 1.3 percent, respectively.
Neither figure is statistically significant. Row 8 of table 1 shows that these
abnormal returns remain virtually unchanged when we adjust for beta as well
as size [eq. (7)l.t
Therefore, when abnormal returns are measured on a size adjusted basis,
there is no evidence of market inefficiency. Furthermore, in spite of substantial
performance differences between publications, none achieve statistically sig-
nificant outperformance over the one- and two-year post-publication periods.
The cross-sectional dispersion of tipsters post-publication performance is thus
probably attributable to sampling variation. It would be inappropriate to
claim that an advisory column with a favourable record can confidently be
expected to repeat its success.
The last two rows of table 1 show the CARS estimated from the SIMM
using either the FTA or EW Index [eqs. (8) and (9)]. The tip month abnormal
This similarity between the results in rows 7 and 8 could be because the risk adjustment
procedure is ineffective in eliminating the influence of the market from the size adjusted returns.
However, a regression of size and beta adjusted abnormal returns on the corresponding excess
returns on the market discloses no relationship between the two. In fact, risk adjustment adds
little to the size based procedure because the betas of recommended securities are representative of
the stocks capitalisation deciles [see Dimson and Marsh (1984b)]. It is possible that for other
studies, risk adjustment would play a larger role.
132 E. Dimson and P. Mursh, Ecent studies and the si:e effect
return of around 4 percent is close to the figures obtained with the other
methods. In a single month, the alpha contributes relatively little, so that the
SIMM results resemble the corresponding CAPM results based on the FTA
and EW Index (rows 2 and 6). Over longer periods, however, the alpha plays a
more important role, causing market model results to deviate from their
CAPM counterparts (see fig. 1). Longer-term SIlMM performance is largely
insensitive to the choice of index [as in Brown and Warner (1980)], and the
post-publication CAR is -3 percent over one year, and - 5 to -6 percent
over two years.
Since, on a size adjusted basis, post-publication performance is neutral, the
SIMM results seem somewhat surprising. The general expectation in using the
market model is that each sample will act as its own control, with the mean
alpha estimate encapsulating factors such as the size effect. On the other hand,
the SIMM two-year performance figures are statistically significant at the 10
percent (FTA Index) and 1 percent (EW Index) levels, whereas the corre-
sponding CAPM results (rows 2 and 6) are more extreme and more significant,
at the 1 and 0.1 percent levels. In this study, the market model avoids some of
the bias which aflIicts the CAPM-type methodologies, but the SIMM results
are not as neutral as those obtained with size adjustment.
By definition, the SIMM takes account of the samples sensitivity to a single
factor, the market return, while also incorporating an estimate of the mean
non-market return per period via the constant term, alpha. If there are other,
extra-market factors at work, such as the size effect, then the SIMM is
misspecified. The magnitude of any resulting bias depends on the stochastic
nature of the omitted variables (we focus here simply on the size effect). It also
depends on the stability of any omitted coefficient, or in this case, the samples
sensitivity to the size factor, as measured by the average size of the event
securities. At one extreme, if the size effect is constant, and the average
(relative) size of the event securities remains stable over time, then the
estimated SIMM alphas wilI purge abnormal returns of all size-related bias.
On the other hand, if the size effect is volatile and/or non-stationary, then the
mean alphas will vary over time. Unless event dates are spread evenly over a
As Brown and Warner (1980) point out, one would expect market model results to be
insensitive to the choice of index. However, as their empirical results attest, this may not hold true
for simple market adjustments or CAPM-type, beta adjusted methods. They examine a randomly
selected sample of events with equal weight given to each, and tind that a capitafisation weighted
index incorrectly rejects the null hypothesis a whopping 20.4 percent of the time. Controlling for
beta reduces the rejection rate to 14 percent, but the latter would have been much higher with a
larger sample size (their sample contains only 50 events) and an observation period with a greater
size effect (their events span 1944-71 when the size effect was relatively small). Brown and
Warners conclusion that an improper use of the value weighted index was shown to cause
considerable problems which have not been recognised in extant event studies is consistent with
our findings here.
E. Drmon and P. Mush. Ecent studies and the st:e effect 133
very long sample period, this can lead to biased performance measures.
Similarly, bias may arise if the average size of the event securities changes
between the estimation and prediction periods.
Given that our sample differs from the indexes in terms of size composition,
our results may therefore reflect variability in the size effect. Certainly, the UK
size effect does exhibit considerable year-to-year variation. When proxied by
the difference between the annual returns on the EW and FTA Indexes, it
ranges from a low of -49 percent (in 1975) to a high of +42 percent (in 1977)
during the period spanned by this study. If the size effect is generally larger
during the estimation period than in the prediction period, this could explain
our results. While recommendations are spread roughly evenly across years,
problems could still arise from starting and ending effects. In fact, however.
the average monthly size effect during the five-year estimation period was 0.47
percent per month, while the corresponding figure during the two-year post-
publication period was 10 basis points larger at 0.57 percent per month. If
anything, this should cause tipsters to slightly outperform the SIMM
benchmark.
An alternative explanation is that the average sensitivities of our sample
securities to the size factor may vary systematically between the estimation
and prediction periods. l3 For example, if event securities experience positive
abnormal returns over the (pre-event) estimation period, the alphas will be
biased. At the same time, this would be accompanied by an increase in relative
size for the sample stocks, making them, on average, larger in the prediction
than in the estimation period. This would accentuate the upward bias in the
alphas, resulting in spurious negative abnormal returns over the post-event
period. If correct, this implies that the exclusion period is inappropriately
chosen. However, we verified that the results are insensitive to whether we
halve or double the length of the exclusion period. In this sense, our results
appear robust. The bias to which we refer is more subtle than a simple failure
to exclude the immediate pre-recommendation period in which event-related
abnormal returns occur.
*The bias arises because event date clustering causes the weights given to various calendar
periods to differ between the pre- and post-event periods. This problem would be exacerbated by a
US-type seasonal in the size effect, particularly if event dates cluster around the turn of the year.
As already noted, however, the UK size premium exhibits no such seasonal over our sample
period.
13This is analogous to the well-documented problems arising from systematic changes in beta
[e.g.. see Mandelker (1974)l. Indeed, the negative abnormal returns in our own studv could arise
from betas shifting over time in an even; related manner. For example, if tipsters tend to
recommend stocks which they feel are becoming relatively less risky, then betas in the estimation
period will overstate betas in the prediction period, leading to apparent underperformance against
the SIMM benchmark. Overestimation by 0.15 to 0.20 would be sufficient to explain away the
observed results. However, we can reject this explanation, since the mean post-publication beta is
in fact underestimated (by 0.05) during the estimation period.
134 E. Dmson und P. Marsh, Event rruf~es und the see efect
The SIMM cannot reveal whether an estimation period alpha arises from
size (or other) factors or from event related abnormal returns. However, by
assuming the validity of a factor based model of the return generating process,
we can investigate whether the sample securities exhibit abnormal returns after
accounting for this size effect. We therefore estimate the average size adjusted
returns of the event securities, using eq. (8), over six non-overlapping pre-
recommendation periods of one year. Over the pre-tip year, the abnormal
return is 12.4 percent, while over the five years of the estimation period, the
abnormal returns are always positive, and average 3.4 percent per year. By
contrast, for a randomly selected set of securities, the abnormal return over
these years averages zero.
This evidence suggests that the SIMM alpha estimates are biased due to an
abnormal return over and above the size effect during the estimation period.
This implies that the sample stocks will have grown larger in (relative) size by
the prediction period, thus further accentuating the upward bias. This pre-event
abnormal return may be a consequence of an inevitable form of selection or
survivorship bias arising from estimating the SIMM parameters for only those
stocks which have (by definition) survived to the event date. Alternatively, it
may indicate that tipsters tend to recommend stocks which have performed
well over a very long prior period.
Unfortunately, we can find no exclusion period, not even one of seven years
or more, which leaves us with unbiased alphas according to our size adjusted
criterion. Some authors have sought to avoid these selection bias problems
associated with pre-event data by estimating SIMM parameters from post-event
data [see Copeland and Mayers (1982)]. Unfortunately this future benchmark
technique precludes making timely statements about investment performance.
Since our most recent event dates are only three years old, it would be
necessary to wait until 1989 to make statements about their two-year perfor-
mance. Nor does the future benchmark technique circumvent problems
arising from variability in the size effect. The best we can do in the context of
the market model is therefore to report the results based on the one-year
exclusion period, while recognising that this figure is a downward biased
estimate.
This poses serious problems when choosing an exclusion period, Ball (1972) suggests succes-
sively omitting months until any abnormal returns appear to be contained entirely within the
exclusion period. However, if there is a size effect, and the sample has deviant size exposure, this
procedure may reach a local optimum, or never converge. Indeed, Ball (1972) had to exclude 158
months of data, and Marsh (1979) could find no exclusion period which met this criterion. Even if
the criterion is met, this may be because event related, abnormal returns counterbalance the size
effect; and when the exclusion period is long, there can be large differences between the mean
capitalisations of the event securities in the estimation and prediction periods. In both instances,
the alphas will be biased.
E. Dmson and P. Marsh. Ecent studies and thesi:e effect 135
Lakonishok and Vermaelen (1983)], and many other well-known event studies
[e.g.. Kaplan and Roll (1972), Gonedes, Dopuch and Penman (1976) and
seven of the fourteen studies in Jensens (1983) Symposium on the market for
corporate control]. Yet if stocks experience an event with a probability less
than proportional to their capitalisation, then when there is a size effect, a
capitalisation weighted index can always be expected to underperform an
equally weighted portfolio of event securities. However, we have yet to identify
any published study in which the probability of an event is at least propor-
tional to capitalisation.
Similarly, an equally weighted index will produce biased results if event
securities are larger or smaller than the typical index constituent. Again, this
appears to be the rule, not the exception. For example, stock recommendations
and acquisitions are more prevalent for large companies, while initial public
offerings and financial distress are typically experienced by small firms. In
addition, event securities may be non-random with respect to size because of
sample selection requirements (e.g. the availability of prior stock price histo-
ries or supplementary accounting data), pre-event abnormal returns (as with
stock splits or rights issues), and the partitioning criteria employed (e.g.,
acquirers versus acquirees, or large versus small changes in dividends or
earnings). Cross-sectional regressions of abnormal returns on size-related
company characteristics can also suffer from all of the problems posed by
partitioning [see Jain (1982)]. Clearly, size mismatching is a potential problem
in many event study settings.
Third, the severity of any bias will depend on the magnitude and volatility
of the size effect. Event studies carried out in countries and periods in which
the size effect is negligible will avoid these problems. This seems unlikely to be
the norm, however, given the considerable evidence that the worlds stock
markets are characterised by substantial small firm effects, which vary consid-
erably over time, ls Event studies spanning short periods, or in countries with
narrowly based stock markets, are likely to suffer most from bias (e.g.,
Holland and Switzerland, where the three largest companies represent over
half of the equity market value). In countries such as the USA, where the size
effect has a seasonal, analogous problems may arise, particularly when events
cluster around the turn of the year. Similarly, event studies may experience
problems from extra-market factors other than size, in markets such as South
Africa, Australia and Norway where gold, natural resources and oil companies
are heavily represented.
See, for example, Banz (1981) and Brown, Kleidon and Marsh (1983) for the USA; Dimson
and Marsh (1985) and Reinganum and Shapiro (1984) for the UK; Brown, Keim, Kleidon and
Marsh (1983) for Australia: Berges, McConnell and Schlarbaym (1982) for Canada; Van den
Bergh and Wessels (1983) for Holland; Wahlroos and Berglund (1983) for Finland; and Nakamura
and Terada (1984) and Kato and Schallheim (1985) for Japan.
E. Dimon and P. Mush. Evenr studies and the xe effecr 137
Appendix
This appendix reports the average of the abnormal returns on the 862
recommended stocks. Up to the day of tip, the return is based on holding all
862 stocks. From the day of tip onwards, the return is based on holding the
792 stocks recommended for purchase plus a short position in the 70 stocks
recommended for sale. Columns (2) to (9) refer to the eight alternative
methods for computing abnormal returns, i.e., eqs. (2) to (9), respectively.
Table 2
Cumulative percentage returns for 862 tips during 1976-82, using eight performance
benchmarks.
- 13 0 0 0 0 0 0 0 0
-12 0.7 0.7 0.5 0.0 0.5 0.2 0.7 0.5
-11 3.0 2.9 3.4 2.0 3.1 2.4 3.5 3.5
-10 3.3 3.3 4.3 2.3 3.3 2.4 3.2 3.7
-9 3.0 3.1 2.6 0.5 2.3 1.4 3.0 3.2
-8 3.9 4.0 3.4 0.9 3.0 2.0 2.9 3.0
-7 4.0 4.1 2.3 -0.3 2.4 1.4 2.5 2.3
-6 2.9 3.2 - -0.1 - 2.9 1.0 0.1 2.0 1.7
- 5 4.3 4.7 1.0 - 2.2 2.6 1.7 1.9 1.8
- 4 5.7 5.9 2.9 -0.7 4.4 3.2 1.9 2.5
-3 7.4 7.6 2.7 - 1.0 5.5 4.3 1.9 1.8
-2 10.2 10.4 4.9 0.6 8.4 7.1 2.7 3.0
- 1 12.0 12.3 6.1 1.7 10.1 8.8 3.2 4.1
Dav of tio 14.8 15.1 7.9 3.5 12.4 11.1 4.7 5.3
E. Dun.wn and P. Marsh. Ecenr studies and the rr:e effecr 139
Table 2 (continued)
Day of tip 0 0 0 0 0 0 0 0
0 4.6 4.7 3.1 3.3 3.9 3.9 4.4 3.3
0 0 0 0 0 0 0 0 0
1 0.2 0.3 -0.2 -0.3 0.0 0.0 -0.1 -0.2
2 -0.1 0.3 -0.9 - 1.3 -0.5 - 0.4 -0.5 -0.9
3 -0.1 0.5 -0.9 - 1.7 -0.6 -0.4 -0.8 -1.1
4 0.2 0.8 - 1.9 -2.7 -1.0 -0.9 - 0.9 - 1.6
5 0.4 0.9 - 2.3 -3.0 -1.1 - 0.9 - 1.2 - 1.7
6 1.0 1.4 - 2.0 -2.9 -0.7 -0.6 - 0.6 -0.8
7 1.0 1.7 - 2.1 -3.1 - 0.6 -0.5 - 1.1 -1.0
8 0.4 1.2 -3.1 -3.8 -1.1 -0.8 - 2.0 - 1.4
9 0.6 1.4 - 4.3 -4.8 -1.3 - 0.9 - 2.6 - 2.2
10 0.5 1.2 -4.6 -5.3 -1.1 -0.8 -3.3 - 2.8
11 1.8 2.5 -4.7 - 5.6 - 0.6 -0.4 -3.1 - 2.8
12 2.5 3.2 -5.6 - 7.2 -0.5 -0.4 - 2.6 - 3.2
13 3.1 4.0 -6.2 - 8.2 -0.6 -0.5 - 2.0 -3.2
14 2.5 3.4 - 7.9 - 10.3 - 1.7 - 1.7 - 2.8 -4.2
15 2.4 3.4 - 8.4 - 11.3 - 2.0 - 2.1 - 2.8 - 4.0
16 2.8 3.9 - 9.6 - 12.4 - 2.0 - 2.0 -3.3 -4.7
17 3.3 4.4 - 9.8 - 12.7 - 1.9 - 1.8 -3.6 - 4.8
18 3.9 5.2 - 10.3 - 13.1 - 1.6 - 1.4 -3.5 -4.7
19 3.3 4.7 -11.3 - 14.5 - 2.3 -2.2 -4.1 -4.8
20 3.5 4.9 - 11.7 - 14.9 -2.0 - 2.0 - 5.0 -5.2
21 4.2 5.5 - 12.6 - 15.9 -1.7 - 1.7 - 5.6 -6.0
22 4.8 6.2 - 12.2 - 15.8 - 1.1 -1.0 - 5.7 -5.7
23 5.0 6.4 - 14.0 - 17.8 - 1.8 -1.8 -6.1 -6.7
24 6.9 8.5 - 14.1 - 18.9 - 1.3 -1.3 -4.8 -6.4
The abbreviations used here are as follows: FTA denotes the capitalisation weighted Financial
Times-Actuaries All Share Index; EW denotes the Dimson-Marsh (1985) Equally Weighted
Index of all UK equities; CAPM denotes the Capital Asset Pricing Model; SIMM denotes the
Single Index Market Model.
References
Ball. R., 1972, Changes in accounting technique and stock prices, Empirical Research in
Accounting, Journal of Accounting Research 10 (suppl.), l-38.
Ball. R.. 1978, Anomalies in relationships between securities yields and yield-surrogates, Journal
of Financial Economics 6, 103-126.
Banz. R.W., 1981, The relationship between return and market value of common stock, Journal of
Financial Economics 9, 3-18.
Barry, C.B. and S.J. Brown, 1984, Differential information and the small firm effect, Journal of
Financial Economics 13.283-294.
Basu, S., 1983, The relationship between earnings yield, market value and the return for NYSE
common stocks, Journal of Financial Economics 12, 129-156.
Beaver, W.H., 1981, Econometric properties of alternative security return methods, Journal of
Accounting Research 19, 163-184.
140 E. Damson and P. Marsh. Ecenr srudies and the sr:e efect
Gonedes, N.J., N. Dopuch and S.H. Penman. 1976, Disclosure rules, information-production, and
capital market equilibrium: The case of forecast disclosure rules, Journal of Accounting
Research 14, 89-137.
Jaffee. J.. 1974. The effect of regulation changes on insider trading. Bell Journal of Economics 5,
93-121.
Jain, P.C., 1982. Cross-sectional association between abnormal returns and firm specific variables,
Journal of Accounting and Economics 4, 205-228.
Jensen, M.C., 1978, Some anomalous evidence regarding market efficiency, Journal of Financial
Economics 6, 95-101.
Jensen. M.C., ed., 1983, Symposium on the market for corporate control: The scientific evidence,
Journal of Financial Economics 11.
Kaplan. R.S. and R. Roll, 1972, Investor evaluation of accounting information: Some empirical
evidence, Journal of Business 45, 225-257.
Kato, K. and J.S. Schallheim, 1985, Seasonal and size anomalies in the Japanese stock market,
Journal of Financial and Quantitative Analysis 20, 243-260.
Keim. D.B., 1983, Size related anomalies and stock return seasonahty: Further empirical evidence.
Journal of Financial Economics 12, 3-32.
Lakonishok, J. and S. Smidt, 1984, Volume and turn-of-the-year behaviour. Journal of Financial
Economics 13.435-455.
Levis, M., 1985, Are small firms big performers?, The Investment Analyst 76, 21-27.
Lintner, J., 1965, The valuation of risk assets and the selection of risky investments in stock
portfolios and capital budgets, Review of Economics and Statistics 47, 13-37.
Litzenberger, R.H. and K. Ramaswamy, 1979, The effect of personal taxes and dividends on
capital asset pricing theory and empirical evidence. Journal of Financial Economics 7.163-195.
Litzenberger, R.H. and K. Ramaswamy, 1982, The effect of dividends on common stock prices:
Tax effects or information effects, Journal of Finance 37, 429-443.
Mandelker. G., 1974. Risk and return: The case for merging firms. Journal of Financial
Economics 7, 110-120.
Marsh, P.R., 1979, Equity rights issues and the efficiency of the UK stock market, Journal of
Finance 34, 839-862.
Masulis. R., 1980, The effects of capital structure change on security prices: A study of exchange
offers, Journal of Financial Economics 8, 139-177.
Miller, M.H. and M. Scholes, 1982, Dividends and taxes, Journal of Financial Economics 6,
333-364.
Nakamura, R. and N. Terada, 1984, The size effect and seasonality in Japanese stock returns
(Nomura Research Institute, Tokyo).
Phillips. S.M. and C.W. Smith, 1980, Trading costs for listed options: The implications for market
efficiency, Journal of Financial Economics 8, 179-201.
Reinganum, M.R., 1981, Misspecification of capital asset pricing: Empirical anomalies based on
earnings yields and market values, Journal of Financial Economics 9, 19-46.
Reinganum, M.R. and A.C. Shapiro, 1984, Taxes and stock return seasonality: Evidence from the
London Stock Exchange, Working paper (University of Southern California, Los Angeles, CA).
Roll. R., 1983, On computing mean returns and the small firm premium, Journal of Financial
Economics 12, 371-386.
Ruback, R.S., 1983, Assessing competition in the market for corporate acquisitions, Journal of
Financial Economics 11, 141-153.
Scholes. M.. 1972, The market for securities: substitution versus price pressure and the effects of
information on share prices, Journal of Business 45, 179-211.
Schultz. P., 1983, Transactions costs and the small firm effect: A comment, Journal of Financial
Economics 12, 81-88.
Schwert, G.W., 1983, Size and stock returns and other empirical regularities, Journal of Financial
Economics 12, 3-12.
Seymour Pierce, 1984. Annual survey of new year tips (Seymour Pierce & Co., London).
Sharpe. W.F., 1963, A simplified model for portfolio analysis, Management Science 9, 277-293.
Sharpe. W.F., 1964, Capital asset prices: A theory of market equilibrium under conditions of risk,
Journal of Finance 19, 425-442.
Stanley. K.L., W.G. Lewellen and G. Schlarbaum, 1980, Investor response to investment research,
142 E. Dumon and P. Marsh. O>ent rtuhes and thesi:e effect