The Risk and Return of Venture Capital: Article in Press

ARTICLE IN PRESS
Journal of Financial Economics 75 (2005) 352

www.elsevier.com/locate/econbase
The risk and return of venture capital$

John H. Cochrane
Graduate School of Business, University of Chicago, 5807 S. Woodlawn, Chicago, IL 60637, USA
Received 30 April 2003; received in revised form 14 July 2003; accepted 12 March 2004
Available online 14 October 2004
Abstract
This paper measures the mean, standard deviation, alpha, and beta of venture capital
investments, using a maximum likelihood estimate that corrects for selection bias. The biascorrected estimation neatly accounts for log returns. It reduces the estimate of the mean log
return from 108% to 15%, and of the log market model intercept from 92% to 7%: The
selection bias correction also dramatically attenuates high arithmetic average returns: it
reduces the mean arithmetic return from 698% to 59%, and it reduces the arithmetic alpha
from 462% to 32%. I conrm the robustness of the estimates in a variety of ways. I also nd
that the smallest Nasdaq stocks have similar large means, volatilities, and arithmetic alphas in
this time period, conrming that the remaining puzzles are not special to venture capital.
Published by Elsevier B.V.
JEL classification: G24
Keywords: Venture capital; Private equity; Selection bias
$
I am grateful to Susan Woodward, who suggested the idea of a selection-bias correction for venture
capital returns, and who also made many useful comments and suggestions. I gratefully acknowledge the
contribution of Shawn Blosser, who assembled the venture capital data. I thank many seminar participants
and two anonymous referees for important comments and suggestions. I gratefully acknowledge research
support from NSF grants administered by the NBER and from CRSP. Data, programs, and an appendix
describing data procedures and algebra can be found at http://gsbwww.uchicago.edu/fac/john.cochrane/
research/Papers/.
Corresponding author.
E-mail address: [email protected] (J.H. Cochrane).
0304-405X/$ - see front matter Published by Elsevier B.V.

doi:10.1016/j.jneco.2004.03.006
ARTICLE IN PRESS
4
J.H. Cochrane / Journal of Financial Economics 75 (2005) 352
1. Introduction
This paper measures the expected return, standard deviation, alpha, and beta of
venture capital investments. Overcoming selection bias is the central hurdle in
evaluating such investments, and it is the focus of this paper. We observe valuations
only when a rm goes public, receives new nancing, or is acquired. These events are
more likely when the rm has experienced a good return. I overcome this bias with a
maximum-likelihood estimate. I identify and measure the increasing probability of
observing a return as value increases, the parameters of the underlying return
distribution, and the point at which rms go out of business.
I base the analysis on measured returns from investment to IPO, acquisition, or
additional nancing. I do not attempt to ll in valuations at intermediate dates. I
examine individual venture capital projects. Since venture funds often take 23%
annual fees and 2030% of prots at IPO, returns to investors in venture capital
funds are often lower. Fund returns also reect some diversication across projects.
The central question is whether venture capital investments behave the same way
as publicly traded securities. Do venture capital investments yield larger riskadjusted average returns than traded securities? In addition, which kind of traded
securities do they resemble? How large are their betas, and how much residual risk
do they carry?
One can cite many reasons why the risk and return of venture capital might differ
from the risk and return of traded stocks, even holding constant their betas or
characteristics such as industry, small size, and nancial structure (leverage, book/
market ratio, etc.). First, investors might require a higher average return to
compensate for the illiquidity of private equity. Second, private equity is typically
held in large chunks, so each investment might represent a sizeable fraction of the
average investors wealth. Finally, VC funds often provide a mentoring or
monitoring role to the rm. They often sit on the board of directors, or have the
right to appoint or re managers. Compensation for these contributions could result
in a higher measured nancial return.
On the other hand, venture capital is a competitive business with relatively free
(though not instantaneous; see Kaplan and Shoar, 2003) entry. Many venture capital
rms and their large institutional investors can effectively diversify their portfolios.
The special relationship, information, and monitoring stories that suggest a
restricted supply of venture capital might be overblown. Private equity could be
just like public equity.
I verify large and volatile returns if there is a new nancing round, IPO, or
acquisition, i.e., if we do not correct for selection bias. The average arithmetic return
to IPO or acquisition is 698% with a standard deviation of 3,282%. The distribution
is highly skewed; there are a few returns of thousands of percent, many more modest
returns of only 100% or so, and a surprising number of losses. The skewed
distribution is well described by a lognormal, but average log returns to IPO or
acquisition still have a large 108% mean and 135% standard deviation. A CAPM
estimate gives an arithmetic alpha of 462%; a market model in logs still gives an
alpha of 92%.
ARTICLE IN PRESS
The selection bias correction dramatically lowers these estimates, suggesting that
venture capital investments are much more similar to traded securities than one
would otherwise suspect. The estimated average log return is 15% per year, not
108%. A market model in logs gives a slope coefcient of 1.7 and a 7:1%; not
+92%, intercept. Mean arithmetic returns are 59%, not 698%. The arithmetic alpha
is 32%, not 462%. The standard deviation of arithmetic returns is 107%, not
3,282%.
I also nd that investments in later rounds are steadily less risky. Mean returns,
alphas, and betas all decline steadily from rst-round to fourth-round investments,
while idiosyncratic variance remains the same. Later rounds are also more likely to
go public.
Though much lower than their selection-biased counterparts, a 59% mean
arithmetic return and 32% arithmetic alpha are still surprisingly large. Most
anomalies papers quarrel over 12% per month. The large arithmetic returns result
from the large idiosyncratic volatility of these individual rm returns, not from a
2
large mean log return. If s 1 (100%), em1=2s is large (65%), even if m 0:
Venture capital investments are like options; they have a small chance of a huge
payoff.
One naturally distrusts the black-box nature of maximum likelihood, especially
when it produces an anomalous result. For this reason I extensively check the
facts behind the estimates. The estimates are driven by, and replicate, two central sets
of stylized facts: the distribution of observed returns as a function of rm age, and
the pattern of exits as a function of rm age. The distribution of total (not
annualized) returns is quite stable across horizons. This nding contrasts strongly
with the typical pattern that the total return distribution shifts to the right and
spreads out over time as returns compound. A stable total return is, however, a
signature pattern of a selected sample. When the winners are pulled off the top
of the return distribution each period, that distribution does not grow with time.
The exits (IPO, acquisition, new nancing, failure) occur slowly as a function of rm
age, essentially with geometric decay. This fact tells us that the underlying
distribution of annual log returns must have a small mean and a large standard
deviation. If the annual log return distribution had a large positive or negative mean,
all rms would soon go public or fail as the mass of the total return distribution
moves quickly to the left or right. Given a small mean log return, we need a large
standard deviation so that the tails can generate successes and failures that grow
slowly over time.
The identication is interesting. The pattern of exits with time, rather than the
returns, drives the core nding of low mean log return and high return volatility. The
distribution of returns over time then identies the probability that a rm goes public
or is acquired as a function of value. In addition, the high volatility, rather than a
high mean return, drives the core nding of high average arithmetic returns.
Together, these facts suggest that the core ndings of high arithmetic returns and
alphas are robust. It is hard to imagine that the pattern of exits could be anything
but the geometric decay we observe in this dataset, or that the returns of individual
venture capital projects are not highly volatile, given that the returns of traded small
ARTICLE IN PRESS
6
growth stocks are similarly volatile. I also test the hypotheses a 0 and ER 15%
and nd them overwhelmingly rejected.
The estimates are not just an artifact of the late 1990s IPO boom. Ignoring all data
past 1997 leads to qualitatively similar results. Treating all rms still alive at the end
of the sample (June 2000) as out of business and worthless on that date also leads to
qualitatively similar results. The results do not depend on the choice of reference
return: I use the S&P500, the Nasdaq, the smallest Nasdaq decile, and a portfolio of
tiny Nasdaq rms on the right-hand side of the market model, and all leave high,
volatility-induced arithmetic alphas. The estimates are consistent across two basic
return denitions, from investment to IPO or acquisition, and from one round of
venture investment to the next. This consistency, despite quite different features of
the two samples, gives credence to the underlying model. Since the round-to-round
sample weights IPOs much less, this consistency also suggests there is no great return
when the illiquidity or other special feature of venture capital is removed on IPO.
The estimates are quite similar across industries; they are not just a feature of
internet stocks. The estimates do not hinge on particular observations. The central
estimates allow for measurement error, and the estimates are robust to various
treatments of measurement error. Removing the measurement error process results
in even greater estimates of mean returns. An analysis of inuential data points nds
that the estimates are not driven by the occasional huge successes, and also are not
driven by the occasional nancing round that doubles in value in two weeks.
For these reasons, the remaining average arithmetic returns and alphas are not
easily dismissed. If venture capital seems a bit anomalous, perhaps similar traded
stocks behave the same way. I nd that a sample of very small Nasdaq stocks in this
time period has similarly large mean arithmetic returns, largeover 100%
standard deviations, and large53%!arithmetic alphas. These alphas are
statistically signicant, and they are not explained by a conventional small-rm
portfolio or by the Fama-French three-factor model. However, the beta of venture
capital on these very small stocks is not one, and the alpha is not zero, so venture
capital returns are not explained by these very small rm returns. They are similar
phenomena, but not the same phenomenon.
Whatever the explanationwhether the large arithmetic alphas reect the
presence of an additional factor, whether they are a premium for illiquidity, etc.
the fact that we see a similar phenomenon in public and private markets suggests
that there is little that is special about venture capital per se.
2. Literature
This papers distinctive contribution is to estimate the risk and return of venture
capital projects, to correct seriously for selection bias, especially the biases induced
by projects that remain private at the end of the sample, and to avoid imputed
values.
Peng (2001) estimates a venture capital index from the same basic data I use, with
a method-of-moments repeat sales regression to assign unobserved values and a
ARTICLE IN PRESS
reweighting procedure to correct for the still-private rms at the end of the sample.
He nds an average geometric return of 55%, much higher than the 15% I nd for
individual projects. He also nds a very high 4.66 beta on the Nasdaq index.
Moskowitz and Vissing-Jorgenson (2002) nd that a portfolio of all private equity
has a mean and standard deviation of return close to those of the value-weighted
index of traded stocks. However, they use self-reported valuations from the survey of
consumer nances, and venture capital is less than 1% of all private equity, which
includes privately held businesses and partnerships. Long (1999) estimates a
standard deviation of 24.68% per year, based on the return to IPO of nine
successful VC investments.
Bygrave and Timmons (1992) examine venture capital funds, and nd an average
internal rate of return of 13.5% for 19741989. The technique does not allow any
risk calculations. Venture Economics (2000) reports a 25.2% ve-year return
and 18.7% ten-year return for all venture capital funds in their database as of 12/21/
99, a period with much higher stock returns. This calculation uses year-end values
reported by the funds themselves. Chen et al. (2002) examine the 148 venture
capital funds in the Venture Economics data that had liquidated as of 1999. In these
funds they nd an annual arithmetic average return of 45%, an annual compound
(log) average return of 13.4%, and a standard deviation of 115.6%, quite similar
to my results. As a result of the large volatility, however, they calculate that one
should only allocate 9% of a portfolio to venture capital. Reyes (1990) reports
betas from 1.0 to 3.8 for venture capital as a whole, in a sample of 175 mature
venture capital funds, but using no correction for selection or missing intermediate
data. Kaplan and Schoar (2003) nd that average fund returns are about the
same as the S&P500 return. They nd that fund returns are surprisingly persistent
over time.
Gompers and Lerner (1997) measure risk and return by examining the investments
of a single venture capital rm, periodically marking values to market. This sample
includes failures, eliminating a large source of selection bias but leaving the survival
of the venture rm itself and the valuation of its still-private investments. They nd
an arithmetic average annual return of 30.5% gross of fees from 19721997. Without
marking to market, they nd a beta of 1.08 on the market. Marking to market, they
nd a higher beta of 1.4 on the market, and 1.27 on the market with 0.75 on the small
rm portfolio and 0.02 on the value portfolio in a Fama-French three-factor
regression. Clearly, marking to market rather than using self-reported values has a
large impact on risk measures. They do not report a standard deviation, though
pone
can infer from b 1:4 and R2 0:49 a standard deviation of 1:4 16= 0:49
32%: (This is for a fund, not the individual projects.) Gompers and Lerner nd an
intercept of 8% per year with either the one-factor or three-factor model. Ljungqvist
and Richardson (2003) examine in detail all the venture fund investments of a single
large institutional investor, and they nd a 19.8% internal rate of return. They
reduce the sample selection problem posed by projects still private at the end of the
sample by focusing on investments made before 1992, almost all of which have
resolved. Assigning betas, they recover a 56% premium, which they interpret as a
premium for the illiquidity of venture capital investments.
ARTICLE IN PRESS
8
Discount rates applied by VC investors might be informative, but the contrast

between high discount rates applied by venture capital investors and lower ex post
average returns is an enduring puzzle in the venture capital literature. Smith
and Smith (2000) survey a large number of studies that report discount rates of 35%
to 50%. However, this puzzle depends on the interpretation of expected cash
ows. If expected means what will happen if everything goes as planned, it is
much larger than a conditional mean, and a larger discount rate should be
applied.
3. Overcoming selection bias

We observe a return only when the rm gets new nancing or is acquired, but this
fact need not bias our estimates. If the probability of observing a return were
independent of the projects value, simple averages would still correctly measure the
underlying return characteristics. However, projects are more likely to get new
nancing, and especially to go public, when their value has risen. As a result, the
mean returns to projects that get additional nancing are an upward-biased estimate
of the underlying mean return.
To understand the effects of selection, suppose that every project goes public when
its value has grown by a factor of 10. Now, every measured return is exactly 1,000%,
no matter what the underlying return distribution. A mean return of 1,000% and a
zero standard deviation is obviously a wildly biased estimate of the returns facing an
investor!
In this example, however, we can still identify the parameters of the underlying
return distribution. The 1,000% measured returns tell us that the cutoff for going
public is 1,000%. Observed returns tell us about the selection function, not the return
distribution. The fraction of projects that go public at each age then identies the
return distribution. If we see that 10% of the projects go public in one year, then we
know that the 10% upper tail of the return distribution begins at a 1,000% return.
Since the mean grows with horizon and the standard deviation grows with the square
root of horizon, the fractions that go public over time can separately identify the
mean and the standard deviation (and, potentially, other moments) of the underlying
return distribution.
In reality, the selection of projects to get new nancing or be acquired is not a step
function of value. Instead, the probability of obtaining new nancing is a smoothly
increasing function of the projects value, as illustrated by PrIPOjValue in Fig. 1.
The distribution of measured returns is then the product of the underlying return
distribution and the rising selection probability. Measured returns still have an
upward-biased mean and a downward-biased volatility. The calculations are more
complex, but we can still identify the underlying return distribution and the selection
function by watching the distribution of observed returns as well as the fraction of
projects that obtain new nancing over time.
I have nothing new to say about why projects are more likely to get new nancing
when value has increased, and I t a convenient functional form rather than impose
ARTICLE IN PRESS
Pr(IPO|Value)
Return = Value at year 1
Measured Returns
Fig. 1. Generating the measured return distribution from the underlying return distribution and selection
of projects to go public.
a particular economic model of this phenomenon. Its not surprising: good news
about future productivity raises value and the need for new nancing. The standard
q theory of investment also predicts that rms will invest more when their values rise.
(MacIntosh (1997, p. 295) discusses selection.) I also do not model the fact that more
projects are started when market valuations are high, though the same motivations
apply.
3.1. Maximum likelihood estimation
My objective is to estimate the mean, standard deviation, alpha, and beta of
venture capital investments, correcting for the selection bias caused by the fact that
we do not see returns for projects that remain private. To do this, I have to develop a
model of the probability structure of the datahow the returns we see are generated
from the underlying return process and the selection of projects that get new
nancing or go out of business. Then, for each possible value of the parameters, I
can calculate the probability of seeing the data given those parameters.
The fundamental data unit is a nancing round. Each round can have one of three
basic fates. First, the rm can go public, be acquired, or get a new round of
nancing. These fates give us a new valuation, so we can measure a return. For this
discussion, I lump all three fates together under the name new nancing round.
Second, the rm can go out of business. Third, the rm can remain private at the end
of the sample. We need to calculate the probabilities of these three events, and the
probability of the observed return if the rm gets new nancing.
ARTICLE IN PRESS
10
Fig. 2 illustrates how I calculate the likelihood function. I set up a grid for the log
of the projects value logV t at each date t. I start each project at an initial value
V 0 1; as shown in the top panel of Fig. 2. (Im following the fate of a typical dollar
invested.) I model the growth in value for subsequent periods as a lognormally
distributed variable,
ln

V t1
f
g ln Rft dln Rm
t1 ln Rt et1 ;
Vt
et1 N0; s2 :
(1)
I use a time interval of three months, balancing accuracy and simulation time. Eq. (1)
is like the CAPM, but using log rather than arithmetic returns. Given the extreme
skewness and volatility of venture capital investments, a statistical model with
normally distributed arithmetic returns would be strikingly inappropriate. Below, I
derive and report the implied market model for arithmetic returns (alpha and beta)
from this linear lognormal statistical model. From Eq. (1), I generate the probability
distribution of value at the beginning of period 1, PrV 1 as shown in the second
panel of Fig. 2.
Time zero value = $1
Pr(out|value)
Pr(new round|value)
Value at beginning of time 1
Pr(new round at time 1)
k
Pr(out of bus. at time 1)
Pr(still private at end of time 1)
Value at beginning of time 2

Pr(new round at time 2)
-1
-0.5
0.5
log value grid
Fig. 2. Procedure for calculating the likelihood function.
1.5
ARTICLE IN PRESS
11
Next, the rm could get a new nancing round. The probability of getting a new
round is an increasing function of value. I model this probability as a logistic
function,
Prnew round at tjV t 1=1 ealnV t b :
(2)
This function rises smoothly from 0 to 1, as shown in the second panel of Fig. 2.
Since I have started with a value of $1, I assume here that selection to go public
depends on total return achieved, not size per se. A $1 investment that grows to
$1,000 is likely to go public, where a $10,000 investment that falls to $1,000 is not.
Now I can nd the probability that the rm gets a new round with value V t ;
Prnew round at t; value V t PrV t Prnew round at tjV t :
This probability is shown by the bars on the right-hand side of the second panel of
Fig. 2. These rms exit the calculation of subsequent probabilities.
Next, the rm can go out of business. This is more likely for low values. I model
Prout of business at tjV t as a declining linear function of value V t ; starting from
the lowest value gridpoint and ending at an upper bound k, as shown by
Proutjvalue on the left side of the second panel of Fig. 2. A lognormal process such
as (1) never reaches a value of zero, so we must envision something like k if we are to
generate a nite probability of going out of business.1 Multiplying, we obtain the
probability that the rm goes out of business in period 1,
Prout of business at t; value V t
PrV t 1 Prnew round at tjV t Prout of business at tjV t :
These probabilities are shown by the bars on the left side of the second panel
of Fig. 2.
Next, I calculate the probability that the rm remains private at the end of period
1. These are just the rms that are left over,
Prprivate at end of t; value V t
PrV t 1 Prnew roundjV t 1 Prout of businessjV t :
This probability is indicated by the bars in the third panel of Fig. 2.
Next, I again apply (1) to nd the probability that the rm enters the second
period with value V 2 ; shown in the bottom panel of Fig. 2,
X
PrV t1 jV t Prprivate at end of t; V t :
(3)
PrV t1
Vt
PrV t1 jV t is given by the lognormal distribution of (1). As before, I nd the

probability of a new round in period 2, the probability of going out of business in
1
The working paper version of this article used a simpler specication that the rm went out of business
if V fell below k. Unfortunately, this specication leads to numerical problems, since the likelihood
function changes discontinuously as the parameter k passes through a value gridpoint. The linear
probability model is more realistic, and results in a better-behaved continuous likelihood function. A
smooth function like the logistic new nancing selection function would be prettier, but this specication
requires only one parameter, and the computational cost of extra parameters is high.
ARTICLE IN PRESS
12
period 2, and the probability of remaining private at the end of period 2. All of these
are shown in the bottom panel of Fig. 2. This procedure continues until we reach the
end of the sample.
3.2. Accounting for data errors
Many data points have bad or missing dates or returns. Each round results in one
of the following categories: (1) new nancing with good date and good return data,
(2) new nancing with good dates but bad return data, (3) new nancing with bad
dates and bad return data, (4) still private at end of sample, (5) out of business with
good exit date, (6) out of business with bad exit date.
To assign the probability of a type 1 event, a new round with a good date and
good return data, I rst nd the fraction d of all rounds with new nancing that have
good date and return data. Then, the probability of seeing this event is d times the
probability of a new round at age t with value V t ;
Prnew financing at age t; value V t ; good data
d Prnew financing at t; value V t :
I assume here that seeing good data is independent of value.

A few projects with normal returns in a very short time have astronomical
annualized returns. Are these few data points driving the results? One outlier
observation with probability near zero can have a huge impact on maximum
likelihood. As a simple way to account for such outliers, I consider a uniformly
distributed measurement error. With probability 1 p; the data record the true
value. With probability p; the data erroneously record a value uniformly distributed
over the value grid. I modify Eq. (4) to
Prnew financing at age t; value V t ; good data
d 1 p Prnew financing at t; value V t
X
1
Prnew financing at t; value V t :
d p
#gridpoints V
t
This modication fattens up the tails of the measured value distribution. It allows a
small number of observations to get a huge positive or negative return by
measurement error rather than force a huge mean or variance of the return
distribution to accommodate a few extreme annualized returns.
A type 2 event, new nancing with good dates but bad return data, is still
informative. We know how long it takes this investment round to build up the
kind of value that typically leads to new nancing. To calculate the probability of a
type 2 event, I sum across the vertical bars on the right side of the second panel of
Fig. 2,
Prnew financing at age t; no return data
X
Prnew financing at t; value V t :
1 d
Vt
ARTICLE IN PRESS
13
A type 3 event, new nancing with bad dates and bad return data, tells us that at
some point this project was good enough to get new nancing, though we know only
that it happened between the start of the project and the end of the sample. To
calculate the probability of this event, I sum over time from the initial round date to
the end of the sample as well,
Prnew financing, no date or return data
XX
Prnew financing at t; valueV t :
1 d
t
Vt
To nd the probability of a type 4 event, still private at the end of the sample,
I simply sum across values at the appropriate age
Prstill private at end of sample
X
Prstill private at t end of sample start date; V t :
Vt
. Type 5 and 6 events, out of business, tell us about the lower tail of the return
distribution. Some of the out of business observations have dates, and some do not.
Even when there is apparently good date data, a large fraction of the out-of-business
observations occur on two specic dates. Apparently, there were periodic data
cleanups of out-of-business observations prior to 1997. Therefore, when there is an
out-of-business date, I interpret it as this rm went out of business on or before
date t, summing up the probabilities of younger out-of-business events, rather than
on date t. This assignment affects the results: since one of the cleanup dates comes
on the heels of a large positive stock return, using the dates as they are leads to
negative beta estimates. To account for missing date data in out-of-business rms, I
calculate the fraction of all out-of-business rounds with good date data c. Thus, I
calculate the probability of a type 5 event, out of business with good date
information, as
Prout of business on or before age t; date data
t X
X
Prout of business at t; V t :
c
t1
Vt
Finally, if the date data are bad, all we know is that this round went out of
business at some point before the end of the sample. I calculate the probability of a
type 6 event as
Prout of business, no date data
1 c
end X
X
t1
Prout of business at t; V t :
Vt
Based on the above structure, for given parameters fg; d; s; k; a; b; pg; I can
compute the probability that we see any data point. Taking the log and adding up
over all data points, I obtain the log likelihood. I search numerically over values
fg; d; s; k; a; b; pg to maximize the likelihood function.
ARTICLE IN PRESS
14
Of course, the ability to separately identify the probability of going public and the
parameters of the return process requires some assumptions. Most important, I
assume that the selection function Pr(new round jV t ) is the same for rms of all ages
t. If the initial value doubles in a month, we are just as likely to get a new round as if
it takes ten years to double the initial value. This is surely unrealistic at very short
and very long time periods. I also assume that the return process is i.i.d. One might
specify that value creation starts slowly and then accelerates, or that betas or
volatilities change with size. However, identifying these tendencies without much
more data will be tenuous.
4. Data
I use the VentureOne database from its beginning in 1987 to June 2000. The
dataset consists of 16,613 nancing rounds, with 7,765 companies and a total of
$112,613 million raised. VentureOne claims to have captured approximately 98% of
nancing rounds, mitigating survival bias of projects and funds. However, the
VentureOne data are not completely free of survival bias. VentureOne records a
nancing round if it includes at least one venture capital rm with $20 million or
more in assets under management. Having found a qualifying round, they search for
previous rounds. Gompers and Lerner (2000, 288pp.) discuss this and other potential
selection biases in the database. Kaplan et al. (2002) compare the VentureOne data
to a sample of 143 VC nancings on which they have detailed information. They nd
as many as 15% of rounds omitted. They nd that post-money values of a nancing
round, though not the fact of the round, are more likely to be reported if the
company subsequently goes public. This selection problem does not bias my
estimates.
The VentureOne data do not include the nancial results of a public offering,
merger, or acquisition. To compute such values, we use the SDC Platinum
Corporate New Issues and Mergers and Acquisitions (M&A) databases, MarketGuide, and other online resources.2 We calculate returns to IPO using offering
prices. There is usually a lockup period between IPO and the time that venture
capital investors can sell shares, and there is an active literature studying IPO
mispricing, post-IPO drift and lockup-expiration effects, so one might want to study
returns to the end of the rst day of trading, or even include six months or more of
market returns. However, my objective is to measure venture capital returns, not to
contribute to the large literature that studies returns to newly listed rms. For this
purpose, it seems wisest to draw the line at the offering price. For example, suppose
that I include rst-day returns, and that this inclusion substantially raises the
resulting mean returns and alphas. Would we call that the risk and return of
venture capital or IPO mispricing? Clearly the latter, so I stop at offering prices
to focus on the former. In addition, all of these new-listing effects are small
compared to the returns (and errors) in the venture capital data. Even a 10% error in
2
We here includes Shawn Blosser, who assembled the venture capital data.
ARTICLE IN PRESS
15
nal value would have little effect on my results, since it is spread over the many
years of a typical VC investment. A 10% error is only four days of volatility at the
estimated nearly 100% standard deviation of return.3
The basic data consist of the date of each investment, dollar amount invested,
value of the rm after each investment, and characteristics including industry and
location. VentureOne also notes whether the company has gone public, been
acquired, or gone out of business, and the date of these events. We infer returns by
tracking the value of the rm after each investment. For example, suppose rm XYZ
has a rst round that raises $10 million, after which the rm is valued at $20 million.
We infer that the VC investors own half of the stock. If the rm later goes public,
raising $50 million and valued at $100 million after IPO, we infer that the VC
investors portion of the rm is now worth $25 million. We then infer their gross
return at $25M/$10M = 250%. We use the same method to assess dilution of initial
investors claims in multiple rounds.
The biggest potential error of this procedure is that if VentureOne misses
intermediate rounds, the extra investment is credited as a return to the original
investors. For example, the edition of VentureOne I used to construct the data
missed all but the seed round of Yahoo, resulting in a return even more enormous
than reality. I run the data through several lters4 and I add the measurement error
process p to try to account for this kind of error.
Venture capitalists typically obtain convertible preferred rather than common
stock. (See Kaplan and Stromberg (2003). Admati and Peiderer (1994) have a nice
summary of venture capital arrangements, especially mechanisms designed to insure
that valuations are arms length.) These arrangements are not noted in the
VentureOne data, so I treat all investments as common stock. This approximation is
not likely to introduce a large bias. The results are driven by the successes, not by
liquidation values in the surprisingly rare failures, or in acquisitions that produce
losses for common stock investors, where convertible preferred holders can retrieve
their capital. In addition, the bias will be to understate estimated VC returns, while
3
The unusually large rst-day returns in 1999 and 2000 are a possible exception. For example,
Ljungqvist and Wilhelm (2003, Table, II) report mean rst-day returns for 19962000 of 17%, 14%, 23%,
73%, and 58%, with medians of 10%, 9%, 10%, 39%, and 30%. However, these are reported as
transitory anomalies, not returns expected when the projects are started. We should be uncomfortable
adding a 73% expected one-day return to our view of the venture capital value creation process. Also, I
nd below quite similar results in the pre-1997 sample, which avoids this anomalous period. See also Lee
and Wahal (2002), who nd that VC-backed rms have larger rst-day returns than other rms.
4
Starting with 16,852 observations in the base case of the IPO/acquisition sample (numbers vary for
subsamples), I eliminate 99 observations with more than 100% or less than 0% inferred shareholder value,
and I eliminate 107 investments in the last period, the second quarter of 2000, since the model cant say
anything until at least one period has passed. In 25 observations, the exit date comes before the VC round
date, so I treat the exit date as missing.
For the maximum likelihood estimation, I treat 37 IPO, acquisition, or new rounds with zero returns
as out of business (0 blows up a lognormal), and I delete four observations with anomalously high returns
(over 30,000%) after I hand-checking them and nding that they were errors due to missing intermediate
rounds. I similarly deleted four observations with a log annualized return greater than 15
(100 e15 1 3:269 108 %) on the strong suspicion of measurement error in the dates. All of these
observations are included in the data characterization, however. I am left with 16,638 data points.
ARTICLE IN PRESS
16
the puzzle is that the estimated returns are so high. Gilson and Schizer (2003)
argue that the practice of issuing convertible preferred stock to VC investors is
not driven by cash ow or control considerations, but by tax law. Management
is typically awarded common shares at the same time as the venture nancing
round. Distinguishing the classes of shares allows managers to underreport the value
of their share grants, taxable immediately at ordinary income rates, and thus to
report this value as a capital gain later on. If so, then the distinction between
common and convertible preferred shares makes even less of a difference for my
analysis.
I model the return to equity directly, so the fact that debt data are unavailable
does not generate an accounting mistake in calculating returns. Firms with different
levels of debt can have different betas, however, which I do not capture.
4.1. IPO/acquisition and round-to-round samples
The basic data unit is a nancing round. If a nancing round is followed by
another round, if the rm is acquired, or if the rm goes public, we can calculate a
return. I consider two basic sample denitions for these returns. In the round-toround sample, I measure every return from a nancing round to a subsequent
nancing round, IPO, or acquisition. Thus, if a rm has two nancing rounds and
then goes public, I measure two returns, from round 1 to round 2, and from round 2
to IPO. If the rm has two rounds and then fails, I measure a positive return from
round 1 to round 2 but then a failure from round 2. If the rm has two rounds and
remains private, I measure a return from round 1 to round 2, but round 2 is coded as
remaining private.
One might be suspicious of returns constructed from such round-to-round
valuations. A new round determines the terms at which new investors come in but
almost never the terms at which old investors can get out. The returns to investors
are really the returns to acquisition or IPO only, ignoring intermediate nancing
rounds. In addition, an important reason to study venture capital is to examine
whether venture capital investments have low prices and high returns due to their
illiquidity. We can only hope to see this fact in returns from investment to IPO, not
in returns from one round of venture investment to another, since the latter returns
retain the illiquid character of venture capital investments. More basically, it is
interesting to characterize the eventual fate of venture capital investments as well as
the returns measured in successive nancing rounds.
For all these reasons, I emphasize a second basic data sample, denoted IPO/
acquisition below. If a rm has two rounds and then goes public, I measure two
returns, round 1 to IPO, and round 2 to IPO. If the rm has two rounds and then
fails, I measure two failures, round 1 to failure and round 2 to failure. If it has two
rounds and remains private, both rounds are coded as remaining private with no
measured returns. In addition to its direct interest, we can look for signs of an
illiquidity or other premium by contrasting these round-to-IPO returns with the
above round-to-round returns. Different rounds of the same company overlap in
time, of course, and I deal with the econometric issues raised by this overlap below.
ARTICLE IN PRESS
17
Table 1
The fate of venture capital investments
IPO/acquisition
Round to round
Fate
Return
No return
Total
IPO
Acquisition
Out of business
Remains private
IPO registered
New round
16.1
5.8
9.0
5.3
14.6
21.4
20.4
9.0
45.5
3.7
45.5
3.7
Return
5.9
2.9
4.2
28.3
No return
2.0
6.3
23.3
1.2
25.9
Total
7.9
9.2
4.2
23.3
1.2
54.2
Note: Table entries are the percentage of venture capital nancing rounds with the indicated fates. The
IPO/acquisition sample tracks each investment to its nal fate. The round-to-round sample tracks each
investment to its next nancing round only. Return indicates rounds for which we can measure a return,
No reference indicates rounds in the given category (e.g., IPO) but for which data are missing and we
cannot calculate a return.
Table 1 characterizes the fates of venture capital investments. We see that 21.4%
of rounds eventually result in an IPO and 20.4% eventually result in acquisition.
Unfortunately, I am able to assign a return to only about three quarters of the IPO
and one quarter of the acquisitions. We see that 45.5% remain private, 3.7% have
registered for but not completed an IPO, and 9% go out of business. There are
surprisingly few failures. Moskowitz and Vissing-Jorgenson (2002) nd that only
34% of their sample of private equity survive ten years. However, many rms go
public at valuations that give losses to VC investors, and many more are acquired on
such terms. (Weighting by dollars invested yields quite similar numbers, so I lump
investments together without size effects in the estimation.)
I measure far more returns in the round-to-round sample. The average company
has 2.1 venture capital nancing rounds (16; 638 rounds=7; 765 companies), so the
fractions that end in IPO, acquisition, out of business, or still private are cut in half,
while 54.2% get a new round, about half of which result in return data. The smaller
number that remain private means less selection bias to control for, and less worry
that some of the still-private rms are living dead, really out of business.
5. Results
Table 2 presents characteristics of the subsamples. Table 3 presents parameter
estimates for the IPO/acquisition sample, and Table 4 presents estimates for the
round-to-round sample. Table 5 presents asymptotic standard errors.
5.1. Base case results
The base case is the All sample in Table 3. The mean log return in Table 3 is a
sensible 15%, just about the same as the 15.9% mean log S&P500 return in this
ARTICLE IN PRESS
18
Table 2
Characteristics of the samples
Rounds
All
Industries
Subsamples
Health
Info
Retail
Other
Pre 97
Dead 00
IPO/acquisition sample
Number
16,638 7,668
Out of bus. 9
9
IPO
21
17
Acquired
20
20
Private
49
54
c
95
93
d
48
38
4,474
9
21
21
49
97
49
2,453
9
26
21
43
98
57
1,234
9
31
19
41
96
62
3,915
9
27
18
46
96
51
9,190
10
21
25
45
94
49
3,091
7
15
10
68
96
38
442
12
22
29
38
94
26
5,932
5
33
26
36
75
48
16,638
58
21
20
0
99
52
Round-to-round sample
Number
16,633 7,667
Out of bus. 4
4
IPO
8
5
Acquired
9
8
New round 54
59
Private
25
25
c
93
88
d
51
42
4,471
4
7
9
55
25
96
55
2,453
5
11
11
50
23
99
61
1,234
5
18
11
41
25
98
66
3,912
4
9
8
59
20
94
55
9,188
4
8
11
55
22
93
52
3,091
4
7
5
45
39
94
41
442
7
10
13
52
18
90
39
6,764
2
12
11
69
7
67
54
16,633
29
8
9
54
0
99
52
Note: All entries except Number are percentages. c percent of out of business with good data. d
percent of new nancing or acquisition with good data. Private are rms still private at the end of the
sample, including rms that have registered for but not completed an IPO.
period. (I report average returns, alphas and standard deviations as annualized

percentages, by multiplying averages and alphas by 400 and multiplying standard
deviations by 200.) The standard deviation of log return is 89%, much larger than
the 14.9% standard deviation of the log S&P500 return in this period. These are
individual rms, so we expect them to be quite volatile compared to a diversied
portfolio such as thepS&P500.
The 89% annualized standard deviation might be
easier to digest as 89= 365 4:7% daily standard deviation, which is typical of very
small growth stocks.
The intercept g is negative at 7:1% The slope d is sensible at 1.7; venture capital
is riskier than the S&P500. The residual standard deviation s is large at 86%.
The volatility of returns comes from idiosyncratic volatility, not from a large
slope coefcient. The implied regression R2 is a very small 0:075:
(1:72 14:92 =1:72 14:92 892 0:075:) Systematic risk is a small component of
the risk of an individual venture capital investment.
(I estimate the parameters g; d; s directly. I calculate E ln R and s ln R in the rst
two columns using the mean 19872000 Treasury bill return of 6.8%, and the
S&P500 mean and standard deviation of 15.9% and 14.9%, e.g., E ln R 7:1
6:8 1:7 15:9 6:8 15%: I present mean log returns rst in Tables 3 and 4, as
the mean is better estimated, more stable, and more comparable across specications
than is its decomposition into an intercept and a slope.)
ARTICLE IN PRESS
19
Table 3
Parameter estimates in the IPO/acquisition sample
E ln R s ln R g
All, baseline
Asymptotic s
Bootstrap s
15
Health
Info
Retail
Other
1
2
3
4
59
sR a
97
96
103
105
7.7
0.3
27
1.2
0.9
0.5
93
92
100
66
69
67
72
121
119
129
134
19
12
8.0
0.8
96
98
98
99
3.7
1.6
4.4
12
1.0
0.8
0.6
0.5
95
97
98
99
71
65
60
51
120
120
120
119
17
15
17
25
67
108
127
62
8.7 0.2
5.2 1.4
11 0.1
13
0.6
67
105
127
61
42
79
111
46
6.7
86
0.6
7.0
ER
1.7
0.04
0.4
Nasdaq
14
Nasdaq Dec1
17
Nasdaq o$2M 8.2
No d
11
Round
Round
Round
Round
7.1
0.7
1.7
2.4
89
5.9
107
1.9 25
1.0
0.02 0.02
0.4 3.6 0.08
3.8
0.06
0.28
9.6
0.6
1.9
39
45
22
1.4 20
1.0 22
0.5 14
11
0.7
0.7
0.7
0.8
5.0
5.4
5.0
4.3
5.7
6.3
4.1
4.2
53
49
46
39
1.1
0.9
0.7
0.5
17
16
17
13
1.0
1.0
0.8
1.1
4.2
3.6
3.9
2.5
8.0
5.0
2.9
5.5
76 33
139 55
181 106
71 33
0.2
1.7
0.1
0.6
36
14
11
53
0.7
0.8
0.4
0.4
5.1
4.3
10.0
10.0
7.8
4.3
2.9
13
11
32
9.4
Note: Returns are calculated from venture capital nancing round to eventual IPO, acquisition, or failure,
ignoring intermediate venture nancing rounds.
Columns: E ln R; s ln R are the parameters of the underlying lognormal return process. All means,
standard deviations and alphas are reported as annualized percentages, e.g., 400 E
ln R;
200 s ln R;
V
g ln Rft
400 ER 1; etc. g; d; and s are the parameters of the market model in logs, ln Vt1
t
f
m
2
dln Rt1 ln Rt et1 ; et1 N0; s : E ln R; s ln R are calculated from g; d; s using the sample mean
and variance of the three-month T-bill rate Rf and S&P500 return Rm ; E ln R g E ln Rft
2
E ln Rft and p
s2
ln R d2 s2 ln Rm
dE ln Rm
t s : ER; sR are average arithmetic returns ER
1 t2
eE ln R2s ln R ; sR ER es2 ln R 1: a and b are implied parameters of the discrete time regression
f
i
model in levels, V it1 =V it a Rft bRm
t1 Rt vt1 : k; a; b are estimated parameters of the selection
function. k is point at which rms start to go out of business, expressed as a percentage of initial value. a; b
govern the selection function PrIPO, acq. at tjV t 1=1 ealnV t b : Given that an IPO/acquisition
occurs, there is a probability p that a uniformly distributed value is recorded instead of the correct value.
Rows: All includes all nancing rounds. Asymptotic standard errors are based on second derivatives of
the likelihood function. Bootstrap standard errors are based on 20 replications of the estimate, choosing
the sample randomly with replacement. Nasdaq, Nasdaq Dec1, Nasdaq o$2M, and No d use
the indicated reference returns in place of the S&P500. Round i considers only investments in nancing
round i. Health, Info, Retail, Other are industry classications.
The asymptotic standard errors in the second row of Table 3 indicate that all these
numbers are measured with great statistical precision. The bootstrap standard errors
in the third row are a good deal larger than asymptotic standard errors, but still
show the parameters to be quite well estimated. The bootstrap standard errors are
larger in part because there are a small number of outlier data points with very large
likelihoods. Their inclusion or exclusion has a larger effect on the results than the
asymptotic distribution theory suggests. The asymptotic standard errors also ignore
ARTICLE IN PRESS
20
Table 4
Parameter estimates in the round-to-round sample
E ln R
All, baseline
Asymptotic s
Bootstrap s
20
Nasdaq
Nasdaq Dec1
Nasdaq o$2M
No d
15
22
16
21
Round
Round
Round
Round
s ln R
84
Health
Info
Retail
Other
ER
sR
100
45
0.6
21
0.4
3.8
1.7
0.02
0.2
1.3
0.02
0.3
4.7
0.4
0.8
7.6
1.1
4.7
0.6
0.1
0.5
84
0.8
6.4
59
91
90
91
85
4.9
7.3
4.5
1.1
0.7
0.2
87
88
90
61
68
62
61
110
112
111
102
35
49
37
1.2
0.7
0.2
18
24
16
20
1.5
0.6
1.6
1.6
1.5
3.3
1.4
1.4
3.4
2.5
3.5
4.2
26
20
15
8.8
90
83
77
84
11
7.5
3.6
0.1
0.8
0.6
0.5
0.2
89
82
77
83
72
58
47
46
112
99
89
97
55
44
35
37
1.0
0.7
0.5
0.2
16
22
29
21
1.9
1.6
1.4
1.3
1.3
1.4
1.4
1.4
4.3
3.6
4.6
3.7
24
23
25
8.0
62
95
121
64
15
12
11
3.9
0.3
0.5
0.7
0.6
62
94
121
63
46
74
111
29
70
119
171
70
36
62
96
16
0.3
0.5
0.8
0.6
48
19
14
35
0.3
0.7
0.5
0.5
7.6
2.9
4.1
5.2
4.6
2.2
0.5
3.6
1.1
1
2
3
4
7.2
7.5
11
5.7
0.5
Note: Returns are calculated from venture capital nancing round to the next event: new nancing, IPO,
acquisition, or failure. See the note to Table 3 for row and column headings.
cross-correlation between individual venture capital returns, since I do not specify a

cross-correlation structure in the data-generating model (1).
So far, the estimates look reasonable. If anything, the negative intercept is
surprisingly low. However, the CAPM and most asset pricing and portfolio theory
species arithmetic, not logarithmic, returns. Portfolios are linear in arithmetic, not
log, returns, so diversication applies to arithmetic returns. The columns ER; sR; a;
and b of Table 3 calculate implied characteristics of arithmetic returns.5 The mean
5
We want to nd the model in levels implied by Eq. (1), i.e.
V it1
f
i
Rft a bRm
t1 Rt vt1 :
V it
I nd b from b covR; Rm =varRm ; and then a from a ER Rf bERm Rf : The formulas are
2
b egd1Eln R
f
m ln Rf s2 =2s2 1s2 =2

m
2 2
edsm 1
2
esm
1
(6)
f
a elnR fegdEln R ln R d sm =2s =2 1 bemm ln R sm =2 1g;

where s2m s2 ln Rm : The continuous time limit is simpler, b d; se sv; and
1
1
dd 1s2m s2 :
2
2
I present the discrete time computations in the tables; the continuous time results are quite similar.
ag
(7)
ARTICLE IN PRESS
21
Table 5
Asymptotic standard errors for Tables 3 and 4
IPO/acquisition (Table 3)
Round to round (Table 4)
All, baseline
Bootstrap
0.7
1.7
0.04
0.37
0.6
7.0
0.02
3.57
0.02
0.08
0.06
0.28
0.6
1.9
1.1
4.7
0.08
0.46
0.8
6.4
0.4
3.8
0.02
0.20
0.02
0.33
0.4
0.8
Nasdaq
Nasdaq Dec1
Nasdaq o$2M
No d
1.0
1.0
1.7
0.7
0.05
0.04
0.02
1.1
1.1
0.8
1.0
0.81
0.62
0.35
0.15
0.02
0.02
0.01
0.02
0.13
0.15
0.08
0.11
0.5
0.6
0.5
0.6
0.7
1.1
1.2
0.7
0.01
0.04
0.01
0.9
1.1
0.5
0.8
0.4
1.0
0.2
0.6
0.03
0.01
0.03
0.03
0.03
0.02
0.02
0.03
0.4
0.4
0.3
0.3
Round
Round
Round
Round
1.2
2.4
3.2
5.3
0.05
0.20
0.24
0.38
1.8
2.3
2.5
3.3
0.94
1.18
1.16
1.57
0.04
0.06
0.08
0.08
0.11
0.16
0.31
0.20
1.0
1.2
1.1
1.8
1.4
2.3
2.4
4.0
0.09
0.16
0.16
0.26
1.3
1.8
1.7
2.9
0.7
1.4
1.3
2.0
0.06
0.07
0.08
0.12
0.03
0.05
0.08
0.14
0.5
0.8
0.9
1.1
1.7
1.7
1.9
3.5
0.14
0.13
0.08
0.26
1.4
1.6
3.5
4.6
2.06
0.69
1.27
7.52
0.03
0.01
0.00
0.01
0.19
0.06
0.00
0.14
1.2
0.8
1.2
4.6
1.9
1.7
5.5
6.8
0.15
0.13
0.30
0.46
1.5
1.3
3.8
5.0
2.2
0.8
1.7
6.1
0.01
0.01
0.02
0.10
0.20
0.04
0.10
1.07
0.8
0.4
0.5
4.0
Health
Info
Retail
Other
1
2
3
4
arithmetic return ER in Table 3 is a whopping 59%, with a 107% standard

deviation. Even the 1.9 arithmetic b and the large S&P500 return in this period do
not generate a return that high, leaving a 32% arithmetic a:
The large mean arithmetic returns and alphas result from the volatility rather
than the mean of the underlying lognormal return distribution. The mean
2
arithmetic return is ER eE ln R1=2 s ln R : With s2 ln R on the order of
100%, usually negligible 12 s2 terms generate 50% per year arithmetic returns by
themselves. Venture capital investments are like call options; their arithmetic
mean return depends on the mass in the right tail, which is driven by volatility
more than by drift. I examine the high arithmetic returns and alphas in great detail
below.
The out-of-business cutoff parameter k is 25%, meaning that the chance of going
out of business rises to 12 at 12.5% of initial value. This is a low number, but
reasonable. Venture capital investors are likely to hang in there and wait for the nal
payout despite substantial intermediate losses.
The parameters a and b control the selection function. b is the point at which there
is a 50% probability of going public or being acquired per quarter, and it occurs at a
substantial 380% log return. Finally, the measurement error parameter p is about
10% and statistically signicant. The estimation accounts for a small number of
large positive and negative returns as measurement error rather than treat them as
extreme values of a lognormal process.
ARTICLE IN PRESS
22
The round-to-round sample in Table 4 gives quite similar results. The average log
return is slightly higher, 20% rather than 15%, with quite similar volatility, 84%
rather than 89%. The average log return splits in to a lower slope, 0.6, and thus a
higher intercept, 7.6%. As we will see below, IPOs are more sensitive to market
conditions than new rounds, so an estimate that emphasizes new rounds sees a lower
slope. As in the IPO/acquisition sample, the average arithmetic returns, driven by
large idiosyncratic volatility, are huge at 59%, with 100% standard deviation and
45% arithmetic a: The selection function parameter b is much lower, centering that
function at 130% growth in log value. The typical rm builds value through several
rounds before IPO, so this is what we expect. The measurement error p is lower,
showing the smaller fraction of large outliers in the round to round valuations. The
asymptotic standard errors in Table 5 are quite similar to those of the IPO/
acquisition sample. Once again, the bootstrap standard errors are larger, but the
parameters are still well estimated.
5.2. Alternative reference returns
Perhaps the Nasdaq or small-stock Nasdaq portfolios provide better reference
returns than the S&P500. We are interested in comparing venture capital to similar
traded securities, not in testing an absolute asset pricing model, so a performance
attribution approach is appropriate. The next three rows of Tables 3 and 4 address
this case.
In the IPO/acquisition sample of Table 3, the slope coefcient declines from 1.7 to
1.2 using Nasdaq and to 0.9 using the CRSP Nasdaq decile 1 (small) stocks. We
expect betas nearer to one if these are more representative as reference returns.
However, the residual standard deviation actually goes up a little bit, so the implied
R2 s are even smaller. The mean log returns are about the same, and the arithmetic
alphas rise slightly.
Nasdaq o$2M is a portfolio of Nasdaq stocks with less than $2 million in
market capitalization, rebalanced monthly. I discuss this portfolio in detail below.
It has a 71% mean arithmetic return and a 62% S&P500 alpha, compared to the
22% mean arithmetic return and statistically insignicant 12% alpha for the CRSP
Nasdaq decile 1, so a b near one on this portfolio would eliminate the arithmetic
alpha in venture capital investments. This portfolio is a little more successful.
The log intercept declines to 27%; but the slope coefcient is only 0.5 so it only
cuts the arithmetic alpha down to 22%. In the round-to-round sample of Table 4,
there are small changes in the slope and log intercept g from changing the reference
return, but the 60% mean arithmetic return and 45% arithmetic alpha are basically
unchanged.
Perhaps the complications of the market model are leading to trouble. The No d
rows of Tables 3 and 4 estimate the mean and standard deviation of log returns
directly. The mean log returns are just about the same. In the IPO/acquisition sample
of Table 3, the standard deviation is even larger at 105%, leading to larger mean
arithmetic returns, 72% rather than 59%. In the round-to-round sample of Table 4,
all means and standard deviations are just about the same with no d:
ARTICLE IN PRESS
23
5.3. Rounds
The Round i subsamples in Tables 3 and 4 break the sample down by
investment rounds. Its interesting to see whether the different rounds have different
characteristics, i.e., whether later rounds are less risky. Its also important to do
this for the IPO/acquisition sample, for two reasons. First, the model taken
literally should not be applied to a sample with several rounds of the same rm,
since we cannot normalize the initial values of both rst and second rounds
to a dollar and use the same probability of new nancing as a function of value.
Applying the model to each round separately, we avoid this problem. The selection
function is rather at, however, so mixing the rounds might not make much
practical difference. Second, the use of overlapping rounds from the same rm
induces cross-correlation between observations, ignored by my maximum likelihood
estimate. This should affect standard errors and not bias point estimates. When we
look at each round separately, there is no overlap, so standard errors in the round
subsamples will indicate whether this cross-correlation in fact has any important
effects.
Table 2 already suggests that later rounds are slightly more mature. The chance of
ending up as an IPO rises from 17% for the rst round to 31% for the fourth round
in the IPO/acquisition sample, and from 5% to 18% in the round-to-round sample.
However, the chance of acquisition and failure is the same across rounds.
In the IPO/acquisition sample of Table 3, later rounds have progressively lower
mean log returns, from 19% to 0:8%; steadily lower slope coefcients, from 1.0 to
0.5, and steadily lower intercepts, from 3.7% to 12%: All of these estimates paint
the picture that later rounds are less riskyand hence less rewardinginvestments.
These ndings are consistent with the theoretical analysis of Berk et al. (2004). The
asymptotic standard error of the intercept g (Table 5) grows to ve percentage points
by round 4, however, so the statistical signicance of this pattern that g declines
across rounds is not high. The volatilities are huge and steady at about 100%, so we
still see large average arithmetic returns and alphas in all rounds. Still, even these
decline across rounds; arithmetic mean returns decline from 71% to 51% and
arithmetic alphas decline from 53% to 39% from rst-round to fourth-round
investments. The cutoff for going out of business k declines for later rounds, the
center point of the selection function b declines from 4.2 to 2.5, and the measurement
error p declines, all of which suggest less risky and more mature projects in later
rounds.
These patterns are all conrmed in the round-to-round sample of Table 4. As we
move to later rounds, the mean log return, intercept, and slope all decline, while
volatility is about the same. The mean arithmetic returns and alphas are still high,
but means decline from 72% to 46% and alphas decline from 55% to 37% from the
rst to fourth rounds.
In Table 5, the standard errors for round 1 (with the largest number of
observations) of the IPO/acquisition sample are still quite small compared to
economically interesting variation in the coefcients. The most important change is
the standard error of the intercept g which rises from 0.67 to 1.23. Thus, even if there
ARTICLE IN PRESS
24
is perfect cross-correlation between rounds, in which case additional rounds give no

additional information, the coefcients are well measured.
5.4. Industries
Venture capital is not all dot-com. Table 2 shows that roughly one-third of the
sample is in health, retail, or other industry classications. Perhaps the unusual
results are conned to the special events in the dot-com sector during this sample.
Table 2 shows that the industry subsamples have remarkably similar fates, however.
Technology (info) investments do not go public much more frequently, or fail any
less often, than other industries.
In Tables 3 and 4, mean log returns are quite similar across industries, except that
Other has a slightly larger mean log return (25% rather than 1517%) in the IPO/
acquisition sample, and a much lower mean log return (8% rather than 25%) in the
round-to-round sample. However, the small sample sizes mean that these estimates
have high standard errors in Table 5, so these differences are not likely to be
statistically signicant.
In Table 3, we see a larger slope d 1:4 for the information industry, and a
correspondingly lower intercept. Firms in the information industry went public
following large market increases, more so than in the other industries.
The main difference across industries is that information and retail have much
larger residual and overall variance, and lower failure cutoffs k. Variance is a key
parameter in accounting for success, especially early successes, as a higher variance
increases the mass in the right tail. Variance together with the cutoff k accounts for
failures, as both parameters increase the left tail. Thus, the pattern of higher variance
and lower k is driven by the larger number of early and highly protable IPOs in the
information and retail industries, together with the fact that failures are about the
same across industries.
Since the volatilities are still high, we still see large mean arithmetic returns and
arithmetic alphas, and the pattern is conrmed across all industry groups. The retail
industry in the IPO/acquisition sample is the champion, with a 106% arithmetic
alpha, driven by its 127% residual standard deviation and slightly negative beta. The
large arithmetic returns and alphas occur throughout venture capital, and do not
come from the high tech sample alone.
6. Facts: fates and returns

Maximum likelihood gives the appearance of statistical purity, yet it often leaves
one unsatised. Are there robust stylized facts behind these estimates? Or are they
driven by peculiar aspects of a few data points? Does maximum likelihood focus on
apparently well-measured but economically uninteresting moments in the data, at
the expense of capturing apparently less well-measured but more economically
important moments? In particular, the nding of huge arithmetic returns and alphas
sits uncomfortably. What facts in the data lie behind these estimates?
ARTICLE IN PRESS
25
As I argue earlier, the crucial stylized facts are the pattern of exitsnew nancing,
acquisition, or failurewith project age, and the returns achieved as a function of
age. It is also interesting to contrast the selection-biased, direct return estimates with
the selection-bias-corrected estimates above. So, let us look at the observed returns,
and at the speed with which projects get a return or go out of business.
6.1. Fates
Fig. 3 presents the cumulative fraction of rounds in each categorynew nancing
or acquisition, out of business, or still privateas a function of age, for the IPO/
acquired sample. The dashed lines give the data, while the solid lines give the
predictions of the model, using the baseline estimates from Table 3.
The data paint a picture of essentially exponential decay. About 10% of the
remaining rms go public or are acquired with each year of age, so that by ve years
after the initial investment, about half of the rounds have gone public or been
acquired. (The pattern is slightly speeded up in later subsamples. For example,
100
90
80
Still private
70
IPO, acquired
Percentage
60
50
Data
Model
40
30
20
10
Out of business
Years since investment
Fig. 3. Cumulative probability that a venture capital nancing round in the IPO/acquired sample will end
up IPO or acquired, out of business, or remain private, as a function of age. Dashed lines: data. Solid lines:
prediction of the model, using baseline estimates from Table 3.
ARTICLE IN PRESS
26
projects that start in 1995 go public and out of business at a slightly faster rate than
projects that start in 1990. However, the difference is small, so age alone is a
reasonable state variable.)
The model replicates these stylized features of the data reasonably well. The major
discrepancy is that the model seems to have almost twice the hazard of going out of
business seen in the data, and the number remaining private is correspondingly
lower. However, this comparison is misleading. The data lines in Fig. 3 treat out-ofbusiness dates as real, while the estimate treats data that say out of business on date
t as went out of business on or before date t, recognizing VentureOnes
occasional cleanups. This difference means that the estimates recognize failures
about twice as fast as in the VentureOne data, and that is the pattern we see in Fig. 3.
Also, the data lines characterize only the sample with good date information, while
the model estimates are chosen to t the entire sample, including rms with bad date
data. And, of course, maximum likelihood does not set out to pick parameters that
t this one moment as well as possible.
Fig. 4 presents the same picture for the round-to-round sample. Things
happen much faster in this sample, since the typical investment has several
rounds before going public, being acquired, or failing. Here roughly 30% of
the remaining rounds go public, are acquired, or get a new round of nancing
each year. The model provides an excellent t, with the same understanding of the
out-of-business lines.
6.2. Returns
Table 6 characterizes observed returns in the data, i.e., when there is a new
nancing or acquisition. The column headings give age bins in years. For example,
the 12 year column summarizes all investment rounds that went public or were
acquired between one and two years after the venture capital nancing round, and
for which I have good return and date data. The average log return in all age
categories of the IPO/acquisition sample is 108% with a 135% standard deviation.
This estimate contrasts strongly with the selection-bias-corrected estimate of a 15%
mean log return in Table 3. Correcting for selection bias has a huge impact on
estimated mean log returns.
Fig. 5 plots smoothed histograms of log returns in age categories. (The
distributions in Fig. 5 are normalized to have the same area; they are the
distribution of returns conditional on observing a return in the indicated time frame.)
The distribution of returns in Fig. 5 shifts slightly to the right and then stabilizes.
The average log returns in Table 6 show the same pattern: they increase slightly with
horizon out to 12 years, and then stabilize. These are total returns, not annualized.
This behavior is unusual. Log returns usually grow with horizon,
p so we expect veyear returns ve times as large as the one-year returns, and 5 times as spread out.
Total returns that stabilize are a signature of a selected sample. In the simple
example that all projects go public when they have achieved 1,000% growth, the
distribution of measured total returns is the samea point mass at 1,000%for all
horizons. Fig. 5 dramatically makes the case that we should regard venture capital
ARTICLE IN PRESS
27
100
90
Data
IPO, acquired, or new round
Still private
80
70
Model
Percentage
60
50
40
30
20
10
Out of business
0
4
5
Fig. 4. Cumulative probability that a venture capital nancing round in the round-to-round sample will
end up IPO, acquired, or new round; out of business; or remain private, as a function of age. Dashed lines:
data. Solid lines: prediction of the model, using baseline estimates from Table 4.
projects as a selected sample, with a selection function that is stable across project
ages.
Fig. 5 shows that, despite the 108% mean log return, a substantial fraction of
projects go public or are acquired at valuations that generate losses to the venture
capital investors, even on projects that go public or are acquired soon after the
venture capital investment (01 year bin). Venture capital has a high mean return,
but it is not a gold mine.
Fig. 6 presents the histogram of log returns as predicted by the model, using the
baseline estimate of Table 3. The model captures the return distributions of Fig. 5
quite well. In particular, note how the model return distributions settle down to a
constant at ve years and above.
Fig. 6 also includes the estimated selection function, which shows how the model
accounts for the pattern of observed returns across horizons. In the domain of the 3
month return distribution, the selection function is low and at. A small fraction of
projects go public, with a return distribution generated by the lognormal with a small
mean and a huge volatility, and little modied by selection. As the horizon increases,
the underlying return distribution shifts to the right, and starts to run in to the
ARTICLE IN PRESS
28
Table 6
Statistics for observed returns
Age bins
1 month1
(1) IPO/acquisition sample
Number
3,595
16 month
334
612 month 12 year 23 year 34 year 45 year 5 year1
476
877
706
525
283
413
(a) Log returns, percent (not annualized)

Average
108
63
Std. dev.
135
105
Median
105
57
93
118
86
104
130
100
127
136
127
135
143
131
118
146
136
97
147
113
(b) Arithmetic returns, percent

Average
698
306
Std. dev.
3,282
1,659
Median
184
77
399
881
135
737
4,828
172
849
2,548
255
1067
4,613
272
708
1,456
288
535
1,123
209
(c) Annualized arithmetic returns, percent

Average
3.7e+09
4.0e+10 1,200
Std. dev.
2.2e+11
7.2e+11 5,800
373
4,200
99
133
62
76
38
44
20
28
122
160
73
94
52
57
39
42
27
33
15
24
2,108
59
73
2,383
46
81
550
44
105
174
55
119
75
67
96
79
43
162
Average log returns, percent

48
57
55
81
51
84
50
113
84
42
94
24
26
110
46
44
91
39
55
99
44
14
99
0
(d) Annualized log returns, percent

Average
72
201
Std. dev.
148
371
(2) Round-to-round sample
(a) Log returns, percent
Number
6,125
Average
53
Std. dev.
85
(b) Subsamples.
New round
IPO
Acquisition
945
59
82
Note: The IPO/acquisition sample consists of all venture capital nancing rounds that eventually result
in an IPO or acquisition in the indicated time frame and with good return data. The round-to-round
sample consists of all venture capital nancing rounds that get another round of nancing, IPO, or
acquisition in the indicated time frame and with good return data.
steeply rising part of the selection function. Since the winners are removed from the
sample, the measured return distribution then settles down to a constant. The risk
facing a venture capital investor is as much when his or her return will occur as how
much that return will be.
ARTICLE IN PRESS
29
0-1
1-3
3-5
5+
-400
-300
-200
-100
0
100
Log Return
200
300
400
500
Fig. 5. Smoothed histogram of log returns by age categories, IPO/acquisition sample. Each point is a
normally weighted kernel estimate.
The estimated selection function is actually quite at. In Fig. 6, it only rises from a
20% to an 80% probability of going public as log value rises from 200% (an
arithmetic return of 100 e2 1 639%) to 500% (an arithmetic return of
100 e5 1 14; 741%). If the selection function were a step function, we would
see no variance of returns conditional on IPO or acquisition. The smoothly rising
selection function is required to generate the large variance of observed returns.
6.3. Round-to-round sample
Table 6 presents means and standard deviations in the round-to-round sample;
Fig. 7 presents smoothed histograms of log returns for this sample, and Fig. 8
presents the predictions of the model, using the round-to-round sample baseline
estimates. The average log returns are about half of their value in the IPO/
acquisition sample, though still substantial at about 50%. Again, we expect this
result since most rms have several venture rounds before going public or being
acquired. The standard deviation of log returns is still substantial, around 80%. As
the round to round means are about half the IPO/acquisition means, the round to
ARTICLE IN PRESS
30
1
0.9
0.8
3 mo.
Scale for Pr(IPO,acq.|V)
0.7
Pr(IPO,acq.|V)
0.6
1 yr.
0.5
2 yr.
5, 10 yr.
0.4
0.3
0.2
0.1
-400
-300
-200
-100
0
100
Log returns (%)
200
300
400
500
Fig. 6. Distribution of returns conditional on IPO/acquisition, predicted by the model, and estimated
selection function. Estimates from All subsample of IPO/acquisition sample.
round variances are about half the IPO/acquisition

variances, and round to round
p
standard deviations are lower by about 2: The return distribution is even more
stable with horizon in this case than in the IPO/acquisition sample. It does not even
begin to move to the right, as an unselected sample would do. The model captures
this effect, as the model return distributions are even more stable than in the IPO/
acquisition case.
6.4. Arithmetic returns
The second group of rows in the IPO/acquisition part of Table 6 presents
arithmetic returns. The average arithmetic return is an astonishing 698%. Sorted by
age, it rises from 306% in the rst six months, peaking at 1,067% in year 34 and
then declining a bit to 535% for years 5+. The standard deviations are even larger,
3,282% on average and also peaking in the middle years.
Clearly, arithmetic returns have an extremely skewed distribution. Median net
returns are half or less of mean net returns. The high average reects the small
possibility of earning a truly astounding return, combined with the much larger
probability of a more modest return. Summing squared returns really emphasizes the
ARTICLE IN PRESS
31
1-3
0-1
3-5
5+
-400
-300
-200
-100
0
100
Log Return
200
300
400
500
Fig. 7. Smoothed histogram of log returns, round-to-round sample. Each point is a normally weighted
kernel estimate. The numbers give age bins in years.
few positive outliers, leading to standard deviations in the thousands. These extreme
arithmetic returns are just what one would expect from the log returns and a
2
lognormal distribution: 100 e1:081=21:35 1 632%; close to the observed
698%. To make this point more clearly, Fig. 9 plots a smoothed histogram of log
returns and a smoothed histogram of arithmetic returns, together with the
distributions implied by a lognormal, using the sample mean and variance. This
plot includes all returns to IPO or acquisition. The top plot shows that log returns
are well modeled by a normal distribution. The bottom plot shows visually that
arithmetic returns are hugely skewed. However, the arithmetic returns coming from
a lognormal with large variance are also hugely skewed, and the tted lognormal
captures the right tail quite well. The major discrepancy is in the left tail, but kernel
density estimates are not good at describing distributions in regions where they slope
a great deal, and that is the case here.
Though the estimated 59% mean and 107% standard deviation of arithmetic
returns in Table 3 might have seemed surprisingly high, they are nothing like the
698% mean and 3,282% standard deviation of arithmetic returns with no sample
ARTICLE IN PRESS
32
1
3 mo.
0.9
0.8
Scale for Pr(new fin.|V)
0.7
Pr(New fin.|V)
0.6
1 yr.
0.5
2 yr.
0.4
5, 10 yr.
0.3
0.2
0.1
-400
-300
-200
-100
0
100
Log returns (%)
200
300
400
500
Fig. 8. Distribution of returns conditional on new nancing predicted by the model, and selection
function. All estimate of the round-to-round sample.
selection correction. The sample selection correction has a dramatic effect on

estimates of the arithmetic mean return.
6.5. Annualized returns
It might seem strange that so far I have presented total returns without
annualizing. The next two rows of Table 6 show annualized returns. The average
annualized return is 3:7 109 percent, and the average in the rst six months is
4:0 1010 percent. These must be the highest average returns ever reported in the
nance literature, which just dramatizes the severity of selection bias in venture
capital. The mean and volatility of annualized returns then decline sharply with
horizon.
The extreme annualized returns result from a small number of sensible returns that
occur over very short time periods. If you experience a moderate (in this dataset)
100% return, but it happens in two weeks, the result is a 100 224 1
1: 67 109 percent annualized return. Many of these outliers were checked by
hand, and they appear to be real. There is some question whether they represent
ARTICLE IN PRESS
-500
-400
-300
200
-200
400
-100
600
0
100
100 log return
200
800
1000
1200
Percent arithmetic return
300
1400
33
400
500
600
1600
1800
2000
Fig. 9. Smoothed histograms (kernel density estimates) and distributions implied by a lognormal. The top
panel presents the smoothed histogram of all log returns to IPO or acquisition (solid), using a Gaussian
kernel and s 0:20; together with a normal distribution using the sample mean and variance of the log
returns (dashed). The bottom panel presents a smoothed histogram of all arithmetic returns to IPO or
acquisition, using a Gaussian kernel and s 0:25; together with a lognormal distribution tted to the
mean and variance of log returns.
arms-length transactions, however. Ebay is a famous story (though not in the

dataset). Dissatised with the offering price, Ebay got one last round of venture
nancing at a high valuation, and then went public a short time later at an even
larger value. More typically, the dataset contains seed nancings quickly followed by
rst-stage nancings involving the same investors. It appears that in many cases, the
valuation in the initial seed nancing is a matter of little consequence, as the overall
allocation of equity will be determined at the time of the rst round, or the decision
could be made not to proceed with the start-up. (See for instance, the discussion in
Halloran, 1997.) While not data errors per se, huge annualized returns from seed to
rst round in such cases clearly do not represent the general rate of return to venture
capital investments. (This is analogous to the calendar time vs. event time issue
in IPO returns.) Below, I check the sensitivity of the estimates to these observations
in several ways.
ARTICLE IN PRESS
34
However, the log transformation again gives sensible numbers, so the large
average annualized returns are fundamentally a story of extreme volatility, not a
story about outliers or data errors. Average annualized log returns also decline
roughly with the inverse of the horizon. Again, this is what we expect from a selected
sample. For an unselected sample, we expect annualized returns to be stable across
horizon, and total returns to grow with horizon. In a selected sample, total returns
are stable with horizon, so annualized returns decline with horizon.6
In the round-to-round sample, arithmetic returns and annualized returns (not
shown) behave in the same way: arithmetic returns are large and very skewed with
huge standard deviations; annualized arithmetic returns are huge for short horizons,
and annualized returns decline quickly with horizon.
There is no right and wrong here. Statistics are just statistics. Skewed arithmetic
returns are what one expects from roughly lognormal returns with extreme variance.
The constant total returns and declining annualized returns with horizon are what
one expects from a roughly constant total log return distribution, generated by a
selection function of value and not of horizon. It is clear from this analysis that one
cannot do much of anything with the observed returns without correcting for
selection effects.
6.6. Subsamples
How different are returns to a new round, IPO, or acquisition? In addition to the
direct interest in these questions, I lumped outcomes together in the estimation, and
its important to check that this procedure is not unreasonable. The nal rows of
Table 6 present mean log returns across horizons for these subsamples of the roundto-round sample, and Fig. 10 collects the distribution of returns for different
outcomes, summing over all ages.
The mean log returns to IPO are a bit larger (81%) than returns to a new round or
acquisition (50%). Except for good returns to acquisitions in the rst six months,
and poor returns to new rounds and acquisitions after ve years, each category is
reasonably stable over horizon. Fig. 10 shows that the modal return to acquisition is
about the same as the modal return to IPO; the lower mean return to acquisition
comes from the larger left tail of acquisitions. The largest difference is the
surprisingly greater volatility of acquisition returns, and the much lower volatility of
new round returns. I conclude that lumping the three outcomes together is not a
gross violation of the data, and not worth xing at the large cost of adding
parameters to the already complex ML estimation. Most important, the gure
conrms that IPOS are similar to other fundings and revaluations, and not a
qualitatively different jackpot as in popular perception.
6
Also, we should not expect the average annualized arithmetic returns of Table 6 to be stable across
horizons, even in an unselected sample. In such a sample, the annualized average return is independent of
2
2
2
1=t
horizon, not the average annualized return. ER0!t Ee1=t ln R0!t emt=tts =2t ems =2t ; while
1=t
mt1=2s2 t 1=t
ms2 =2
e
: Small s and large t approximations do not work well in a dataset
ER0!t e
with huge s and occasionally very small t.
ARTICLE IN PRESS
35
New round
IPO
Acquired
-400
-300
-200
-100
0
100
Log return
200
300
400
500
Fig. 10. Smoothed histogram of returns, all ages, subsamples of the round-to-round sample. New
round, IPO, and Acquired are the returns for all rounds whose next nancing is a new round, IPO,
or acquisition, from initial investment to the indicated event.
7. How facts drive the estimates

Having seen estimates and a collection of stylized facts, it is time to see how the
stylized facts drive the estimates. This discussion can give us condence that the
estimates not driven by a few data points or by odd and untrustworthy aspects of the
data.
7.1. Stylized facts for mean and standard deviation
Table 6 nds average log returns of about 100% in the IPO/acquisition sample,
stable across horizons, and Fig. 3 shows about 10% of nancing rounds going public
or being acquired per year in the rst few years. These facts allow us to make a
simple back-of-the-envelope estimate of the mean and variance of venture capital
returns, correcting for selection bias. The same general ideas underlie the more
realistic, but hence more complex, maximum likelihood estimation, and this simple
ARTICLE IN PRESS
36
calculation shows how some of the rather unusual results are robust features of the
data.
Consider the very simple selection model: we see a return as soon as the log value
exceeds b. We can calibrate b to the average log return, or about 100%. Once again,
returns identify the selection function. The fraction
of projects that go public by year

btm
t is given by the right tail of the normal F p ; where m and s denote the mean and
ts
standard deviation of log returns. The 10% right tail of a standard normal is 1.28, so
the fact that 10% go public in the rst year means 1m
s 1:28:
1
0:78 or 78% would
A small mean m 0 with a large standard deviation s 1:28
generate the right tail. However, a small standard deviation s 0:1 or 10% and a
huge mean m 1 0:1 1:28 0:87 or 87% would also work. Which is it? The
second year separately identies m and s:
With a zero mean and a 78% standard
p 18% of rms have gone public,
deviation, we should see that by year 2, F 120
0:78 2
i.e., an additional 8% in year 2, which is roughly what we see. With a huge mean
m 87% and asmall standard
deviation s 10%; we predict that by year 2,

120:86
p
F5:2 100 8 106 %) rms have gone public.
essentially all (F
0:10 2
This is not at all what we seemore than 80% are still private at the end of year two.
To get rid of the high mean arithmetic returns, despite high variance, we need a
strongly negative mean log return. The same logic rules out this option. Given
1m
1 2
s 1:28; the lowest value of m 2 s we can achieve is given by m 64% and
1 2
s 128% (min m 12 s2 s.t. 1m
s 1:28), leading to m 2 s 0:18 and a reasonable
0:18
mean arithmetic return 100 e 1 20%: But a strong negative mean implies
that IPOs quickly cease and practically every rm goes out of business in short order
as the distribution
marches to the left. With m 64% and s 128%; we predict
120:64
p
F1:26 10:4% go public in two years. But 10% go public in
that F
1:28 2
the rst year,

so only

0.4% more go public in the second year. After that, things get
130:64
p
worse. F
F1:32 9:3% go public by year 3. Since 10% went public
1:28 3
already in year 1, this number reveals a distribution moving quickly to the left and
the oversimplication of this back-of-the-envelope calculation that ignores
intermediate exits.
To see the problem with failures, start with the fact from Fig. 3 that a steady small
percentageroughly 1%fail each year. The simplest failure model is a step
function at k, just like our step function at b for going public. The 1% tail of the
normal is 2.33 standard deviations from the mean, so to get 1% to go out of business
in one year, we need km
s 2:33: Using m 64% and s 128%; that means
k 2:33 1:28 0:64 3:62: A rm goes out of business when value declines to
100 e3:62 2:7% of its original value, which is both sensible and close to the
formal estimates in Tables 2 and 5. (The latter reports essentially twice this value,
since the selection for out of business is a linearly declining function of value
rather

p
than a xed cutoff.) But at these parameters, in two years, F k2m
2 s
ARTICLE IN PRESS
2:3420:64
p
2 1:28
37

p
p F 2:3430:64
F0:47 32% fail, and by 3 years F k3m
3 s
3 1:28
F0:12 45% fail!

In sum, the fact that rms steadily go public and fail, as seen in Fig. 3, means we
must have a log return distribution with a small meanno strong tendency to move
to the left or rightand a high variance. Then the tails, which generate rms that go
public or out of business, grow gradually with time. Alas, a mean log return near
zero and large variance imply very large arithmetic returns. This logic is compelling,
and suggests that these central ndings are not specic to the sample period.
The round-to-round sample has lower average returns, about 50% in Table 6. We
also see more frequent new nancings, about 30% per year in Fig. 7. The 30% right
tail is 0.52 standard deviations above the mean. Thus, we know from the rst year
that 0:5m
s 0:52: With a mean m 0; this implies s 0:50=0:52 96%: The lower
observed returns and greater probability of seeing a return are offsetting, giving
about the same estimate of standard deviation as for the IPO/acquisition sample.
It is comforting to see and understand the same underlying mean and standard
deviation parameters in the two samples, despite their quite different observed
means, standard deviations, and histories. This simple calculation shows why, and
why it is a robust feature of natural stylized facts.
7.2. Stylized facts for betas
How can we identify and measure betas? In the simple model that all rms go
public at b value, we would identify b by an increased fraction that go public
following a large market return, not by any change in return, since all observed
returns are the same (b: With a slowly rising selection function, we will see increased
returns as well, since the underlying value distribution shifts to the right. The formal
estimate also relies on more complex effects. For example, after a runup in the
market, many rms will go public, so the distribution of remaining project values
will be different than it would have been otherwise. These dynamic effects are harder
to characterize as stylized facts.
We can anticipate that these tendencies will be difcult to measure, so that beta
estimates might not be precise or robust. With 100% per year idiosyncratic risk, a
typical 15% (1s) rise in the market is a small risk, and shifts the distribution of
returns only a small amount. In the simple model, a 15% rise
10in
the market raises the
F1:28 10% to
fraction
of
rms
that
go
public
in
one
year
from
F
0:78

F1:09 14%: The actual selection functions rise slowly, so moving the
F 10:15
0:78
return distribution to the right 15% will push even fewer rms over the border.
Similarly, the huge residual standard deviation means that the R2 s are low, so
market model return regression estimates will be imprecise.
Still, let us see what facts can be documented about returns and fates conditional
on index returns. Table 7 presents regressions of observed returns on the S&P500
index return. With arithmetic returns, the intercepts (alpha) are huge. At 462%, this
is probably the largest alpha ever claimed in a nance paper, though it surely reects
the severe selection bias in this sample rather than a golden-egg-laying goose. The
ARTICLE IN PRESS
38
Table 7
Market model regressions
a (%)
sa
IPO/acq. arithmetic
IPO/acq. log
Round to round, arithmetic
Round to round, log
462
92
111
53
111
3.6
67
1.8
Round only, arithmetic

Round only, log
IPO only, arithmetic
IPO only, log
Acquisition only, arithmetic
Acquisition only, log
128
49
300
66
477
77
67
1.8
218
4.8
95
9.8
sb
R2 (%)
2.0
0.4
1.3
0.0
0.6
0.1
0.6
0.1
0.2
0.8
0.1
0.0
0.7
0.0
2.1
0.7
0.8
0.8
0.6
0.1
1.5
0.2
0.5
0.3
0.3
0.0
0.0
2.1
0.3
2.6
Note: Market model regressions are Rt!tk a bRm

t!tk et!tk (arithmetic) and ln Rt!tk a
b ln Rm
t!tk et!tk (log). For an investment made at date t and a new valuation (new round, IPO,
acquisition) at t k; I regress the return on the corresponding S&P500 index return for the period
t ! t k: Standard errors are plain OLS ignoring any serial or cross correlation.
32% arithmetic alpha in the selection-bias-corrected Table 3 pales by comparison.

Once again, though the selection-bias-corrected estimates leave some puzzle, the
selection bias correction has dramatic effects on the uncorrected estimates.
The beta for arithmetic returns is large at 2.0. There is a tendency for market
returns to coincide with even larger venture capital returns. Log returns trim the
outliers, however, and produce a lower beta of 0.4. The R2 values in these regressions
are tiny, as expected. For this reason, betas are poorly measured, despite the huge
sample and optimistic plain-vanilla OLS standard errors.
The round-to-round regressions produce lower betas still, suggesting that much of
the measured beta comes from a tendency to go public at high market valuations
rather than a tendency for new rounds to be more highly valued when the market is
high. Splitting into new round, IPO, and acquisition categories we see this pattern
clearly. The positive betas come from the IPOs.
Fig. 11 graphs the time series of the fraction of outstanding rms in the IPO/
acquisition sample that go public each quarter, along with the previous years
S&P500 returns. (The fraction that goes public is a two-quarter moving average.) If
you look hard, you can see that IPOs increase following good market returns in
19921993, 19961997, and 19992000. (There was a huge surge in IPOs in the last
two years of the sample. However, there was also a huge surge of new projects, so the
fraction of outstanding rms that go public only rises modestly as shown.) 1992 and
19961997 also show a modest correlation between average IPO returns and the
S&P500 index. For the IPOs, increased numbers rather than larger returns drive the
estimated betas.
Fig. 12 graphs the same time series for rms in the IPO/acquisition sample that are
acquired. Here we see no tendency at all for the frequency of acquisitions to rise
following good market returns. However, we do see that the returns to acquired
ARTICLE IN PRESS
39
10
Percent IPO
150
Avg. IPO returns
100
S&P 500 return
25
75
1988
1990
1992
1994
1996
1998
2000
Fig. 11. Percentage of outstanding projects going public, percent average log returns for projects going
public (right scale), and previous years percent log S&P500 return. Percentage of projects going public
and their returns are two-quarter moving averages. IPO/acquisition sample.
rms track the S&P500 index well, with a scale factor of about two or three. This
graph suggests that returns rather than greater frequency of acquisitions drive a beta
estimate among acquisitions. However, this picture is not conrmed in Table 7,
which found negative betas. There are more observations in later years, so the
regression and this graph weight observations differently.
A similar gure for new rounds in the round-to-round sample shows no tendency
for an increased frequency of nancing, and a barely discernible tendency towards
higher values on the tail of stock market rises. The maximum likelihood beta
estimates in the round-to-round sample are correspondingly lower and less precise,
and are driven by the acquisition and IPO outcomes in that sample.
In sum, the correlation of observed returns with market returns, and the
correlation of the frequency of observed new nancing or acquisition with market
returns, form the basic stylized facts behind beta estimates. The stylized facts are
there: the frequency of IPOs rises when the market rises, and the valuation of
acquisitions rises when the market rises. However, the stylized facts are much weaker
than those that drive average returns and the variance of returns. This weakness
explains why the intercept and beta estimates of the formal model are not
ARTICLE IN PRESS
40
6
100
Percent acquired
80
Average return
60
40
30
S&P500 return
20
0
20
10
-10
1988
1990
1992
1994
1996
1998
2000
Fig. 12. Percent of outstanding projects acquired, average log returns of acquisitions (right scale), and
previous years S&P500 return, IPO/acquisition sample. Percent acquired is a two-quarter moving average.
particularly well estimated or stable across subsamples or variations in technique,

while the average and standard deviation of log returns are quite stable. This
weakness also explains why I have not extended the estimation, for example to threefactor betas or other risk corrections.
8. Testing a 0
An arithmetic return of 59% and a 32% arithmetic alpha are still uncomfortably
large. We have already seen that they result from a mean log return near zero, the
1 2
large volatility of log returns, and em2s : We have seen in a back-of-the-envelope

way that m 50% would produce IPOs that cease after a few years and all rms
soon failing. But perhaps the more realistic model and formal estimate do not speak
so strongly against a 0: What if we change all the parameters? In particular, can
we accept the high mean arithmetic return, but imagine a b of three to ve so that the
high mean return is explained? The stylized facts behind high volatility are
compelling, but those driving us to a small beta are not so convincing. Can we
imagine that the data are wrong in simple ways that would overturn the nding of a
ARTICLE IN PRESS
41
Table 8
Additional estimates and tests for the IPO/acquisition sample
E ln R s ln R g
ER sR
All, baseline
15
89
7.1
1.7
86 59
a0
ER 15%
0.9
3.3
82
60
30
3.3
2.5
73 34
60 15
Pre-1997
Dead 2000
No p
11
36
11
81
59
115
11
0.8 80 46
27
0.3 59 58
4.0
0.9 114 85
107 32
93
64
0.0
94 48
69 48
152 67
1.9
25
1.0 3.8
2.6
23
28
0.9 3.9 15
1
3.4 28
w2
9.6
0.8
9.6 1.0 3.6 4.4
0.3 150
0.7 4.9 31
1.1 11
0.6 5.8
1,428
2,523
170
Note: a 0 imposes a 0 on the estimation, by always choosing g so that, given the other parameters,
the arithmetic a calculation is zero. ER 15% imposes that value on the no-d estimation, choosing g so
that the arithmetic average return calculation is always 15%. w2 gives the likelihood ratio statistic for these
parameter restrictions. Each statistic is w2 1 with a 5% critical value of 3.84. Pre-1997 limits the data
sample to January 1 1997, treating as still private any exits past that date. Dead 2000 assumes that
any project still private at the end of the sample goes out of business. No p removes the measurement
error.
Table 9
Additional estimates for the round-to-round sample
E ln R
s ln R
21
1.7
1.3
4.7
1.9
21
19
1.5
2.2
1.4
1.0
6.8
9.9
0.4
1.0
1.0
19
108
11
0.4
0.3
1.2
5.1
6.4
1.8
2.2
7.2
ER
sR
45
0.6
All, baseline
20
84
7.6
0.6
84
59
100
a0
ER 15%
3.6
8.9
77
69
27
8.9
1.9
72
69
27
15
86
74
Pre-1997
Dead 2000
No p
21
32
16
75
76
104
10
16
1.6
0.4
0.9
0.9
75
74
103
52
65
77
87
91
133
0.0
40
47
60
w2
1,807
3,060
864
Note: See note to Table 8.
high a? All these questions point naturally to an estimate with restricted parameters
such that a 0; and a likelihood ratio test.
Table 8 presents additional estimates for the IPO/acquisition sample, starting with
a test of a 0: (I solve Eq. (7) for the value of g that, given the other parameters,
results in a 0; and I x g at that value in the estimation.) Table 10 collects
asymptotic standard errors. Imposing a 0 lowers the mean log return from 15% to
0:9%: Together with a slightly lower standard deviation, the mean arithmetic
return is cut in half, from 59% to 34%. However, imposing a 0 changes the
decomposition of mean log return, lowering the log model intercept from 7:1% to
30%; and raising the slope coefcient d from 1.7 to 2.5. Interestingly, the estimate
ARTICLE IN PRESS
42
Table 10
Asymptotic standard errors for Tables 8 and 9 estimates
IPO/acquisition sample
g
a0
ER 15%
Pre-1997
Dead 2000
No p
1.2
0.7
1.1
Round-to-round sample
0.06
0.7
0.6
0.59
0.65
0.03
0.01
0.13
0.01
0.8
1.1
0.11
0.05
0.08
1.1
1.2
1.1
0.42
0.08
0.37
0.04
0.03
0.02
0.12
0.16
0.17
0.8
1.1
1.3
1.0
1.2
0.01
0.6
0.6
0.4
0.3
0.04
0.02
0.03
0.01
0.4
0.6
0.12
0.06
0.08
1.1
1.1
0.8
0.9
1.1
0.2
0.01
0.00
0.02
0.06
0.02
0.03
0.4
0.5
does not just raise beta. It achieves half of the alpha decline via the difcult route of
lowering mean returns. Apparently, there is strong sample evidence against the highbeta parameterization, despite the apparent weakness of stylized facts seen in the last
section. The estimate also increases measurement error, to try to handle observations
that now cause trouble. Alas, the statistical evidence against this parameterization is
strong. Imposing a 0; the log likelihood declines by 1,428/2. Compared to the 5%
w2 1 critical value of 3.84, the a 0 restriction is spectacularly rejected.
The round-to-round sample in Table 9 behaves similarly. The average log return
declines from 20% to 3:6%; and the average arithmetic return is cut in half from
59% to 27%. The intercept declines dramatically from g 7:6% to g 27%;
and the slope rises from 0:6 to 1:9: But the w2 1 likelihood ratio statistic is 1,807, an
even more spectacular rejection.
So far, the estimates raise slope coefcients a good deal in order to lower alphas.
We might want to keep the estimate from following this path in order to examine the
evidence against the core troubling estimate of high average arithmetic returns,
rather than to excuse such returns by large poorly measured betas. In the ER 15%
rows of Tables 8 and 9, I impose an average arithmetic return of 15%, the same as
the S&P500 in this sample, estimating the mean and variance of returns directly, i.e.,
restricting the no d estimate of Tables 3 and 4. The model might have kept the
high standard deviation, and matched it with a 50% or so mean log return in order
to reduce m 12 s2 : Instead, the dynamic evidence for a mean log return near zero is
so strong that the estimate keeps it, with E ln R 3:3% in the IPO/acquisition
sample and E ln R 8:9% in the round-to-round sample. The estimate reduces
standard deviation accordingly, to 60% in the IPO/acquisition sample and 69% in
the round-to-round sample. These variance reductions are just enough to produce
1 2
the desired 15% mean arithmetic return via em2s : However, this reduction in
variance does great damage to the models ability to t the dynamic pattern of new
nancing. The measurement error probabilities rise to 28% in the IPO/acquisition
sample, and to 9.9% in the round-to-round sample. The w2 1 likelihood ratio
statistics are 2,523 for IPO/acquisition and 3,060 in the round-to-round samples,
even more decisively rejecting the ER 15% restriction.
ARTICLE IN PRESS
60
43
Dash: IPO/Acquisition
Solid: Out of business
50
=0, others unchanged
Percentage
40
30
=0
20
10
Data
4
5
Fig. 13. Simulated fates of venture capital investments, imposing a 0: In the a 0 case (triangles), I
impose the condition that the arithmetic a is zero and maximize likelihood over the remaining parameters.
In the a 0; others unchanged case (no symbols), I change the intercept g to produce a 0; leaving
other parameters unchanged. Squares give the data. Dashed lines plot IPO/acquisitions; solid lines plot
failures.
Where is the great violence to the data indicated by these likelihood ratio
statistics? Fig. 13 compares the simulated fates in the a 0 restricted models to the
simulated fates with the baseline estimates, and Table 11 characterizes the simulated
distribution of observed returns with the various restricted models. Table 11 is meant
to convey the same information as the return distribution in Figs. 6 and 8 in more
compact form.
I start by lowering the intercept g to produce a 0 with no change in the other
parameters. The a 0; others unchanged lines of Fig. 13 shows that this
restriction produces far too many bankruptcies and too few IPO/acquisitions. (The
dashed line with no symbols, representing the estimates prediction of IPO/
acquisitions, is well below the dashed line with squares, representing the data; and
the solid line with no symbols, representing the estimates prediction of out of
business projects, is well above the solid line with squares, representing failures in the
data.) As the back-of-the-envelope calculation suggested, a low mean log return
implies that the distribution of values moves to the left over time, so we have an
inadequate right tail of successes and too large a left tail of failures.
ARTICLE IN PRESS
44
Table 11
Moments of simulated returns to new nancing or acquisition under restricted parameter estimates
1. IPO/acquisition sample
Horizon (years):
1/4
2. Round-to-round sample
10
1/4
10
(a) E log return (%)

Baseline estimate 21
a0
11
ER 15%
8
78
42
29
128
72
50
165
101
70
168
103
71
30
16
19
70
39
39
69
34
31
57
14
13
55
10
11
(b) s log return (%)

Baseline estimate 18
a0
13
ER 15%
9
68
51
35
110
90
62
135
127
91
136
130
94
16
12
11
44
40
30
55
55
38
60
61
44
60
61
44
The a 0 lines of Fig. 13 present the simulated histories when I impose a 0;

but allow ML to search over the other parameters, in particular raising the slope
coefcient b and measurement error, so as to give a 0 while keeping a large mean
arithmetic return. Now the estimate can match the pattern of successes (the dashed
lines with triangles and squares are close), but it still predicts far too many failures
(the solid line with triangles is far above the solid line with squares). The 20% lower
mean log and 30% lower mean arithmetic return in this estimate still leave a
distribution that marches off to the left too much. The ER 15% restriction and the
round-to-round sample behave similarly.
In Table 11, the mean returns to new nancing or acquisition under the restricted
models are often less than half the mean returns under the unrestricted model and in
the data. The standard deviations are often a poor match in some of the
parameterizations. The restricted estimates also miss facts underlying the beta
estimates, although I do not graph this phenomenon.
In sum, the data speak strongly against lowering the arithmetic alpha to zero,
either by lowering mean arithmetic returns or by raising betas. To believe such a
parameterization, we must believe that beta is much larger than estimated, we must
believe that the data are measured with much more error, we must believe that the
data substantially understate the frequency and timing of failure (as indicated by
Fig. 13), and we must believe that the sample systematically overstates the returns to
IPO, acquisition, and new nancing by as much as a factor of two, as indicated by
Table 11.
9. Robustness
I check that the anomalous IPO market at the end of the sample, measurement
error, and the imputation of returns to out-of-business projects do not affect the
results.
ARTICLE IN PRESS
45
9.1. End of sample

We might suspect that the results depend crucially on the anomalous behavior
of the IPO market during the late 1990s, and the unfortunate fact that the sample
stops in June of 2000, just after the rst Nasdaq crash. (This fact is not a
coincidencethe data collection for this project was commissioned by a now defunct
dot-com.)
To address this concern, the pre-1997 subsample uses no information
after January 1997. I ignore all rounds that are not started by January 1997,
and I treat all rounds started before then that have not yet gone public,
been acquired, had another round, or gone out of business by January 1997
as still private. The Dead 2000 sample assumes that all rms still private
as of June 2000 go out of business on that date. This experiment also provides
a way to address the living dead bias: some rms that are reported as still
alive are probably really inactive and worthless. Assuming all inactive rms
are worthless by the end of the sample gives a bound on how severe that bias
could be.
As Table 2 shows, about two-thirds of the venture capital nancing rounds
begin after January 1997, so the concern that the results are special to the subsample
is not unfounded. However, the main difference in fates is that rms are much more
likely to fail in the post-1997 sample. The fraction that go public, etc., is virtually
identical. If we assume that all rms alive in June 2000 go out of business, we
increase the out-of-business fraction dramatically, at the expense of the still
private category.
In Tables 8 and 9, the mean and standard deviation of log returns are essentially
the same in the pre-1997 sample as in the base case. The main difference is the split
of the mean return between slope and intercept for the IPO/acquisition sample.
The slope coefcient d switches sign from 1:7 to 0:8; and the intercept g rises
from 7:1% to +11%. The association of IPOs with the stock price rise of the late
1990s is the major piece of information identifying the slope d: Since volatility is
unchanged and the mean log return is unchanged, mean arithmetic returns
are essentially unchanged in the pre-1997 sample. Since the slopes decline,
arithmetic alphas actually increase in the pre-1997 sample, to 48% in Table 8 and
40% in Table 9.
Assuming that all rms still private in June 2000 go out of business on that date
plays havoc with the estimate. The failure cutoff k increases to 150% of its initial
value in Table 8 and 108% in Table 9, naturally enough, as the chance of failure
increases dramatically. The other parameters change a bit, as they must still account
for the successes in the pre-2000 data despite much higher k. Both samples show a
much larger mean log return, and the IPO/acquisition sample shows a somewhat
smaller variance.
In the end, the mean arithmetic returns and alphas are the same or higher in the
pre-1997 and Dead 2000 samples. As always, the idiosyncratic variance remains large
and it is not paired with a huge negative mean. Thus, neither the late 1990s boom nor
a living dead bias is behind the central results.
ARTICLE IN PRESS
46
9.2. Measurement error and outliers

How does the measurement error process affect the estimates? In the rows of
Tables 8 and 9 marked no p, I remove the measurement error process. This change
raises the estimated standard deviation from 89% to 115% in Table 8 and from 84%
to 104% in Table 9. Absent measurement error, we need a larger variance to
accommodate tail returns. The mean log return is unaffected. Higher variance alone
would drive more rms to failure, so the failure cutoff k drops from 25% to 11% in
Table 8 and 21% to 11% in Table 9. All the other estimated parameters are basically
unchanged. Raising the variance raises mean arithmetic returns to 85% and a to
67% in Table 8, and mean arithmetic returns to 65% and a to 60% in Table 9.
Repeating the whole set of estimations without measurement error, the largest
difference, in addition to the larger variance, is much less stability in slope
coefcients d across subsamples. A few large returns, very unlikely with a lognormal
distribution, drive the d estimates without measurement error. The estimates vary as
the few inuential data points jump in and out of subsamples.
The likelihood ratio statistic for p 0 is 170 in the IPO/acquisition sample and
864 in the round-to-round sample. The 5% critical value for a w21 is 3.84. Whether
we interpret the measurement error process as such, or as a device to induce a fatter
tail in the true return process,7 the model wants it.
The measurement error process does not just throw out large returns, which are
plausibly the most interesting part of venture capital. It largely throws out
reasonable returns that occur in a very short time period, leading to very large
annualized returns. Even if not errors, these events are a separate phenomenon from
what most of us think as the central features of venture capital. Venture capital is
about the possibility of earning a very large return in a few years, not about the
chance of only doubling your money in a month.
To document this interpretation, I examine outliers, the data points that
contribute the greatest (negative) amount to the log likelihood in each estimate.
Without measurement error, the biggest outliers are IPO/acquisitions that have
moderately large positiveand negativereturns in a short time span, not large
returns per se. With measurement error, the outliers are old (eight to ten years)
projects that eventually go public or are acquired with very low returns5% to 20%
of initial value. This nding is sensible. Since low values exit, and the probability of
going public is very low at a low value, it is hard to attain a very low value and go
public without failing along the way.
To check further that the high mean returns and alphas are not driven by
anomalous quick successes, implying huge annualized returns, I try replacing the
actual age of returns in the rst year with 1 year or less, i.e. summing the
7
We cannot interpret this exact specication of the measurement error process as a fat-tailed return
distribution. The measurement error distribution is applied only once, and does not cumulate. A fat-tailed
return distribution, or, equivalently, the addition of a jump process, is an interesting extension, but one I
have not pursued to keep the number of parameters down and to preserve the ease of making
transformations such as log to arithmetic based on lognormal formulas.
ARTICLE IN PRESS
47
probability of an IPO over the rst year rather than using the probability of
achieving the IPO on the reported date. This variant has practically no effect at all
on the estimated parameters.
In sum, the measurement error complication does not drive the large alphas. Quite
the opposite, adding measurement error reduces the volatility-induced mean
arithmetic returns and alphas by accounting for the occasional quick large returns.
9.3. Returns to out-of-business projects
So far, I have implicitly assumed that when a rm goes out of business, the
investor receives whatever value is left. What if, instead, investors get nothing when
the rm goes out of business? This change adds a lumpy left tail to the return
distribution. Perhaps this lumpy left tail is enough to get rid of the troublesome
alphas. Mean log returns become 1; and the standard deviation of log returns
1; but we can still characterize the mean and standard deviation of arithmetic
returns.
To answer this question, I simulate the model at the baseline parameter estimates,
and nd the probability and value of all the various outcomes. I then calculate the
annualized expected arithmetic return, assuming that investors get zero return for
any project that goes out of business. (Since we are aggregating payoffs at different
horizons, I calculate the arithmetic discount rate that sets the present value of the
cash ows to one.) The average arithmetic return declines only from 58:72% to
58:38%: This modication has so little effect because the failure values k are quite
low, around 10% of the initial investment, and only 9% of rms fail. Losing the last
10%, in the 9% of investments that are down to 10% of initial value, naturally has a
small effect on average returns.
10. Comparison to traded securities

If we admit large arithmetic mean returns, standard deviations, and arithmetic
alphas in venture capital, are these ndings unique, or do similar traded securities
behave the same way?
Table 12 presents means, standard deviations, and market model regressions for
individual small Nasdaq stocks. To form the subsamples, I take all stocks that have
market value below the indicated cutoffs in month t, and I examine their returns
from month t 1 to month t 2: I lag by two months to ensure that erroneously low
prices at t do not lead to spuriously high returns from t to t 1; though results with
no lag (selection in t, return from t to t 1) are in fact quite similar. I examine
market value cutoffs of two million, ve million, 10 million and 50 million dollars.
The average venture capital nancing round in my sample raises $6.7 million. The
rst ve deciles of all Nasdaq market value observations in this period occur at 5, 10,
1
1
2
; 10
; 10
; and 12
17, 27, and 52 million dollars, so my cutoffs are approximately the 20
quantiles of market value. Small Nasdaq stocks have a large number of missing
return observations in CRSP data, most due to no trading. I ignore missing returns
ARTICLE IN PRESS
48
Table 12
Characteristics of monthly returns for individual Nasdaq stocks
N
MEo$2M, arithmetic
MEo$2M, log
ME o$5M, arithmetic
MEo$5M, log
ME o$10M, arithmetic
MEo$10M, log
All Nasdaq, arithmetic
All Nasdaq, log
22,289
72,496
145,077
776,290
ER
sR
ERvw
62
1.9
37
5.1
24
5.8
14
3.4
175
113
139
103
118
93
81
72.2
54
53
15
27
26
12
31
3.1
22
29
16
(4.0)
(2.6)
(1.8)
(1.3)
(1.1)
(0.9)
(0.3)
(0.3)
R2 (%)
0.49
0.40
0.60
0.57
0.76
0.66
0.91
0.97
0.18
0.30
0.44
0.77
0.99
1.3
3.1
4.6
Note: Mean returns, alphas, and standard deviations are annualized

p percentages: means and alphas are
multiplied by 1,200, and standard deviations are multiplied by 100 12: ERvw denotes the value-weighted
m
tb
i
Rit Rtb
mean return. a; b; and R2 are from market model regressions,
t a bRt Rt et for

i
tb
m
tb
m
i
arithmetic returns and ln Rt ln Rt a b ln Rt ln Rt et for log returns, where R is the
S&P500 return and Rtb is the three-month T-bill return. The sample consists of all Nasdaq stocks on
CRSP, January 1987December 2001, including delisting returns. Each set of rows considers a different
upper limit for market value (ME) in month t. Returns are then calculated from month t 1 to month
t 2: Missing return data (both regular and delisting) are ignored if the security is still listed the following
period, or if the previous period included a valid delisting return. Other missing returns are assumed to be
100%: Log returns are computed ignoring observations with 100% returns. Parentheses present simple
pooled OLS standard errors ignoring serial or cross correlation.
when the company remains listed at t 3; or if month t 1 is a delisting return. I

treat all remaining missing returns as 100%; to give the most conservative estimate
possible. This sample is different than the CRSP deciles in several respects. First, the
cutoff for inclusion is a xed dollar value rather than a decile in the selection month;
as a result the numbers and fraction of the Nasdaq in each category uctuate over
time. Second, I rebalance each month rather than once per year. I make both
changes in order to better control the characteristics of the sample. With 100%
standard deviations, stocks do not keep their capitalizations for long.
The estimates in Table 12 for the smallest Nasdaq stocks are surprisingly similar to
the venture capital estimates. First, the mean arithmetic return of the smallest
Nasdaq stocks is 62%, comparable to the 59% mean return in the baseline venture
capital estimates of Tables 3 and 4. As we increase the size cutoff, mean returns
gradually decline. The mean return of all Nasdaq stocks is only 14.2%, similar to the
S&P500 return in this time period. Value-weighted averages are lower, but still quite
highthe basic result is not a feature of only the smallest stocks in each category.
Second, these individual stock returns are very volatile. The smallest Nasdaq
stocks have a 175% annualized arithmetic return standard deviation, even larger
than the 107% from the baseline venture capital estimate of Table 3. The $5 million
and $10 million cutoffs produce 139% and 118% standard deviations, quite similar
to the 107% of Table 3. Even stocks in the full Nasdaq sample have a quite high
80.7% standard deviation.
Third, as in the venture capital estimates, large arithmetic mean returns come from
the large volatility of log returns, not a large mean log return. The mean and
ARTICLE IN PRESS
49
standard deviation of the small stock log returns are 1:9% and 113%, comparable
to the 15% and 89% venture capital estimates in Table 3.
Last, and most important, the simple market model regression for the smallest
Nasdaq stocks leaves a 53% annualized arithmetic alpha, even larger than the 32%
venture capital alpha of Table 3. This is a feature of only the very smallest stocks. As
we move the cutoff to $5 and $10 million, the alpha declines rapidly to 27%
comparable to the 32% of Table 3and 12%, and it has disappeared to 3% in the
overall Nasdaq. The slope coefcients b are not large at 0.49, gradually rising with
size to 0.91. These values are smaller than the 1.9 b of Table 3. As in Table 3, the log
market models leave substantial negative intercepts, here 15% declining with size
to 22%; similar to the 7% of Table 3. As in Table 3, the arithmetic alpha is
induced by volatility, not by a large intercept in the log market model. The R2 are of
course tiny with such large standard deviations, though the higher R2 for the log
models suggest that they are a better statistical t than the arithmetic market models.
The nding of high arithmetic returns and alphas in very small Nasdaq stocks is
unusual enough to merit a closer look. In Table 13, I form portfolios of the stocks in
the samples of Table 12. By the use of portfolio average returns, the standard errors
control for cross correlation that the pooled statistics and regressions of Table 12
ignore. In the rst panel of Table 13 we still see the very high average returns in the
small portfolios, declining quickly with size. Despite the large standard deviations,
the mean returns appear statistically signicant.
The large average returns (and alphas) are not seen in the CRSP small decile. The
CRSP small Nasdaq decile is comparable to the $10 million cutoff. To see the high
returns, you have to look at smaller stocks and control the characteristics more
tightly than annual rebalancing in CRSP portfolios allows.
In the second panel of Table 13, I run simple market model regressions for these
portfolios on the S&P500 return. The market model regressions of the individual
stocks in Table 12 are borne out in portfolios here. The smallest stocks show a 61.6%
arithmetic alpha, and the second portfolio still shows a 31.6% arithmetic a: Though
the standard errors are greater than those of Table 12, the alphas are statistically
signicant. By contrast, the relatively much smaller 12% a of the CRSP rst (small)
decile is not statistically signicant. Again, the behavior of my smallest group is not
reected in the CRSP small decile.
Is the behavior of very small stocks just an extreme size effect, explainable by a
large beta on a small stock portfolio? In the third panel of Table 13, I run regressions
with the CRSP decile 1 on the right-hand side. Alas, this hope is not borne out.
Though regressions of the small stock portfolios on the CRSP decile 1 return give
substantial betas, up to 1.4, they also leave a substantial alpha. The alpha is 43% in
the $2 million or less portfolio, declining quickly to 18% in the $5M portfolio and
vanishing for larger portfolios. The small Nasdaq portfolio returns lose the
correlation with the CRSP small Nasdaq decile, also suggesting something more
than an extreme size effect.
Perhaps these very small stocks are value stocks as well as small stocks. The
nal panel of Table 13 runs a regression of the small Nasdaq portfolios on the three
FamaFrench factors. Though the SMB loading is quite large, up to 1.9, nonetheless
ARTICLE IN PRESS
50
Table 13
Characteristics of portfolios of very small Nasdaq stocks
Equally weighted, MEo
Value weighted, MEo
CRSP Dec1
$2M
$5M
$10M
$50M
$2M
$5M
$10M
$50M
ER
s.e.
sR
22
8.2
32
71
14
54
41
9.4
36
25
8.0
31
15
6.2
24
70
14
54
22
9.1
35
18
7.5
29
10
5.8
22
Rt Rtb
t
a
sa
b
a b RtS&P500 Rtb
t et
12
62
32
7.7
14
9.0
0.73
0.65
0.69
16
7.6
0.67
5.4
5.5
0.75
60
14
0.73
24
8.6
0.71
8.5
7.0
0.69
0.6
4.8
0.81
Rt Rtb
t
r
a
sa
b
a b Dec1t Rtb
t et
1.0
0.79
0.92
0
43
18
8.4
3.6
1
1.4
1.1
0.96
4.7
2.1
0.9
0.96
2.7
1.9
0.7
0.78
43
8.9
1.3
0.92
11
3.5
1.0
0.96
2.3
2.0
0.9
0.91
5.7
2.5
0.7
Rt Rtb
t
a
sa
b
s
h
a b R MRFt s SMBt h HMLt et

5.1
57
26
10
1.9
5.5
12
7.6
5.8
3.5
0.8
0.6
0.7
0.7
0.8
1.7
1.9
1.6
1.5
1.4
0.5
0.2
0.3
0.4
0.4
55
12
0.7
1.8
0.1
18
7.3
0.7
1.5
0.3
1.9
5.2
0.7
1.5
0.4
7.0
2.7
0.9
1.3
0.4
Note: Means, standard deviations, and alphas are annualized percentages. Portfolios are re-formed
monthly. Market equity at date t is used to form a portfolio for returns from t 1 to t 2: CRSP Dec1 is
the CRSP smallest decile Nasdaq stock portfolio. Rtb is the three-month T-bill return. The sample is for
the period January 1987 to December 2001.
the smallest portfolio still leaves a 57% alpha, and even the $5 million portfolio still
has a 25% alpha. As expected, the alphas disappear in the larger, and especially
value-weighted, portfolios.
In conclusion, it seems that something unusual happens to the very smallest of
small Nasdaq stocks in this period. It could well be a period-specic event rather
than a true ex ante premium. It could represent exposure to an unusual risk factor.
These Nasdaq stocks are small, thinly traded, and illiquid; the CRSP data show
months of no trading for some of them. For the purposes of this paper, however, the
main point is that alphas of 30% or more are observed in traded securities with
similar characteristics as the venture capital investments.
The small-stock portfolios are natural candidates for a performance attribution of
venture capital investments. It would be gratifying if the venture capital investments
showed a beta near one on the small-stock portfolios. Then the 50% or more average
arithmetic returns in venture capital would be explained, in a performance
ARTICLE IN PRESS
51
attribution sense, by the 50% or more average arithmetic returns in the small-stock
portfolios during this time period. Alas, Tables 3 and 4 only found b of 0.5 and 0.2
on the small stock portfolios, and leave substantial alphas. Venture capital betas are
poorly measured, so perhaps the b really is larger. However, the only conclusion
from the evidence is that venture capital shows an anomalously large arithmetic
alpha in this period, and the very smallest Nasdaq stocks also show an anomalously
large arithmetic alpha. The two events are similar, but they are not the same event.
11. Extensions
There are many ways that this work can be extended, though each involves a
substantial investment in programming and computer time, and could strain the
stylized facts that credibly identify the model.
My selection function is crude. I assume that IPO, acquisition, and failure are only
a function of the rms value. One might desire separate selection functions for IPO,
acquisition, and new rounds at the (not insubstantial) cost of four more parameters.
The decision to go public could well depend on the market, as well as on the value of
the particular rm, and on rm age or other characteristics. (Lerner (1994) nds that
rms are more likely to go public at high market valuations, and more likely to
employ private nancing when the market is low.) I allow for missing data, but I
assume that data errors are independent of value. It might be useful to estimate
additional selection functions for missing data, i.e., rms that subsequently go public
are more likely to have good round data in the VentureOne dataset, or that rms
which go public at large valuations are more likely to have good nal valuation data.
I do no modeling of the decision to start venture capital projects, yet it is clear in the
data that this is an endogenous variable.
My return process is simple. The risks (betas, standard deviation) of the rm are
likely to change as its value increases, as the breakout by nancing round suggests.
Multiple risk factors, or evaluation with reference to carefully tailored portfolios of
traded securities, are obvious generalizations. I do not attempt to capture crosscorrelation of venture capital returns, other than through identiable common
factors. Such residual cross-correlation is of course central to the portfolio question.
More and better data will certainly help. Establishing the dates at which rms
actually go out of business is important to this estimation procedure. Many more
projects were started in the late 1990s with very high market valuations than before
or since. When we learn what happened to the post-crash generation of venture
capital investments, the picture might change.
References
Admati, A., Peiderer, P., 1994. Robust nancial contracting and the role of venture capitalists. Journal of
Finance 49, 371402.
Berk, J., Green, R., Naik, V., 2004. Valuation and return dynamics of new ventures. Review of Financial
Studies 17, 135.
ARTICLE IN PRESS
52
Bygrave, W., Timmons, J., 1992. Venture Capital at the Crossroads. Harvard Business School Press,
Boston.
Chen, P., Baierl, G., Kaplan, P., 2002. Venture capital and its role in strategic asset allocation. Journal of
Portfolio Management 28, 8390.
Gilson, R., Schizer, D., 2003. Understanding venture capital structure: a tax explanation for convertible
preferred stock. Harvard Law Review 116, 874916.
Gompers, P., Lerner, J., 1997. Risk and reward in private equity investments: the challenge of performance
assessment. Journal of Private Equity, 512.
Gompers, P., Lerner, J., 2000. Money chasing deals? The impact of fund inows on private equity
valuations. Journal of Financial Economics 55, 281325.
Halloran, M., 1997. Venture Capital and Public Offering Negotiation, third ed. Aspen, Gaithersbug.
Kaplan, S., Schoar, A., 2003. Private equity performance: returns, persistence and capital ows. Journal of
Finance, forthcoming.
Kaplan, S., Stromberg, P., 2003. Financial contracting theory meets the real world: an empirical analysis
of venture capital contracts. Review of Financial Studies, forthcoming.
Kaplan, S., Sensoy, B., Stromberg, P., 2002. How well do venture capital databases reect actual
investments? Unpublished working paper, University of Chicago.
Lee, P., Wahal, S., 2002. Grandstanding, certication, and the underpricing of venture capital backed
IPOs. Unpublished working paper, Emory University.
Lerner, J., 1994. Venture capitalists and the decision to go public. Journal of Financial Economics 35,
293316.
Ljungqvist, A., Richardson, M., 2003. The cash ow, return and risk characteristics of private equity.
NBER Working Paper 9454.
Ljungqvist, A., Wilhelm, W., 2003. IPO pricing in the dot-com bubble. Journal of Finance 58-2, 577608.
Long, A., 1999. Inferring period variability of private market returns as measured by s from the range of
value (wealth) outcomes over time. Journal of Private Equity 5, 6396.
MacIntosh, J., 1997. Venture capital exits in Canada and the United States. In: Halpern, P. (Ed.),
Financing Growth in Canada. University of Calgary Press, Calgary.
Moskowitz, T., Vissing-Jorgenson, A., 2002. The returns to entrepreneurial investment: a private equity
premium puzzle? American Economic Review 92, 745778.
Peng, L., 2001. Building a venture capital index. Unpublished working paper, University of Cincinnati.
Reyes, J., 1990. Industry struggling to forge tools for measuring risk. Venture Capital Journal, Venture
Economics, Investment Benchmarks: Venture Capital.
Smith, J., Smith, R., 2000. Entrepreneurial Finance. Wiley, New York.
Venture Economics, 2000. Press release, May 1, 2000 at www.ventureeconomics.com.

The Risk and Return of Venture Capital: Article in Press

Uploaded by

Copyright:

Available Formats

The Risk and Return of Venture Capital: Article in Press

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

The Risk and Return of Venture Capital: Article in Press

Uploaded by

Copyright:

Available Formats

ARTICLE IN PRESS

Journal of Financial Economics 75 (2005) 352

The risk and return of venture capital$

0304-405X/$ - see front matter Published by Elsevier B.V.

J.H. Cochrane / Journal of Financial Economics 75 (2005) 352

J.H. Cochrane / Journal of Financial Economics 75 (2005) 352

J.H. Cochrane / Journal of Financial Economics 75 (2005) 352

Discount rates applied by VC investors might be informative, but the contrast

3. Overcoming selection bias

Return = Value at year 1

Time zero value = $1

Value at beginning of time 1

Pr(new round at time 1)

Pr(still private at end of time 1)

Value at beginning of time 2

Fig. 2. Procedure for calculating the likelihood function.

PrV t1 jV t is given by the lognormal distribution of (1). As before, I nd the

J.H. Cochrane / Journal of Financial Economics 75 (2005) 352

I assume here that seeing good data is independent of value.

J.H. Cochrane / Journal of Financial Economics 75 (2005) 352

period. (I report average returns, alphas and standard deviations as annualized

cross-correlation between individual venture capital returns, since I do not specify a

We want to nd the model in levels implied by Eq. (1), i.e.

m ln Rf s2 =2s2 1s2 =2

a elnR fegdEln R ln R d sm =2s =2 1 bemm ln R sm =2 1g;

Round to round (Table 4)

arithmetic return ER in Table 3 is a whopping 59%, with a 107% standard

J.H. Cochrane / Journal of Financial Economics 75 (2005) 352

J.H. Cochrane / Journal of Financial Economics 75 (2005) 352

is perfect cross-correlation between rounds, in which case additional rounds give no

6. Facts: fates and returns

Years since investment

J.H. Cochrane / Journal of Financial Economics 75 (2005) 352

612 month 12 year 23 year 34 year 45 year 5 year1

(a) Log returns, percent (not annualized)

(b) Arithmetic returns, percent

(c) Annualized arithmetic returns, percent

Average log returns, percent

(d) Annualized log returns, percent

Scale for Pr(IPO,acq.|V)

round variances are about half the IPO/acquisition

Scale for Pr(new fin.|V)

selection correction. The sample selection correction has a dramatic effect on

arms-length transactions, however. Ebay is a famous story (though not in the

J.H. Cochrane / Journal of Financial Economics 75 (2005) 352

7. How facts drive the estimates

the rst year,

F0:12 45% fail!

J.H. Cochrane / Journal of Financial Economics 75 (2005) 352

Round only, arithmetic

Note: Market model regressions are Rt!tk a bRm

32% arithmetic alpha in the selection-bias-corrected Table 3 pales by comparison.

S&P 500 return

particularly well estimated or stable across subsamples or variations in technique,

large volatility of log returns, and em2s : We have seen in a back-of-the-envelope

Note: See note to Table 8.

=0, others unchanged

J.H. Cochrane / Journal of Financial Economics 75 (2005) 352

a b R MRFt s SMBt h HMLt et