Waiting for Returns:

Using Space-Time Duality to

Calibrate Financial Diffusions

Mark Kamstra and Moshe A. Milevsky1

Schulich School of Business
York University, North York
Ontario, M3J 1P3, Canada

November 14, 2004

1 Kamstra and Milevsky are Associate Professors of Finance at the Schulich School of Business,
York University, Toronto. The contact author (Milevsky) can be reached at Tel: (416) 736-2100
ext: 66014, Fax: (416) 225-5034, Email: milevsky@yorku.ca. The authors would like to thank
Eliezer Prisman, Tom Salisbury, Georges Monette, Helen Massam, Tim Krehbiel, Peter Carr and
participants at the York University S.S.B. seminar series in Finance and the participants at the
Eastern Finance Association’s annual meeting for helpful comments and discussions.
 Abstract

In this brief paper we propose an alternative methodology for testing and calibrating diffusion
processes for financial time-serries. The methodology focuses on the duality that exists
between time and space for any given stochastic process. More specifically, we use the First
Passage Time (FPT) which is the amount of time required by a stochastic process to travel
a pre-specified distance. Thus, for example, we demonstrate that testing the hypothesis that
(logarithmic) investment returns are independent and normally distributed is equivalent to
testing the hypothesis that the FPT is Inverse Gaussian distributed. We apply this idea to
calibrate geometric Brownian motion (GBM) parameters for the S&P 500 index over the
period 1952 - 2004.
1 Introduction:
Historically, tests of the Geometric Brownian Motion (GBM) model for security prices – and
for that, matter any diffusion process – have been performed by selecting a fixed interval of
time (one day, one week, one month) ∆t, and then using the increments in logarithmic price
∆ ln[P ] over the predetermined ∆t. Under the classical specification of the GBM model, the
logarithmic price increments ∆ ln[P ] should be statistically independent from each other and
these increments ∆ ln[P ] should be Normally distributed with a mean and variance that is
proportional to the time increment ∆t. This approach has a long tradition in finance, and
research done in the early 1950’s by Kendall (1953) and Osborne (1959) as well as the work
by Fama (1970) all the way through the contemporary work of Lo and MacKinlay (1988)
focuses on a particular time interval ∆t.
Thus, for example, Kendall (1953) looked at a time increment of ∆t = one week on the
New York Stock Exchange, and concluded that the logarithmic price increments ∆ ln[P ] have
a statistically insignificant serial correlation in addition to being (approximately) normally
distributed. In another study, Fama (1970) looked at the 30 Dow Jones Industrial stocks
with a ∆t = one day, and concluded that there is a statistically significant positive serial
correlation in logarithmic price increments ∆ ln[P ]. Poterba and Summers (1988) found
that for a ∆t = three years, the logarithmic price increments ∆ ln[P ] exhibit statistically
significant negative serial correlation which translates into a long term mean reversion in
prices. Among the wealth of more recent studies that document violations of the GBM by
looking at the time series properties of returns to various financial instruments are Bakshi
and Chen (2000), Bollerslev, Chou, Kroner (1992), Cont (2001), Cont and da Fonseca (2002),
and Nelson (1991).
Nevertheless, the broad unifying methodology of this large literature is to select a time
interval and then investigate price increments vis a vis that time interval. Hence, it is quite
common to hear that the GBM model is rejected for hourly data while it is accepted for
monthly data but rejected again for yearly data (or some combination thereof).
In this paper we propose an alternative methodology. We investigate the GBM model
for fixed ∆ ln[P ] intervals as opposed to fixed ∆t intervals. In other words, we start at the
beginning of a time series and judiciously select a price increment ∆ ln[P ] = d (for example,
1 percent) and then measure the amount of time τ1 it takes the security to move the pre-
specified quantity. After the security has moved by ∆ ln[P ] = d we measure the time τ2 at
which the security moves an additional ∆ ln[P ] = d and so on an so forth. The final result is
a collection of time increments (τi+1 − τi ) for each pre- specified ∆ ln[P ]. We then compare
(statistically) the empirical distribution of the (τi+1 − τi )’s to the theoretical distribution
they should obey under the GBM model.

1
 If, indeed, the price increments are normal, then the (τi+1 − τi )’s – for each particular
∆ ln[P ] – should obey the Inverse Gaussian (or Wald) distribution as a result of the Space-
Time duality that exists for Brownian motion.
We select an entire spectrum of ∆ ln[P ]’s (for example from one percent all the way to
fifteen percent) and then extract the appropriate sample of τi ’s (for each ∆ ln[P ]) so as to
measure goodness of fit and estimate confidence intervals for the implied drift and diffusion
coefficients.
This is the first study of its kind – to our knowledge – which attempts to verify a
particular parameteric form and estimate parameters via this duality methodology. This
study will also shed light on the persistence of trend as a function of price momentum as
well as the velocity of the price process. If mean reversion behavior exists in the S&P 500,
larger price increments and their respective collection of τi ’s, will exhibit smaller implied
drifts and diffusion coefficients as well as a “poorer fit” to the Inverse Gaussian distribution.
We find the reverse, with larger drift and diffusion coefficients and a better fit to the Inverse
Gaussian distribution with larger increments.
Our approach should not be confused with, and is very different from, the paradigm of
spectral analysis. Spectral analysis attempts to uncover cycles in the underlying process by
fitting sine and cosine functions to the data. (See for example Granger and Morgenstern
(1963).)
In addition, investigating the data via the “dual” might have implications to option
pricing. Pricing an option that is deeply “in the money” or deeply “out of the money”
via the Black-Scholes formalism is in essence assuming that the increment of the strike
price divided by the current underlying security price is lognormally distributed. Hence
the appropriateness of the Black-Scholes model would depend on whether the increment
inherent in the option is properly modeled as a GBM. Our methodology provides a testable
hypothesis to verify this assumption and thereby explains systematic biases in the Black-
Scholes formula. Whether or not this is the case remains to be seen, nevertheless, this
particular problem can be addressed via such a study.
The rest of this paper is organized as follows. In order for the paper to be self-contained
Section 2 summarizes the theoretical properties of the Inverse Gaussian distribution and
demonstrates its relationship to the First Passage Time of a Brownian Motion. Section 3
discusses the issue of parameter and confidence interval estimation for the Inverse Gaussian
distribution vis-a-vis the First Passage Time distribution. The empirical results are tabulated
in Section 4 for the S&P 500. We will estimate the implied drift and diffusion coefficients by
implementing the algorithm developed in the previous sections. Conclusions and directions
for further research will be offered in Section 5.

2
2 The First Passage Time Distribution:
Let Pt denote the price of a security or index. It is standard in finance to assume that the
dynamics of Pt obey the following stochastic differential equation:

d( ln[Pt ] ) = µ dt + σdBt . (1)

Note that some authors use µ − 0.5σ 2 as the ”drift” term in the above equation although
to keep ”cleaner” notation we will assume that µ is in fact the geometric mean return, in
which case E[P1 ] = P0 exp{µ + 0.5σ 2 }. We will let Xt = ln[Pt ], hence:

dXt = µ dt + σdBt (2)

The logarithm of security (or index) prices obeys a Non-Standard Brownian Motion with
Drift. Let us now start at some point in time denoted by zero, such that X0 = x0 . Further-
more, choose an increment denoted by d. Let:

τ1 = inf{s; Xs ≥ x0 + d} (3)

Likewise, let:
τ2 = inf{s; Xs ≥ Xτ1 + d} (4)

further:
τ3 = inf{s; Xs ≥ Xτ2 + d} (5)

Finally:
τi = inf{s; Xs ≥ Xτi−1 + d} (6)

Thus, (τi+1 − τi ) is the sequence of First Passage Times of the stochastic process Xt to
the barriers demarcated by increments of d. It corresponds to the random amount of time
it takes the stochastic process Pt to move by ed − 1 = D percent. In can be shown, see
Seshadri (1993) or Wasan (1969), that the probability density function of the time increments
is Inverse Gaussian distributed. The probability density function (p.d.f.) of the Inverse
Gaussian (IG) random variable is a two parameter (β, λ) function that can be expressed as
follows: √

λ λ (t − β)2
g(t | β, λ) = √ exp − t > 0 (7)
2 π t3 2 β2 t
The IG p.d.f is defined for β > 0 and λ > 0. The mean (expected value) of the Inverse Gaus-
β3
sian random variable is β, while the variance is λ
. The cumulative distribution function,
(c.d.f.) which we denote by G(T | β, λ), of the Inverse Gaussian random variable can not
be expressed in closed form, however, it can be expressed as a function of the c.d.f. of the

3
standard normal random variable Φ[x] in the following elegant way: (See Chhikara (1989)
for details.)
 
  

T
λ T 2λ λ T
G(T | β, λ) = g(t | β, λ) dt = Φ −1 + exp( ) · Φ − 1+
0 T β β T β
(8)
d d2
For the First Passage Time, the parameters will be β = µ
and λ = σ2
. Thus, the expected
d
amount of time it will take the stochastic process Pt to move D percent is: µ
, the variance
dσ 2
in the amount of time will be: µ3
.

3 Parameter Estimation:
The Maximum Likelihood Estimate for the value of β is:
1
n
β̂ = τi (9)
n i=1
It is also an unbiased estimate for the value of β. The UMVUE for λ is:
n−1
λ̂ =  (10)
n 1 1
i=1 τi
− β̂

See Wasan (1969) for a derivation of the confidence intervals for β, λ.

A (1 − α) percent confidence interval for the value of λ is:

λ̂ λ̂
· χ2α/2 ≤λ≤ · χ21−α/2 (11)
n−1 n−1

Where χ2α/2 denotes the value from the chi square distribution with n − 1 degrees of freedom.
d2
Since λ = σ2
, we can obtain a (1 − α) percent confidence interval for the value of σ:
⎛ ⎞
⎝ d d ⎠
≤σ≤  (12)
2 2
λ̂
n−1
· χ1−α/2 λ̂
n−1
· χα/2

Likewise, a (1 − α) percent confidence interval for β is: (Where t1−α/2 denotes the value from
the student t distribution, with n degrees of freedom.)
⎛ ⎡  ⎤−1 ⎡  ⎤−1 ⎞
⎝ β̂ ⎣1 + β̂ β̂
· t1−α/2 ⎦ ≤ β ≤ β̂ ⎣1 − · t1−α/2 ⎦ ⎠ (13)
nλ̂ nλ̂

provided that nβ̂λ̂ · t1−α/2 < 1. Otherwise, the confidence interval is:
⎛ ⎡  ⎤−1 ⎞
⎝ β̂ ⎣1 + β̂
· t1−α/2 ⎦ ≤β≤ ∞⎠ (14)
nλ̂

4
Now, since β = µd , by inverting the confidence interval for β we can obtain a C.I. for µ.
⎛ ⎡  ⎤ ⎡  ⎤ ⎞
⎝ d ⎣1 − β̂ β̂
· t1−α/2 ⎦ β̂ −1 ≤ µ ≤ d ⎣1 + · t1−α/2 ⎦ β̂ −1 ⎠ (15)
nλ̂ nλ̂

Provided that β̂
nλ̂
· t1−α/2 < 1. Otherwise, the confidence interval for µ is:
⎛ ⎡  ⎤ ⎞
⎝ 0 ≤ µ ≤ d ⎣1 + β̂
· t1−α/2 ⎦ β̂ −1 ⎠ (16)
nλ̂

In general small data sets tend to result in one sided confidence intervals.

4 Empirical Results:
Using the principles set out in the previous section we can now derive point estimates and
confidence intervals for the values of µ, σ – the expected growth rate and volatility of returns
– as implied from β, λ from the first passage time data. We used the daily closing prices
on the S&P 500 cash index, for a period of time spanning January 1952 to December 2003,
resulting in 13109 data points for the stochastic process Pt . We then computed the amount of
time it takes the S&P 500 to move a pre-specified percentage D. Under the Null Hypothesis
that Pt obeys a geometric Brownian motion, the collection of these time increments should
obey an Inverse Gaussian distribution.
Figure 1 is a graphical representation of the confidence interval for µ as a function of
the percent increment. The line indicated with circles is the point estimate of the expected
return, while the solid dotted lines indicate the plus or minus two standard deviation confi-
dence about the mean. As one can see, larger increments in space imply a larger range for
the µ of the diffusion process and somewhat larger point estimates. Figure 2 is a graphical
representation of the confidence interval for σ as a function of the percent increment, again
with the line of circles representing the point estimate, now of the volatility, and the solid
dotted lines the confidence interval about that point estimate. In this case we obtain a more
dramatic result with larger increments in space implying a much larger value for the σ.
Recall that in theory – under the constant parameter GBM assumption – both graphs
should be flat to within statistical variations and the size of the data set. It is important to
note, however, the ”kink” in this graph may be due in part to sample-size truncation issues
involved with using daily returns. After all, if we are searching for 1% moves and are only
looking at daily numbers there is a (strong) chance that the S&P 500 moved up by more than
1% during the course of the day, and then reversed itself to close at a less-than-1% change.
The cumulative effect of this truncation is that we (erroneously) conclude the market did

5
not increase be 1% – when it did – and thus the underlying drift is not as high. This has
far reaching implications beyond just intra-day moves. For example, the market might take
3.2 trading days to increase by 1%, but in our data set it will be recorded as (much longer)
4 trading days which creates an artificial downward bias on the implied µ and σ. Of course,
as we increase the size of D, the extent to which this occurs is much less, since it is highly
unlikely that we ”missed” a 10% move in the S&P 500 because we only examined daily
closing prices.
Table 1 displays the point estimates for the parameters β, λ, µ, σ and associated standard
estimates, together with the number of data points that were observed. Thus, for example,
there were only 27 movements of 15% between 1952 and 2004. This small number may
limit the inferences we can draw from this data set from 15% moves. Table 2 displays the
95% confidence intervals for the above mentioned parameter values. Recall that under the
constant parameter GBM assumption the estimated values for β, λ should only depend on
the (logarithmic) space increment d, via the relationship β = d/µ and λ = d2 /σ 2 . Thus,
if the µ, σ for the return generating process are truly constant, then, for example, the β
value estimated for d = 2% increments should be twice the β value estimated at d = 1%
increments. As Table #1 and Table #2 indicate this is not the case and the parameter
estimates are not scaling by d and d2 . Once again, this is an indication that the underlying
generating process is likely not GBM with constant parameters, though this result may also
be due in part to sample-size truncation issues involved with using daily returns, as discussed
above.
Table 3 displays the results from performing a Kolmogorov-Smirinov (KS) test for goodness-
of-fit of the crossing time intervals to an Inverse Gaussian distribution. It is interesting to
note that within any given increment d above the 1% case, the data does not fail a KS test
for goodness-of-fit to an Inverse Gaussian distribution. And, while some of this might be due
to the low power of the KS test, a casual examination of the data, the plots of the CDF of
the data versus the Inverse Gaussian distribution – Figures #3 and above – reveals a good
match between the empirical and theoretical distribution, in particular where the data is
most dense, up to the 70th percentile or so of the cumulative.

5 Extension to Non-Lognormal Returns

As we mentioned earlier, the First Passage Time (FPT) distribution of the logarithmic
prices Xt to a level D will satisfy an Inverse Gaussian (I.G.) distribution if-and-only-if
the logarithmic prices themselves are Normally distributed. Indeed, when the process Xt
is something other than a non-standard Brownian motion – i.e. when eXt is no longer a
geometric Brownian motion – the collection of time increments τi will not be I.G. And,

6
although it is beyond the scope of this paper to derive and present FPT distributions for all
possible parameterization of Xt , in this section we briefly describe how one could go about
deriving a related probability for general process and thus use the space-time duality method
for investigating more general diffusions.
In order to adhere to common notation and terminology in the continuous-time finance
literature, assume the price process itself obeys the following one-dimensional diffusion:

dYt = ν(Yt , t)Yt dt + ξ(Yt , t)Yt dBt , Y0 = y. (17)

This representation covers our earlier geometric Brownian motion – when ν(Yt , t) = µ+0.5σ 2
and ξ(Yt , t) = σ are constants – as well as more general mean reverting and time-dependent
cases. In this case, the probability H(y, t) that Yt ”hits” or ”breaches” a level denoted by
D during a time period denotes by s, satisfies a so-called Kolmogorov partial differential
equation (PDE), denoted by:

∂H(y, t) ∂H(y, t) 1 2 ∂H 2 (y, t)

+ ν(y, t)y + ξ (y, t)y 2 = 0, (18)
∂t ∂y 2 ∂y 2

with a terminal condition H(D, s) = 1 if D > y and zero otherwise as well as a boundary
condition H(D, t) = 1 if y ≤ D. Thus, for example, under a particular parameterization
of equation (??), we can solve for the probability of observing a D =1% move within a
s =1-day period. We can then compare the theoretical probability dictated by equation (??)
against the observed frequency of 1% moves in 1-day. And although this is not exactly the
FPT density, we can employ standard goodness-of-fit methods to test whether in fact the
original (dual) diffusion Yt satisfies the postulated process in question.
In some cases equation (??) can be solved analytically, as we implicitly did earlier in the
paper. Of course, under the most general cases for ν(y, t) and ξ(y, t) one must resort to
numerical methods. Nevertheless, it is possible to obtain the hitting/crossing probabilities
for processes other than simple Brownian motions which opens the door for an alternative
method of calibrating and testing the return generating process for investment returns.

6 Conclusion:
We have proposed an alternative method for calibrating financial diffusions. We choose a
specific increment in price space, say a 1% return barrier, and examine the amount of time
it takes the stochastic process to move the predetermined increment. This is instead of
focusing on a particular increment in time – such as an hour, day or month – as do most
conventional estimation procedures. This methodology benefits from its ability to capture
changes in distribution that depend on the price (space) increment in question. We also

7
believe this approach better fits the perspective and needs of investors who are interested in
how long they will have to wait in order to achieve pre-specified target returns. Our empirical
results re-enforce previous results obtained in the literature that the stochastic price process
for S&P 500 equity returns does not conform to the standard geometric Brownian motion
(GBM) model as evidenced by the fact that our implied growth and volatility rates are not
constant. Interestingly, we do find a reasonably good fit of the GBM to first passage time
data for any given fixed barrier but these parameters are unstable across different return
barriers. In other words, if we only had access to historical data for how long it took the
S&P 500 to grow x% – as opposed to the daily or monthly returns – we could not reject
the Null Hypothesis that equity returns are lognormally distributed. It is only when we
compare the implied parameters across price increments that the GBM model fails. And,
although some of this instability may come from the coarseness of our data, measured daily,
it is unlikely to be solely due to this truncation time issue since this effect persists at larger
increments as well.
This study also sheds light on mean reversion in returns. If mean reversion behavior
exists in the S&P 500, larger price increments and their respective collection of first passage
times should exhibit smaller implied drifts and diffusion coefficients as well as a “poorer fit”
to the Inverse Gaussian distribution. We find the reverse, with larger drift and diffusion
coefficients and a better fit to the Inverse Gaussian distribution with larger increments.
Further research entails calibrating and testing first passage times for alternative prices
processes – such as currencies, commodities and interest rates – at higher frequency and in
particular on individual stocks. The same principle of space time duality can be used to
derive the distribution of first passage times for other stochastic processes and the authors
are currently working on classifying the FPT distribution for processes such as stochastic
volatility and mean reverting diffusions. Indeed, even if returns are generated by an infinite
variance stable distributions – as originally argued by Mandelbrot (1963) – then the finite
variance first passage times could be analyzed instead of the actual returns.

8
References
[1] G. Bakshi, C. Cao and Z. Chen (2000); “Do call prices and the underlying stock always
move in the same direction?” Review of Financial Studies Vo. 13 pp.549-84.

[2] T. Bollerslev, R.Y. Chou, and K.F. Kroner (1992); “ARCH Modeling in Finance: A
Review of the Theory and Empirical Evidence,” Journal of Econometrics Vol. 52, pp.
5-59.

[3] J.Y. Campbell, A.W. Lo and A. C. MacKinlay (1997); T he Econometrics of Financial

Markets, Princeton University Press, Princeton NJ.

[4] R.S. Chhikara and J.L. Folks (1989); The Inverse Gaussian Distribution, Theory,
Methodology and Applications, Marcel Dekker Inc., New York

[5] R. Cont and J. da Fonseca (2002); “Dynamics of implied volatility surfaces,” Quantita-
tive Finance, Vol 2(2), pp.45-60.

[6] R. Cont (2001); “Empirical properties of asset returns: stylized facts and statistical
issues” Quantitative Finance, Vol 1(2), pp. 223-236.

[7] E.F. Fama (1970); “Efficient Capital Markets: A Review of Theory and Empirical
Work”, The Journal of Finance, Vol. 25(2), pp. 383-417.

[8] C.W.J. Granger and O. Morgenstern (1963); “Spectral Analysis of New York Stock
Market Prices”, Kyklos, Vol. 16, pp. 1-27.

[9] M.G. Kendall (1953); “The Analysis of Economic Time-Series, Part I: Prices”, Journal
of the Royal Statistical Society, Vol. 96, pp. 11-25.

[10] A.W. Lo and A.C. MacKinlay (1988); “Stock Market Prices Do Not Follow Random
Walks: Evidence from a Simple Specification Test”, The Review of Financial Studies,
Vol. 1(1), pp. 41-46.

[11] B.B. Mandelbrot (1963); “The Variation of Certain Speculative Prices”, The Journal
of Business, Vol. 36, pp. 394-419.

[12] D.B. Nelson (1991); “Conditional Heteroskedasticity in Asset Returns: A New Ap-
proach,” Econometrica Vol. 59, pp. 347-370.

[13] P.R. Perry (1983); “More Evidence On The Nature Of The Distribution Of Security
Returns”, Journal of Financial and Quantitative Analysis, Vol. 18(2), pp. 211- 221.

9
[14] J.M. Poterba and L.H. Summers (1988); “Mean Reversion In Stock Market Prices:
Evidence and Implications”, The Journal of Financial Economics, Vol. 22(1), pp. 27-
60.

[15] M.F.M. Osborne (1959); “Brownian Motion in the Stock Market”, Operations Research,
Vol. 7, pp. 145-173.

[16] V. Seshadri (1993); The Inverse Gaussian Distribution: A Case Study in Exponential
Families, Clarendon Press, Oxford.

[17] M.T. Wasan (1969); “First Passage Time Distribution of Brownian Motion with Positive
Drift”, Queen’s Papers in Pure and Applied Mathematics, No. 19, Queen’s University,
Kingston, Ontario.

10
 Table #1
S&P 500 1952/01/01-2003/12/31
Annualized Percentage Returns

Barrier n µ (std) σ (std) β (std) λ (std)

1% 299 6.2 (0.96) 6.69 (0.27) 0.161 (0.025) 0.022 (0.002)
2% 175 7.26 (1.2) 8.35 (0.45) 0.276 (0.046) 0.057 (0.006)
3% 122 7.58 (1.24) 8.61 (0.55) 0.396 (0.065) 0.121 (0.016)
4% 95 7.87 (1.23) 8.55 (0.62) 0.508 (0.079) 0.219 (0.032)
5% 77 8.02 (1.45) 10.03 (0.81) 0.624 (0.113) 0.249 (0.04)
6% 64 8 (1.41) 9.76 (0.86) 0.751 (0.132) 0.378 (0.067)
7% 56 8.16 (1.51) 10.48 (0.99) 0.858 (0.159) 0.447 (0.084)
8% 49 8.16 (1.53) 10.63 (1.07) 0.98 (0.185) 0.565 (0.114)
9% 44 8.26 (1.44) 9.98 (1.06) 1.089 (0.19) 0.813 (0.173)
10 % 40 8.29 (1.56) 10.83 (1.21) 1.206 (0.227) 0.853 (0.191)
11 % 36 8.21 (1.52) 10.58 (1.25) 1.342 (0.25) 1.08 (0.255)
12 % 33 8.37 (1.74) 11.98 (1.47) 1.433 (0.298) 1.002 (0.247)
13 % 31 8.39 (1.63) 11.26 (1.43) 1.55 (0.301) 1.332 (0.338)
14 % 28 8.27 (1.6) 11 (1.47) 1.694 (0.328) 1.621 (0.433)
15 % 27 8.39 (1.71) 11.9 (1.62) 1.787 (0.365) 1.59 (0.433)

Using daily returns from S&P 500 for the period 1950 to 2004, the table displays point
estimates for the parameters β, λ, µ, σ. Note that the β, λ parameters are estimated directly
from the data while the µ, σ are ”solved” by the analytic relationship between the Inverse
Guasian and Normal distribution.

11
Table #2
S&P 500 1952/01/01-2003/12/31
95% Confidence Interval

Barrier n µ σ β λ
1 299 (4.32 8.08) (6.16 7.22) (0.11 0.21) (0.02 0.03)
2 175 (4.91 9.61) (7.47 9.23) (0.19 0.37) (0.05 0.07)
3 122 (5.15 10.01) (7.53 9.69) (0.27 0.52) (0.09 0.15)
4 95 (5.46 10.28) (7.33 9.77) (0.35 0.66) (0.16 0.28)
5 77 (5.18 10.86) (8.44 11.62) (0.4 0.85) (0.17 0.33)
6 64 (5.24 10.76) (8.07 11.45) (0.49 1.01) (0.25 0.51)
7 56 (5.2 11.12) (8.54 12.42) (0.55 1.17) (0.28 0.61)
8 49 (5.16 11.16) (8.53 12.73) (0.62 1.34) (0.34 0.79)
9 44 (5.44 11.08) (7.9 12.06) (0.72 1.46) (0.47 1.15)
10 40 (5.23 11.35) (8.46 13.2) (0.76 1.65) (0.48 1.23)
11 36 (5.23 11.19) (8.13 13.03) (0.85 1.83) (0.58 1.58)
12 33 (4.96 11.78) (9.1 14.86) (0.85 2.02) (0.52 1.49)
13 31 (5.2 11.58) (8.46 14.06) (0.96 2.14) (0.67 1.99)
14 28 (5.13 11.41) (8.12 13.88) (1.05 2.34) (0.77 2.47)
15 27 (5.04 11.74) (8.72 15.08) (1.07 2.5) (0.74 2.44)

The table displays the 95% Confidence Interval for the estimated parameters β, λ, µ, σ.

12
 Table #3
The table displays results from a Kolmogorov Smirinov (K.S.) goodness-of-fit test of
the data – for each level of D – against an Inverse Gaussian (I.G.) distribution. The Null
Hypothesis for our K.S. test is that the data was generated from an I.G. distribution with β, λ
parameters specified in Table #2. The Null Hypothesis is rejected if the test statistic is ”too
large” which means that the distance between the empirical CDF and candidate CDF are
”too far” from each other. At the 5% significance level the critical value of the K.S. statistic
is approximately 1.358 and at the 10% significance the critical value is approximately 1.223.
S&P 500 1952/01/01-2003/12/31

Barrier Data Points K.S. Value R / NR at 10% Sig.

1% 299 1.298 Reject
2% 175 0.783 Do Not Reject
3% 122 0.684 Do Not Reject
4% 95 0.784 Do Not Reject
5% 77 0.611 Do Not Reject
6% 64 0.874 Do Not Reject
7% 56 0.609 Do Not Reject
8% 49 0.511 Do Not Reject
9% 44 0.633 Do Not Reject
10% 40 0.549 Do Not Reject
11% 36 0.629 Do Not Reject
12% 33 0.671 Do Not Reject
13% 31 0.576 Do Not Reject
14% 28 0.675 Do Not Reject
15% 27 0.570 Do Not Reject
Figure 1: S&P 500 1952-2004 Return Mean Estimate, with Confidence Interval
Figure 2: S&P 500 1952-2004 Return Volatility Estimate, with Confidence Interval
Figure 3: Barrier=1%
Figure 4: Barrier=2%
Figure 5: Barrier=3%
Figure 6: Barrier=4%
Figure 7: Barrier=5%
Figure 8: Barrier=10%
Figure 9: Barrier=15%