Modeling VXX

Sebastian A. Gehricke
Department of Accountancy and Finance
Otago Business School, University of Otago
Dunedin 9054, New Zealand
Email: sebastian.gehricke@postgrad.otago.ac.nz

Jin E. Zhang
Department of Accountancy and Finance
Otago Business School, University of Otago
Dunedin 9054, New Zealand
Email: jin.zhang@otago.ac.nz

First Version: June 2014

This Version: 13 September 2014

Keywords: VXX; VIX Futures; Roll Yield; Market Price of Variance Risk; Variance Risk
Premium
JEL Classification Code: G13
 Modeling VXX

Abstract

We study the VXX Exchange Traded Note (ETN), that has been actively traded in
the New York Stock Exchange in recent years. We propose a simple model for the VXX
and derive an analytical expression for the VXX roll yield. The roll yield of any futures

position is the return not due to movements of the underlying, in commodity futures it is
often called the cost of carry. Using our model we confirm that the phenomena of the large
negative returns of the VXX, as first documented by Whaley (2013), which we call the

VXX return puzzle, is due to the predominantly negative roll yield as proposed but never
quantified in the literature. We provide a simple and robust estimation of the market price
of variance risk which uses historical VXX returns. Our VXX price model can be used to

study the price of options written on the VXX.

Modeling VXX 1

1 Introduction

There are three major risk factors which are traded in financial markets: market risk which

is traded in the stock market, interest rate risk which is traded in the bond markets and
interest rate derivative markets, and volatility risk which up until recently was only traded
indirectly in the options market.

It is well accepted in the literature that both equity returns and variance are random
(French, Schwert, and Stambaugh, 1987). It is also well understood that the variance
risk premium is significant and negative (Carr and Wu, 2009). Investors trade volatility

either to take advantage of the opportunity in the variance risk premium or to hedge against
volatility risk. One way investors can trade volatility would be to buy at-the-money (ATM)
options, but these do not necessarily stay at-the-money. When the options are Out-of-the-

money (OTM) and in-the-money (ITM) they have smaller volatility sensitivity (Vega) and
are therefore less effective for trading volatility. Options contracts will not always be able
to meet investors need for volatility risk management as there will not be enough liquidity

in the options markets when the market goes down. Also when investors trade volatility in
the options market they also trade market risk and possibly interest rate risk (for longer
term options), therefore trading volatility through options is often contaminated by these

other risk factors making it inefficient for risk management. Developing a financial market
to trade volatility directly is very important for researchers and practitioners. (Zhu and
Zhang, 2007)

Chen, Chung, and Ho (2011) show that VIX (if tradeable) and VIX options expand
investor opportunity set and are useful for diversification, in a mean-variance optimizing
markowitz framework.

In 2003, the methodology for calculating the VIX index changed and the index using
the old methodology was renamed to VXO. The VIX is now calculated using all out-of-the-
Modeling VXX 2

money options on the S&P 500 which have a bid price. Following this change in 2004, the
CBOE launched the much anticipated VIX futures and in 2006 VIX options also started

trading on the CBOE. Both VIX futures and options have consistently grown in daily dollar
trading volume.
In 2009, S&P Dow Jones Indices started reporting several different VIX futures indices

which represent the returns of different VIX futures positions. One example of a VIX
futures index is the S&P 500 VIX Short-Term Futures Index (SPVXSTR) which tracks
the performance of a position in the nearest and second nearest maturing VIX futures.

The SPVXSTR is rebalanced daily to create a constant one month maturity VIX futures
position. Shortly after the VIX futures indices started were developed Barclays Capital
iPath launched the first ever VIX futures index Exchange Traded Product (ETP), the VXX

Exchange Traded Note (ETN). An ETN is unsecured senior debt that pays no coupons
(interest) and does not have a fixed redemption at maturity but rather its redemption
value is linked to the performance of some underlying (Bao, Li, and Gong, 2012).The

VXX’s redemption value, for example, depends on the value of the SPVXSTR at maturity
less an annual management fee of 0.89%.
There are now many different VIX futures ETNs with different underlying indices, all

of these combined make the VIX futures ETN market. The VIX futures ETN market has
become vastly popular, all of the ETNs combined have a market capitalization of nearly
4 billion US dollars and average daily trading volume in excess of 800 million US dollars

(Whaley, 2013). One of the main drivers of VIX futures ETNs growing popularity may be
that Mutual funds and Hedge funds are often restricted from trading futures and options
but they still have a need to hedge volatility risk therefore they trade in the VIX futures

ETN market.
The VXX is the most popular of the VIX futures ETNs and is now the third most

traded ETP, amongst all ETPs, based on average daily trading volume. The VXX is only
Modeling VXX 3

just behind the iShares MSCI Emerging Markets ETF (EEM), much further behind the
S&P 500 ETF (SPY) and in front of the iShares Russell 2000 ETF in terms of daily trading

volume (IWM)1 .
Whaley (2013) is the first to document the phenomena of the highly negative returns
of the VXX, which we will refer to as the VXX return puzzle. Eraker and Wu (2013) also

show the significant negative performance of VIX futures and VIX futures index ETPs
(including the VXX). Deng, McCann, and Wang (2012) show that ETNs on VIX futures
indices, such as the VXX, are not very effective hedging/diversification tools for equity and

mixed equity and bond portfolios. Hancock (2013) tests the performance of VIX futures
ETNs and compares them to three benchmarks. Hancock (2013) shows that the VXX
and other VIX futures ETNs never consistently outperform benchmarks even when used

to diversify equity portfolios. These findings hold even when different holding periods and
portfolio weighting methods are used. Hancock (2013) suggests that the poor performance
is unique to VIX futures ETNs and is not a property of volatility.

We document the VXX returns in table 1 which shows the summary statistics of the
VXX, SPX (S&P 500 index ETP) and VIX returns from 30th January 2009 to the 27th June
2014. Note the abysmal performance of the VXX as can be seen firstly by the -0.32% average

daily discrete return of the VXX and the average daily continously compounded return of
-0.39% as opposed to the average daily continuously compouned and discrete returns of
the VIX which were -0.09% and 0.15% respectively. Secondly, the Holding Period Return

(HPR) shows that within our sample period the VXX has lost 99.59% of its value, the VIX
has only lost 71.66%. The Compound Annual Growth Rate (CAGR) of -63.49% of the
VXX compared to a CAGR of -20.41% of the VIX, further displays the underperformance

of the VXX. We will later show that the main reason why the VXX does not follow the
1
ETP database website: www.etfdb.com/compare/volume as of the 10th October 2014. The average
daily trading volume is computed as an average of the daily number of shares of that ETP traded over the
previous 3 months.
Modeling VXX 4

VIX, as the constant 30-day maturity VIX futures does, is due to the roll yield.In figure 2
we plot the VIX index, the VXX price and the constant 30-day-maturity VIX futures price,

as in Zhang, Shu, and Brenner (2010), so the difference is visually observable. Even with
the well documented and easily observed underperformance the VXX market has made
great strides in popularity, figure 3 shows us the upward trend in the daily dollar trading

volume and the initial increase in and then levelling off in market capitalization of the VXX
since inception.
Whaley (2013), Deng, McCann, and Wang (2012), Husson and McCann (2011) and Bao,

Li, and Gong (2012) all suggest that the VXX is subject to the roll yield of VIX futures and
that this is the cause for the underperformance of the VXX. None of the aforementioned
articles quantify the roll yield or attempt to measure it, therefore we will create a model for

the VXX which allows the quantification of the roll yield and proves the hypothesis that
the roll yield drives the significant negative returns of the VXX.
The roll yield of any futures position is the return that a futures investor captures when

the futures price converges to the spot price, it is the part of the return which is not due to
changes in the price of the underlying asset or index. When the market is in backwardation
(i.e. downward sloping term-structure) the price rolls up to the spot price, therefore the roll

yield will be positive. When the market is in contango (i.e. upward sloping term-structure)
the price rolls down to the spot price, therefore the roll yield will be negative. The VIX
futures term structure is in contango during normal times and therefore the roll yield is

for example negative. The VIX futures term structure can be in backwardation, usually
during large economic downturns, and the roll yield will become positive which can make
ETPs on VIX futures indices profitable (Whaley, 2013).

We study the VXX price by using the VIX futures price approximation from Zhang, Shu,
and Brenner (2010), which we review in section 3.1, to propose the first stochastic volatility

model of the VXX which accounts for the underlying dynamics of the S&P 500 index (SPX)
Modeling VXX 5

and the VIX index. We believe the relationship between the VXX, the VIX and the S&P
500 is essential in building a comprehensive model. We show that the difference between

the 30-day-maturity VIX futures price change and the VXX price change in figure 2 is in
fact due to the roll yield. We then go further and show that the roll yields sign is driven, on
aggregate, by the negative market price of variance risk, λ. Eraker and Wu (2013) use an

equilibrium model approach to show that the Variance Risk Premium (VRP) is the driver
of the VXX’s negative returns. This is consistent with our finding as the market price
of variance risk, λ, and the VRP are almost proportional as shown by Zhang and Huang

(2010).
In the next section we will explain the methodology for how the SPVXSTR index is
calculated. Then in Section 3 we will review the theory behind pricing the VIX and VIX

futures from Zhang and Zhu (2006) and Zhang, Shu, and Brenner (2010) and use this to
create a stochastic model for the VXX price and examine the roll yield of the VXX. In
section 4 we will use the VXX model to develop a simple way of estimating the market

price of variance risk. In section 5 we will examine the effect of the rebalancing frequency
of the SPVXSTR which will also be a robustness test of our continuous time VXX model.
Finally in section 6 we will conclude and discuss on our findings.

2 The SPVXSTR index

To model the VXX we must first understand the SPVXSTR. In this section we will present

the methodology for calculating the SPVXSTR index as interpreted from S&P Dow Jones
Indices (2012).
The SPVXSTR index seeks to model the outcome of holding a long position in short-

term VIX futures, specifically holding positions in the nearest and second nearest maturing
VIX futures. The position is rebalanced daily to create a constant rolling one-month
Modeling VXX 6

maturity VIX futures position (Barclays, 2013). The index is calculated by

SP V XST Rt = SP V XST Rt−1 (1 + CDRt + T BRt ), (1)

where SP V XST Rt is the index level at time t, SP V XST Rt−1 is the index level at time
t − 1, CDRt is the Contract Daily Return of the VIX futures position and T BRt is the

Treasury Bill Return earned on the notional value of the position. The T BRt is given by

  Delta t
1 91
T BRt = 91 , (2)
1 − 360 T BARt−1
where Deltat is the number of calendar days between the current and previous business
days. T BARt−1 is the Treasury Bill Annual Return, which is equal to the most recent
weekly high discount rate for 91-day US Treasury bills effective on the preceding business

day. Usually the rates are announced by the US Treasury on each Monday, but if the
Monday is a holiday then Fridays rates will apply. The CDRt is calculated by

w1,t−1 FtT1 + w2,t−1 FtT2

CDRt = T1 T2
, (3)
w1,t−1 Ft−1 + w2,t−1 Ft−1
where wi,t−1 is the weight in the ith nearest maturing VIX futures at time t − 1, FtTi is the
Ti
market price of the ith nearest maturing VIX futures contract at time t and Ft−1 is the

market price of the ith nearest maturing VIX futures contract at time t − 1.2 The weights
are adjusted daily to be

dr
w1,t = ,
dt
and
dr
w2,t = 1 − ,
dt
2
In Equation (3) we use we use w1,t−1 and w2,t−1 in the numerator. Deng, McCann, and Wang
(2012)CDRt use w1,t and w2,t which is inconsistent with the methodology from S&P Dow Jones Indices
(2012). When calculating discrete returns of any position the weights should stay constant over the period
you are calculating the return for and only the prices should change.
Modeling VXX 7

where S&P Dow Jones Indices (2012) defines “dr =The total number of business days
within a Roll Period beginning with, and including, the following business day and ending

with, but excluding, the following CBOE VIX Futures Settlement Date. The number
of business days includes a new holiday introduced intra-month up to the business day
preceding such a holiday.” and “dt =The total number of business days in the current Roll

Period beginning with, and including, the starting CBOE VIX Futures Settlement Date
and ending with, but excluding, the following CBOE VIX Futures Settlement Date. The
number of business days stays constant in cases of a new holiday introduced intra-month

or an unscheduled market closure” (S&P Dow Jones Indices, 2012, p. 7) Figure 1 shows
the determination of dr and dt in a diagram for convenience of understanding.

3 Modeling VXX
3.1 Review of VIX and VIX futures model

To model the VXX we need a model for the VIX index and VIX futures. Zhang and Zhu
(2006) and Zhang, Shu, and Brenner (2010) have developed a model for the VIX and VIX

futures, for completeness we review and combine the results from both in this section.
The SPX (S&P 500 index) can be modeled by the following diffusion process with a
stochastic process of instantaneous volatility as described by Heston (1993),

p
P
dSt = µSt dt + Vt St dB1,t , (4)
p
P
dVt = κ(θ − Vt )dt + σv Vt dB2,t , (5)

where St is the SPX, Vt is the instantaneous variance of the SPX, µ is the expected re-
turn from investing in the SPX, θ is the physical measure for the long run mean level of

the instantaneous variance, κ is the physical measure for the speed of mean reversion of
P P
instantaneous variance and σV measures the the variance of variance. B1,t and B2,t are
Modeling VXX 8

two standard Brownian motions that describe the random noise in the SPX return and
variance, respectively, they are correlated by a constant correlation coefficient ρ.

The transformations between physical and risk-neutral parameters are given by

θ∗ κ∗
θ= (6)
κ
and

κ∗ = κ + λ, (7)

where κ∗ is the risk-neutral speed of mean reversion of volatility, θ∗ is the risk-neutral long

run mean level of instantaneous variance and λ is the market price of variance risk. We
can then describe the risk-neutral dynamics of the SPX as follows

p
∗
dSt = rSt dt + Vt St dB1,t , (8)
p
∗
dVt = κ∗ (θ∗ − Vt )dt + σv Vt dB2,t , (9)

∗ ∗
where r is the risk free rate, and dB1,t and dB2,t are two new standard Brownian motions

which are correlated by the constant correlation coefficient, ρ. The VIX is equal to the
variance swap rate (Carr and Wu, 2009), which is equivalent to the conditional expectation
in the risk-neutral measure

 Z t+τ0 
1
V IXt2 = Et∗ Vs ds = (1 − B)θ∗ + BVt , (10)
τ0 t
∗τ
30 1−e−κ 0
where τ0 = 365
and B = κ∗ τ0
. Then the VIX futures price formula is given by

FtT p
= Et∗ (V IXT ) =Et∗ ( (1 − B)θ∗ + BVT )
100 Z +∞ p (11)
= (1 − B)θ∗ + BVT f ∗ (VT |Vt )dVT ,
0
Modeling VXX 9

where the transition probability density as given by Cox et al. (1985) is

 v q/2 √
f ∗ (VT |Vt ) = ce−u−v Iq (2 uv), (12)
u
where

2κ∗ −κ∗ (T −t) 2κ∗ θ∗

c= 2 , u = cVt e , v = cVT , q = − 1,
σV (1 − e−κ∗ (T −t) ) σV2
where Iq (.) is the modified Bessel function of the first kind and of order q. The distribution

function is the non-central chi-square, χ2 (2v; 2q + 2, 2u) with 2q + 2 degrees of freedom and
parameter of non-centrality 2u proportional to Vt . Note that (T − t) is the time to maturity
of the VIX futures contract. (Zhang and Zhu, 2006)

Equation (11) is the accurate formula for the VIX futures price from Zhang and Zhu
(2006) using our own notation. Zhang, Shu, and Brenner (2010) provide us with a very
3
good closed form approximation of equation (11) given by

FtT
= F0 + F1 + F2 , (13)
100
where

∗ (T −t) ∗ (T −t) 1
F0 = [θ∗ (1 − Be−κ ) + Vt Be−κ ] ,
2

σv2 ∗ ∗ ∗ −3
F1 = − [θ (1 − Be−κ (T −t) ) + Vt Be−κ (T −t) ] 2
8 
−κ∗ (T −t) −κ∗ (T −t) 2

2 −κ∗ (T −t) 1 − e ∗ (1 − e )
× B Vt e +θ ,
κ∗ 2κ∗
3
In Zhang, Shu, and Brenner (2010) θ is assumed to be time dependant, θt , but we stick with the
simpler version of the model from Zhang and Zhu (2006) and assume that θ is constant.
Modeling VXX 10

σv4 ∗ ∗ ∗ −5
F2 = [θ (1 − Be−κ (T −t) ) + Vt Be−κ (T −t) ] 2
16 
−κ∗ (T −t) 2 ∗
1 ∗ (1 − e−κ (T −t) )3

3 3 −κ∗ (T −t) (1 − e )
×B Vt e + θ ,
2 κ∗ 2 2 κ∗ 2
where F1 +F2 can be thought of as a convexity adjustment from the Taylor series expansion

of equation (11).

3.2 Nearly 30-day VIX futures

Table 2 presents the values of estimated VIX futures prices using the full formula from
Zhang and Zhu (2006), the closed form approximation of the full formula from Zhang, Shu,
and Brenner (2010), equation (13), and two simplifications of the closed form approxima-

tion, F0 + F1 and just F0 . From table 2 we can see that for 30-day VIX futures prices using
just F0 creates a very small error from the accurate formula, equation (11). The table
shows that the error from using just F0 instead of the accurate formula, , equation (11),

is always within 3% when θ∗ = 0.1 and Vt ranges from 0.04 to 0.2, κ∗ ranges from 4 to 7
and σv ranges from 0.1 to 0.7. There is one outlier when Vt = 0.04, κ∗ = 4 and σV = 0.7,
but the error is only just outside 3% at 3.20%. The Root Mean Squared Error (RMSE) is

1.29% which is very acceptable. The results of the numerical exercise presented in table 2
lead us to proposition 1 below.

Proposition 1 The price of nearly 30-day-to-maturity VIX futures can be given by

FtT ∗ ∗ 1
= [θ∗ (1 − Be−κ (T −t) ) + Vt Be−κ (T −t) ] 2 , (14)
100
with some small error when compared to the accurate VIX futures price formula from Zhang
and Zhu (2006), as demonstrated in Table 2. For example for the range of parameters

σV = 0.1 to 0.7, Vt = 0.04 to 0.20, κ∗ = 4 to 7, constant θ∗ = 0.1 and maturity of 30 days,

the RMSE is only 1.29%.
Modeling VXX 11

We take the natural log of equation (14) to get an expression for the natural log price
of VIX futures given by

FtT
 
1 ∗ ∗
ln = ln[θ∗ (1 − Be−κ (T −t) ) + Vt Be−κ (T −t) ], (15)
100 2
FT
where ln( 100
t
) is the natural log the price of nearly 30-day to maturity VIX futures contract.
In Figure 4 we can see the theoretical term structure of VIX futures using equation (14),
the full approximation of VIX futures prices, equation (13) and only the F0 + F1 segment

of the full approximation. We use parameter estimates of θ∗ = 0.1, κ∗ = 5, σv = 0.1425

and Vt = 0.06 to create an upward sloping VIX futures term structure, as is normal for the
VIX futures market. The difference between the points at t + 30 and t + 29 is equal to the

average one-day roll yield of a 30-day to maturity VIX futures contract. The spot return is
zero when the underlying instantaneous variance is constant which means that any return
that can be seen is due to the roll yield of VIX futures. It can be seen in the diagram

that as you step through time from t+30 to t+29 the return will be negative therefore the
one-day roll yield will be negative when the term structure is upward sloping.

3.3 Model of Contract Daily Return

We can model the change of nearly 30-day log VIX futures price by taking the Taylor series
expansion of our simple log VIX futures price formula, equation (15), this gives us

∂ ln FtT 1 ∂ 2 ln FtT ∂ ln FtT

d ln FtT = dVt + (dV t ) 2
+ dt, (16)
∂Vt 2 ∂Vt2 ∂t
next we substitute in the partial derivatives to get4
4
Seen the Appendix, section A.
Modeling VXX 12

−1
θ∗

1 ∗
d ln FtT = − θ + Vt dVt
2 Be−κ∗ (T −t)
−2
θ∗

1 ∗
− − θ + Vt (dVt )2 (17)
4 Be−κ∗ (T −t)
∗
1 κ∗ (Vt − θ∗ )Be−κ (T −t)
 
+ dt.
2 θ∗ + (Vt − θ∗ )Be−κ∗ (T −t)

Proposition 2 The SPVXSTR index is rebalanced daily to maintain a VIX futures posi-
tion with one month maturity, therefore we can model the contract daily return (CDRt ) of

the SPVXSTR as the log return of a 30-day to maturity VIX futures position. From this
and equation (17) we get

CDRt = d ln FtT = d ln Ftt+τ0 + RYt , (18)

T =t+τ0

where

−1
θ∗

1 ∗
d ln Ftt+τ0 = − θ + Vt dVt
2 Be−κ∗ τ0
−2 (19)
θ∗

1 ∗
− − θ + Vt (dVt )2
4 Be−κ∗ τ0

∗
1 κ∗ (Vt − θ∗ )Be−κ τ0
 
RYt = dt, (20)
2 θ∗ + (Vt − θ∗ )Be−κ∗ τ0
where τ0 = 30/365, d ln Ftt+τ0 is the change in the log price of a constant 30-day to maturity
VIX futures contract and RYt is the roll yield of the SPVXSTR. The roll yield of the
SPVXSTR is the return of the underlying VIX futures position due to the maturity of the

position changing from 30 days to 29 days, from one rebalancing of the position to just
before the next rebalancing.
Modeling VXX 13

3.4 VXX model

Proposition 3 We know that the change in the SPVXSTR index, and therefore the VXX,

is composed of the return of the futures position, the CDRt , and a risk-free return on the
notional of the futures position, T BRt . Therefore we can model the VXX using CDRt
combined with a risk free return r given by

d ln V XXt = CDRt + rdt = d ln FtT + rdt

T =t+τ0 (21)
= d ln Ftt+τ0 + RYt + rdt
where RYt is the one day roll yield of the VXX going from 30-day maturity to 29-day
maturity and r is the risk-free return on the notional value of the futures position.

This model of the log VXX price is to our knowledge the first attempt in the literature
to model the VXX using the underlying dynamics of the SPX. The model can be used to

derive the market price of variance risk, λ, from VXX returns, as is described in Section
4. We could also use this model to price VXX options, which are essentially Asian options
on the underlying instantaneous variance, Vt . In the next section we use our VXX model

to quantify the roll yield and show that it drives the VXX’s returns.

3.5 VXX roll yield

Whaley (2013), Deng, McCann, and Wang (2012) and Husson and McCann (2011) all
suggest the roll yield as the reason the VXX’s returns are so negative. Figure 2 shows
us a comparison between the performance of the VIX, the VXX and a constant 30-day to

maturity VIX futures contract. We can see an obvious difference between the 30-day to
maturity VIX futures contract and VXX. From equation (21) we know that the difference
between the 30-day VIX futures price and the VXX can be explained by the roll yield, RYt .

To examine what drives the roll yield of the VXX we assume that the instantaneous
variance, Vt , is constant at the physical measure long run mean level of instantaneous
Modeling VXX 14

variance, θ, to produce the aggregate upward sloping term structure of the VXX. If Vt is
constant then dVt = 0, and therefore equation (18) simplifies to

∗
1 κ∗ (θ − θ∗ )Be−κ τ0
 
CDR = RYt∗ = ∆t, (22)
2 θ∗ + (θ − θ∗ )Be−κ∗ τ0
RYt∗
where ∆t
is the one day roll yield of the aggregate VXX. To examine what drives the roll
yield to be negative, during normal times, we can use the transformation from the risk-

neutral measure to the physical measure long-run mean level of instantaneous variance,
equation (6) and substituting this into equation (22) we get

∗ ∗
RYt∗ 1 λκκ Be−κ τ0
= . (23)
∆t 2 1 + λκ Be−κ∗ τ0
As all parameters apart from λ are always positive and κ∗ = κ + λ > 0 (Zhang, Shu,
and Brenner, 2010), from equation (23) we can see that λ, the market price of variance

risk, is the driver of sign of the one day roll yield of the VXX, on aggregate. We conclude
that the negative roll yield of the VXX is driven by the usually negative, as shown in table
3, market price of variance risk.

4 The Market Price of Variance Risk, λ

When modelling instantaneous volatility which is not directly tradable it is important to

incorporate the market price of risk. Under the Cohen et al. (1972) model you can always

perfectly hedge your position and therefore you do not need the market price of variance for
your model. When you are modeling something that is not traded this situation changes
as you will not be able to create a perfect risk free portfolio and therefore an investor will

require a premium to compensate for the risk, this is the market price of variance risk.
When implementing a stochastic volatility model, such as in the Heston (1993) frame-
work, estimating the market price of variance risk, λ, is essential. There is no clear consen-
Modeling VXX 15

sus on the estimation of the market price of variance risk,λ. Table 3 shows some different
author’s recent estimates for the market price of variance risk λ, the risk neutral measure

of the mean reverting speed of variance, κ∗ , and the sample period used. We can see from
table 3 that the estimation of λ can vastly vary depending on the estimation methodology
used.

Our model lets us develop a simple method of estimating λ. If we substitute Vt = θ in

equation (21) we get

∗
1 κ∗ (θ − θ∗ )Be−κ τ0
 
d ln V XXt = dt + rdt, (24)
2 θ∗ + (θ − θ∗ )Be−κ∗ τ0
where dVt = 0, so we are isolating the aggregate effect of the roll yield.
We can now take the integral of equation (24) and substitute in the transformation from
risk-neutral to physical measure long run mean level of variance, from equation (6) to get

∗
λ̄Be−κ τ0
 
V XXT 1
R = ln = T + rT, (25)
V XX0 2 (1 + λ̄∗ Be−κ∗ τ0 )
κ

where R is the continuously compounded return on the VXX over the sample. V XXT is
the last VXX price and V XX0 is the starting VXX price, in the sample period, rT is the

risk free return and λ̄ is given by

λκ∗ λκ∗
λ̄ = = ∗ . (26)
κ κ −λ

Proposition 4 We can use the VXX return and a estimate of κ∗ to measure λ, the market

price of variance risk by solving equation (26) for λ, which gives us

λ̄κ∗
λ= , (27)
κ∗ + λ̄
and solving equation (25) for λ̄ we get
Modeling VXX 16

2RE
λ̄ = 2RE
.
(1 − κ∗
)Be−κ∗ τ0
where RE is the annualized excess return of the VXX over the sample period, given by

1 V XXT
RE = ln −r (28)
T V XX0

We use the parameter estimate of κ∗ = 5.4642 from Luo and Zhang (2012) ,as their
estimate of κ∗ is the most recent available one in the literature and the closest to our

sample period, to demonstrate our new methodology of calculating λ. We then use the
VXX prices from inception V XX0 = 6693.12 on 30 Jan 2009 and the VXX price at the
end of our sample V XXT = 28.86, on 27 Jun 2014.5 T = 5.4082 in years and rT is

the cumulative treasury bill return over the same time period, rT = T BR0,T = 0.558% as
defined in equation (2) from section 2 but cumulated over the entire sample. The cumulated
TBR is very small but this is expected as Treasury bill rates have been almost zero since

the financial crisis. We input these parameter estimates into equation (27) and (4) from
proposition 4 to calculate that λ = −6.0211 with very little need for computing power.
This estimate coincides with other authors as it is negative and of similar magnitude, refer

to table 3 for comparison.

This method for estimating the market price of variance risk, λ, makes the calibration
of of any Heston (1993) model much simpler, as λ is now a function of VXX prices and κ∗ .

4.1 The Market Price of Variance Risk and the Variance Risk
Premium

Eraker and Wu (2013) use an economic equilibrium model to show that the abysmal per-
formance of VIX futures and VIX futures index ETPs can be explained by the negative
Variance Risk Premium (V RP ). The V RP and the market price of variance risk, λ, are
5
VXX price data from NASDAQ website: www.nasdaq.com/symbol/vxx/historical.
Modeling VXX 17

similar concepts as they both measure the amount of compensation that risk adverse in-
vestors require for taking on the variance risk. Variance is negatively related with equity

returns and therefore the Variance Risk Premium and market price of variance are both
negative. Investors accept the negative returns during normal times, when taking a long
position in volatility, in order to hedge against times of high volatility where they will

receive a positive return from this long position, such as the 2008 financial crisis. Zhang
and Huang (2010) show that the market price of variance risk, λ, from the Heston (1993)
framework is almost proportional to the Variance Risk Premium, V RP , as defined by Carr

and Wu (2009) as long as λτ0 is small. Their result is shown by

    
1 2 1 1
V RP = κτ0 + O(κ τ0 ) θ + 2 2
− κτ0 + O(κ τ0 ) Vt λτ0 + O(λ2 τ02 ), (29)
6 2 3

where O(·) is a function of order λ2 τ02 (Zhang and Huang, 2010). The first part of the
equation is obviously proportional to λ as it is multiplied by λτ0 . The reason the relationship

between V RP and λ is almost proportional is because of the O(λ2 τ02 ) part of equation (29)
which is not proportional to λ but aslong as λτ0 is small (relative to 1) then λ2 τ02 will be
very small.

Our findings are consistent with those from Eraker and Wu (2013) as we find a negative
market price of variance risk drives the returns of the VXX to be so negative, through the
negative roll yield, and they find a negative Variance Risk Premium as the cause of the

negative returns of VIX futures positions and VIX futures ETNs.

5 Rebalancing Frequency of SPVXSTR

In this section we explore the effect of rebalancing frequency of the SPVXSTR. We start
by replicating the SPVXSTR index using VIX futures prices from the 20th of December
Modeling VXX 18

2005 until the 28th of March 20146 , using the methodology from S&P Dow Jones Indices
(2012). This replicated SPVXSTR time series is displayed in figure 5 along with the actual

SPVXSTR time series, the lines are almost exactly identical showing that our replication
is accurate.
Figure 6 shows four time series of the replicated SPVXSTR index with different rebal-

ancing frequencies of daily, weekly, bi-weekly and monthly rebalancing. The figure shows
that as the rebalancing frequency is decreased from daily to weekly, biweekly and monthly,
the SPVXSTR’s value decreases. If this effect exists going from daily to more frequent

rebalancing, for example hourly, then this would be a problem for our continuous time
model. To examine the effect of the rebalancing frequency on the price of the VXX for
smaller time steps than daily we needed a VIX futures price time series that was intraday,

but real data for this is only available to us for the last 50 days, therefore we chose to
simulate a five year long hourly VIX futures price time series.
To simulate the hourly time series of VIX futures prices we first need a time series

of instantaneous volatility, which we get from the physical measure stochastic process of
instantaneous variance Heston (1993), given by

p
dVt = κ(θ − Vt )dt + σv Vt dB. (30)

We then use the simple VIX futures price approximation, F0 , from equation (13) to
find a time series of nearest and second nearest maturing VIX futures prices. We use

κ∗ = 5.4642 as this is the most recent estimation, we λ = −6.0211 as calculated in section

2. We propose θ = 0.1 and σv = 0.4 as reasonable value. The results of this section are
not sensitive to what parameters are used, aslong as they are reasonable. For simplicity we

assume that VIX futures mature every 28 days, that there are no non-trading days, trading
hours are 24 hours of the day and that the risk free rate is zero.
6
Available at http://cfe.cboe.com/Data/HistoricalData.aspx#VX, accessed on the 20th of April 2014.
Modeling VXX 19

We then use the methodology from section 2 to calculate the SPVXSTR index for five
years with different rebalancing frequencies and a starting value of one.

Figure 7 shows the resulting SPVXSTR hourly time series for different rebalancing
frequencies from hourly to monthly. We can see in figure 7 that the simulated SPVXSTR
time series for hourly and daily rebalancing are almost identical. The rebalancing effect

going from daily to hourly rebalancing is therefore very very small and not a problem for
our model. There is however a rebalancing effect if the index is rebalanced less often than
dail, this is consistent with our findings using market VIX futures prices. To show that our

conclusion on the rebalancing frequency is robust to the term structure of VIX futures we
repeated the above exercise but holding Vt constant at different levels. This allows us to
create a time series of SPVXSTR with a upward sloping (in contango) VIX futures term

structure, as shown in figure 8, and downward sloping (in backwardation) VIX futures term
structure, as shown in figure 9.
From figures 8 and 9 we can see that the rebalancing frequency does not significantly

impact the SPVXSTR for hourly rebalancing. However there is a significcant effect when
going to less frequent rebalancing. These results are robust to the term structure shape
of VIX futures. Therefore we can conclude that shifting from a daily rebalancing to more

frequent rebalancing does not affect the returns of the SPVXSTR significantly. Also in both
figures the VXX model time series estimated using our model is the continuous limit of the
rebalancing time series. The VXX model line is almost identical to the daily rebalancing

time series, showing that our continuous time VXX model is adequate to model the discrete
time VXX.
Figures 8 and 9 also show the importance of the roll yield as a driver of the SPVXSTR

and subsequently the VXX. The two figures isolate the effect of the term-structure on the
returns of the VXX SPVXSTR, by holding Vt constant and we know that the roll yield is
a result of the term structure of VIX futures. When the term structure is upward sloping,
Modeling VXX 20

causing a negative roll yield, the simulated SPVXSTR will tend to 0 as in figure 8, and
when the term structure is downward sloping, causing positive roll yield, the simulated

SPVXSTR is exponentially increasing as in figure 9.

6 Conclusions and Discussions

We study the VXX ETP which has been traded very actively on the New York Stock
Exchange in recent years. We use the VIX futures price approximation from Zhang, Shu,
and Brenner (2010) and simplify it for the nearly 30-day VIX futures contract. From this

simplified formula for VIX futures prices we develop a model for the VXX. Our model is,
to our knowledge, the first ever model of the VXX which encompasses the dynamics of the
SPX index and the VIX index. Our model is the simplest way to model the VXX while

capturing the relationship between the SPX, VIX and the VXX.
Our model explains the large negative returns of the VXX very well and is in line with
the methodology from S&P Dow Jones Indices (2012). Our VXX model allows us to show

that the difference in returns of the constant 30-day maturity VIX futures contract, as in
Zhang, Shu, and Brenner (2010), and the VXX is due to the roll yield as suggested in the
literature. We then examine the roll yield and show that λ, the market price of variance

risk is the main driver of the roll yield. Therefore we have provided an explanation of the
VXX return puzzle as the constant 30-day maturity VIX futures contract does not exhibit
these negative price movements and is closely related to the VIX, therefore the roll yield

drives the negative returns of the VXX.

We have also provided a simple and robust way of measuring the market price of variance
risk, λ using our model and VXX prices. To understand the economic explanation for this

we suggest examining the economic model for VIX Exchange Traded Notes from Eraker and
Wu (2013). Their model finds that the negative return premium (Variance Risk Premium),
Modeling VXX 21

which is almost proportional to λ (Zhang and Zhu, 2006), is an equilibrium outcome because
long VIX futures positions allow investors to hedge against high volatility and low return

states, such as exhibited in a financial crisis.

Our continuously rebalanced VXX model is adequate for modeling the daily rebalanced
VXX as the effect of the rebalancing frequency is only significant at less frequent than daily

rebalancing.
Our model for the VXX is the first of its kind, as it is the only one that includes the
relationship between the SPX, the VIX and the VXX which we believe is fundamental in

understanding the VXX. Our model could also be used by practitioners to price options
written on the VXX, which can be regarded as Asian options written on the underlying
instantaneous variance. Bao, Li, and Gong (2012) have created a model for pricing VXX

options but they do not account for the dynamics of the S&P 500 or the VIX, which is
essential in modeling the VXX.
Our research shows that the roll yield is the main cause for the negative performance

of the VXX, as suggested in the literature. It would be interesting to see whether the roll
yield also plays a large part in the returns of other VIX futures ETPs, we expect that
it would. Our model could be expanded by using the full approximation formula of VIX

futures from Zhang, Shu, and Brenner (2010) and letting θ be time dependant. One could
use a similar approach to ours to explore the effect of the roll yield on other VIX futures
ETNs but we advise caution in using the simplified VIX futures price formula F0 as it

will be prone to more error at longer maturities. Further research is also needed into the
calibration technique best used for our model and its accuracy although it is theoretically
sound. Exploring similar approaches to the one in this article to create models of other

VIX futures Exchange Traded Products could help further develop the literature around
these popular yet mysterious investment products.
Modeling VXX 22

Appendix
A. Solving for CDRt model

From equation (15) and Ito’s lemma we get

∂ ln FtT 1 ∂ 2 ln FtT ∂ ln FtT

d ln FtT = dVt + (dV t ) 2
+ dt, (31)
∂Vt 2 ∂Vt2 ∂t
∂ ln FtT ∂ 2 ln FtT ∂ ln FtT
therefore we need to find each of the partial derivatives ∂Vt
, ∂Vt2
and ∂t
. These
are given by

∂ ln FtT 1 Be−κ(T −t)

=
∂Vt 2 [θ∗ (1 − Be−κ(T −t) ) + Vt Be−κ(T −t) ]
−1 , (32)
θ∗

1 ∗
= − θ + Vt
2 Be−κ∗ (T −t)
−2
∂ 2 ln FtT θ∗

1 ∗
=− − θ + Vt (33)
∂Vt2 2 Be−κ∗ (T −t)
and

∗
∂ ln FtT 1 κ∗ (Vt − θ∗ )Be−κ (T −t)
 
= . (34)
∂t 2 θ∗ + (Vt − θ∗ )Be−κ∗ (T −t)
We then substitute all the partial derivatives into equation (31) giving us the full func-
tion of the log futures return as in equation (17) from section 3.3.
Modeling VXX 23

References

Bao, Qunfang, Shenghong Li, and Donggeng Gong, 2012, Pricing VXX option with default

risk and positive volatility skew, European Journal of Operational Research 223, 246–255.

Barclays, 2013, VXX and VXZ Prospectus, http://www.ipathetn.com/static/pdf/vix-

prospectus.pdf, accessed on 30-6-2014.

Carr, Peter, and Liuren Wu, 2009, Variance risk premiums, Review of Financial Studies
22, 1311–1341.

Chen, Hsuan-Chi, San-Lin Chung, and Keng-Yu Ho, 2011, The diversification effects of
volatility-related assets, Journal of Banking & Finance 35, 1179–1189.

Cohen, Jerome B, Fischer Black, and Myron Scholes, 1972, The valuation of option con-
tracts and a test of market efficiency, The Journal of Finance 27, 399–417.

Cox, John C, Jonathan E Ingersoll Jr, and Stephen A Ross, 1985, A theory of the term
structure of interest rates, Econometrica: Journal of the Econometric Society 53, 385–

407.

Deng, Geng, Craig McCann, and Olivia Wang, 2012, Are VIX futures ETPs effective

hedges?, The Journal of Index Investing 3, 35–48.

Duan, Jin-Chuan, and Chung-Ying Yeh, 2010, Jump and Volatility Risk Premiums implied

by VIX, Journal of Economic Dynamics and Control 34, 2232–2244.

Eraker, Bjørn, and Yue Wu, 2013, Explaining the Negative Returns to VIX Futures and

ETNs: An Equilibrium Approach, Available at SSRN 2340070 .

French, Kenneth R, G William Schwert, and Robert F Stambaugh, 1987, Expected stock
returns and volatility, Journal of financial Economics 19, 3–29.
Modeling VXX 24

Hancock, GD, 2013, VIX Futures ETNs: Three Dimensional Losers, Accounting and Fi-
nance Research 2, 53–64.

Heston, Steven L, 1993, A closed-form solution for options with stochastic volatility with
applications to bond and currency options, Review of financial studies 6, 327–343.

Husson, Tim, and Craig McCann, 2011, The VXX ETN and Volatility Exposure, PIABA
Bar Journal 18, 235–252.

Lin, Yueh-Neng, 2007, Pricing VIX futures: Evidence from integrated physical and risk-
neutral probability measures, Journal of Futures Markets 27, 1175–1217.

Luo, Xingguo, and Jin E Zhang, 2012, The term structure of VIX, Journal of Futures
Markets 32, 1092–1123.

S&P Dow Jones Indices, 2012, S&P 500 VIX futures indices Methodology,
http://us.spindices.com/indices/strategy/sp-500-vix-short-term-index-mcap, accessed

on 30-6-2014.

Whaley, Robert E, 2013, Trading Volatility: At What Cost, The Journal of Portfolio

Management 40, 95–108.

Zhang, Jin E, and Yuqin Huang, 2010, The CBOE S&P 500 three-month variance futures,

Journal of Futures Markets 30, 48–70.

Zhang, Jin E, Jinghong Shu, and Menachem Brenner, 2010, The new market for volatility

trading, Journal of Futures Markets 30, 809–833.

Zhang, Jin E, and Yingzi Zhu, 2006, VIX futures, Journal of Futures Markets 26, 521–531.

Zhu, Yingzi, and Jin E Zhang, 2007, Variance term structure and VIX futures pricing,
International Journal of Theoretical and Applied Finance 10, 111–127.
Modeling VXX 25

Table 1: Summary statistics of the daily returns for the SPY, VIX and VXX.
This table shows the summary statistics and correlations of the VXX, SPX (S&P 500 index
ETP) and the VIX index returns from the 2nd February 2009 to the 13th August 2014.
RD represents estimates using discrete daily returns and RC represents estimates using
continuously compounded daily returns. The √ annualised standard deviation is calculated
by multiplying the standard deviation by 252. The Holding Period Return (HPR) is the
discrete return from the first price to the last price of the sample. The Compound Annual
Growth Rate (CAGR) is the constant yearly growth rate that would lead to the change from
1
the first price to the last price in the sample, it is calculated by CAGR = (HP R + 1) T − 1,
where T is the length of the sample in years.

SPX VIX VXX

RD RC RD RC RD RC
Mean 0.08% 0.07% 0.15% −0.09% −0.32% −0.39%
significance p-value (0.0121) (0.0217) (0.4323) (0.6226) (0.0020) (0.0001)
Standard Deviation (σ) 1.13% 1.13% 7.10% 6.86% 3.81% 3.78%
Annualised σ 18.00% 18.00% 112.70% 108.97% 60.55% 60.06%
Skew −0.1280 −0.2377 1.3281 0.7659 0.7811 0.5213
significance p-value (0.0596) (0.0006) (0.0000) (0.0000) (0.0000) (0.0000)
Excess Kurtosis 4.5210 4.4989 5.6978 3.8815 3.2867 2.8287
significance p-value (0.0000) (0.0000) (0.0000) (0.0000) (0.0000) (0.0000)
Holding Period Return 164.48% 97.26% −71.66% −126.09% −99.55% −540.82%
CAGR 19.25% − −20.41% − −62.43% −

Correlations RD RC
SPY VIX VXX SPY VIX VXX
SPY 1 −0.7659 −0.7834 1 −0.7710 −0.7846
significance p-value (0.0000) (0.0000) (0.0000) (0.0000)
VIX − 1 0.8660 − 1 0.8651
significance p-value (0.0000) (0.0000)
VXX − − 1 − − 1
Modeling VXX 26

Table 2: 30-day VIX futures price estimation. This table shows the different VIX
futures price estimates using four different formulae and range of parameter estimates for Vt ,
σV and κ∗ . For this exercise we keep the time to maturity constant at 30 days, τ = τ0 = 365
30
∗ ∗ ∗
and θ constant at θ = 0.10. The first four columns show the hypothetical θ , Vt , σV and
κ∗ parameters used in the futures price estimates. The first column of estimated futures
prices, labelled by F0 , uses the simple approximation for VIX futures prices, the F0 part of
equation (13). The next column of VIX futures prices, labelled by F0 + F1 , uses the simple
formula of VIX futures prices and the first half of the convexity adjustment, F0 + F1 from
equation (13). The F0 + F1 + F2 column of VIX futures prices uses the full approximation
formula, equation (13), from Zhang, Shu, and Brenner (2010). The final column of VIX
futures prices uses the accurate formula, equation (11), from Zhang and Zhu (2006). The
columns labelled % error, are the percentage difference of the preceding column of prices
from the prices estimated by the accurate formula.

Parameters VIX Futures Price estimates

∗ ∗
θ Vt κ σV F0 % error F0 + F1 % error F0 + F1 + F2 % error Accurate F
0.1 0.04 4 0.1 25.14 0.07% 25.12 0.00% 25.12 0.00% 25.12
0.1 0.04 4 0.4 25.14 1.08% 24.86 −0.02% 24.89 0.09% 24.87
0.1 0.04 4 0.7 25.14 3.20% 24.30 −0.25% 24.56 0.82% 24.36
0.1 0.04 5.5 0.1 26.32 0.05% 26.31 0.00% 26.31 0.00% 26.31
0.1 0.04 5.5 0.4 26.32 0.77% 26.12 −0.02% 26.13 0.05% 26.12
0.1 0.04 5.5 0.7 26.32 2.27% 25.69 −0.17% 25.85 0.42% 25.74
0.1 0.04 7 0.1 27.26 0.04% 27.25 0.00% 27.25 0.00% 27.25
0.1 0.04 7 0.4 27.26 0.57% 27.11 −0.01% 27.12 0.03% 27.11
0.1 0.04 7 0.7 27.26 1.69% 26.78 −0.12% 26.88 0.24% 26.81
0.1 0.12 4 0.1 33.51 0.05% 33.49 0.00% 33.49 0.00% 33.49
0.1 0.12 4 0.4 33.51 0.82% 33.24 0.00% 33.25 0.05% 33.24
0.1 0.12 4 0.7 33.51 2.53% 32.67 −0.02% 32.83 0.46% 32.68
0.1 0.12 5.5 0.1 33.20 0.04% 33.19 0.00% 33.19 0.00% 33.19
0.1 0.12 5.5 0.4 33.20 0.67% 32.98 0.00% 32.99 0.03% 32.98
0.1 0.12 5.5 0.7 33.20 2.04% 32.53 −0.04% 32.64 0.31% 32.54
0.1 0.12 7 0.1 32.95 0.03% 32.94 0.00% 32.94 0.00% 32.94
0.1 0.12 7 0.4 32.95 0.55% 32.77 0.00% 32.78 0.02% 32.77
0.1 0.12 7 0.7 32.95 1.66% 32.39 −0.04% 32.48 0.21% 32.41
0.1 0.2 4 0.1 40.17 0.04% 40.15 0.00% 40.15 0.00% 40.15
0.1 0.2 4 0.4 40.17 0.62% 39.92 0.00% 39.93 0.03% 39.92
0.1 0.2 4 0.7 40.17 1.94% 39.41 0.01% 39.51 0.27% 39.40
0.1 0.2 5.5 0.1 38.88 0.03% 38.87 0.00% 38.87 0.00% 38.87
0.1 0.2 5.5 0.4 38.88 0.55% 38.67 0.00% 38.68 0.02% 38.67
0.1 0.2 5.5 0.7 38.88 1.69% 38.23 0.00% 38.32 0.21% 38.24
0.1 0.2 7 0.1 37.79 0.03% 37.77 0.00% 37.77 0.00% 37.77
0.1 0.2 7 0.4 37.79 0.48% 37.61 0.00% 37.61 0.02% 37.61
0.1 0.2 7 0.7 37.79 1.47% 37.23 −0.01% 37.30 0.16% 37.24
Modeling VXX 27

Table 3: λ and κ∗ estimates by various authors. This table shows the estimated value
of λ, the market price of variance risk, and κ∗ , the risk neutral speed of mean reversion of
variance, from different authors using different sample periods and estimation methods.

Author Data period κ∗ λ

Lin (2007) 21 Apr 2004 - 18 Apr 2006 5.3500 -0.3528
Duan and Yeh (2010) 2 Jan 2001 - 29 Dec 2006 -1.7956 -7.5697
Zhang and Huang (2010) 18 May 2004 - 17 Aug 2007 1.2989 -19.1184
†
Luo and Zhang (2012) 2 Jan 1992 - 31 Aug 2009 5.4642
Our estimation 30 Jan 2009 - 27 Jun 2014 5.4642 -6.0211‡

†
Luo and Zhang (2012) do not give the estimate for lambda, but their article is important
here as we use their κ∗ estimate.
‡
We use the κ∗ = 5.4642 estimate from Luo and Zhang (2012) and assume that it is
accurate for our sample period.
Modeling VXX 28

Figure 1: Understanding the SPVXSTR Roll Period. This diagram shows how dr
and dt are determined for the calculation of the weights in each VIX futures contract of
the SPVXSTR. Ti is the settlement date of the ith nearest maturing VIX futures, which is
30 days before S&P 500 options maturity date (3rd Friday of every month) and is usually
on a Wednesday. Ti − 1 is the day before ith nearest maturing VIX futures settlement and
the last day of the roll period. On the last day of the roll period the nearest settling VIX
futures is eliminated and the second nearest settling VIX futures becomes the nearest. The
dr and dt are the factors used in the calculation of the weights of each of the VIX futures
contracts in the SPVXSTR, as shown in section 2. The roll period represents the time
during which the weight in the nearest settling VIX futures contract is gradually replaced
by a position in the second nearest VIX futures contract. At the end of the roll period
all the weight will be in the second nearest VIX futures contract which then becomes the
nearest as the old nearest matures, then the next roll period starts, and the process is
repeated.
Modeling VXX 29

Figure 2: Historical VIX, 30-day VIX futures price and VXX price. This figure
shows the level of the VIX and the price of 30-day VIX futures on the primary vertical
axis and the VXX price on the secondary vertical axis. The 30-day VIX futures contract
is the linearly interpolated price of a constant 30-day maturity VIX futures contract, as in
Zhang, Shu, and Brenner (2010).
Modeling VXX 30

Figure 3: Market Capitalization and Trading Value of VXX. This figure shows the
daily dollar trading volume and market capitalization of the VXX from the 30th January
2009 to the 27th June 2014 in billion US dollars.
Modeling VXX 31

Figure 4: VIX Term Structure. This figure shows the term structure of VIX futures
prices from 1 day to 50 day maturity calculated using our simple approximation ,F0 , the
approximation with the first part of the convexity adjustment, F0 + F1 and the full approx-
imation from Zhang, Shu, and Brenner (2010) , F0 + F1 + F2 . These estimated VIX futures
prices are calculated using constant parameter estimates of θ∗ = 0.1, κ∗ = 5, σV = 0.1425
and Vt = 0.06 but the time to maturity varies from 1 day to 50 days.
Modeling VXX 32

Figure 5: Replicated vs. Actual SPVXSTR. This figure shows the actual SPVXSTR
time series and our replicated SPVXSTR time series using the methodology from S&P Dow
Jones Indices (2012) from the 20th December 2005 until the 28th March 2014.
Modeling VXX 33

Figure 6: Replicated SPVXSTR, different Rebalancing frequencies. This figure

shows four different time series of our replication of SPVXSTR. SPVXSTR daily corre-
sponds to daily, SPVXSTR weekly to weekly, SPVXSTR bi-weekly to two weekly and
SPVXSTR monthly to monthly rebalancing. The final values of the indices are 1178.63 for
daily, 1088.38 for weekly, 842.24 for bi-weekly and 264.60 for monthly rebalancing.
Modeling VXX 34

Figure 7: Simulated index using physical process for Vt . This figure shows the
simulated SPVXSTR index over our 4 year simulation period using V0 = 0.02, σV = 0.4,
the risk-neutral parameter estimates κ∗ = 5.4642 and θ∗ = 0.1, the physical process of dVt
as described in equation (30) and the simple VIX futures price formula, F0 , from equation
eqrefapproxVIXfuture from Zhang, Shu, and Brenner (2010). The label of each time series
corresponds to the rebalancing frequency used.
Modeling VXX 35

Figure 8: Simulated index using Vt = θ < θ ∗ . This figure shows the time series of
the simulated SPVXSTR, when the instantaneous variance is set constant at Vt = θ =
0.0476 < θ∗ = 0.1 forcing a upward sloping VIX futures term structure. To calculate the
futures prices we use the volatility of volatility σV = 0.4 and the risk-neutral parameter
estimates κ∗ = 5.4642 and θ∗ = 0.1 are used. The label of each time series corresponds to
the rebalancing frequency used.
Modeling VXX 36

Figure 9: Simulated index using Vt = 0.14 > θ ∗ . This Figure shows the time
series of the simulated SPVXSTR, when Vt is set constant at 0.14 which is higher than
θ∗ = 0.1 forcing a downward sloping VIX futures term structure. To calculate the futures
prices we use the volatility of volatility σV = 0.4 and the risk-neutral parameter estimates
κ∗ = 5.4642 and θ∗ = 0.1 are used. The label of each time series corresponds to the
rebalancing frequency used.