Journal of Banking and Finance 82 (2017) 180–190

Contents lists available at ScienceDirect

Journal of Banking and Finance

journal homepage: www.elsevier.com/locate/jbf

Optimal delta hedging for optionsR

John Hull∗, Alan White
Joseph L. Rotman School of Management, University of Toronto, 105 St George Street, Toronto, Ontario, Canada M5S 3E6

a r t i c l e i n f o a b s t r a c t

Article history: As has been pointed out by a number of researchers, the normally calculated delta does not minimize
Received 17 March 2016 the variance of changes in the value of a trader’s position. This is because there is a non-zero correlation
Accepted 15 May 2017
between movements in the price of the underlying asset and movements in the asset’s volatility. The
Available online 16 May 2017
minimum variance delta takes account of both price changes and the expected change in volatility condi-
JEL classification: tional on a price change. This paper determines empirically a model for the minimum variance delta. We
G13 test the model using data on options on the S&P 500 and show that it is an improvement over stochas-
tic volatility models, even when the latter are calibrated afresh each day for each option maturity. We
Keywords: also present results for options on the S&P 100, the Dow Jones, individual stocks, and commodity and
Options
interest-rate ETFs.
Delta
© 2017 The Authors. Published by Elsevier B.V.
Vega
Stochastic volatility This is an open access article under the CC BY-NC-ND license.
Minimum variance (http://creativecommons.org/licenses/by-nc-nd/4.0/)

1. Introduction Delta is by far the most important hedge parameter and for-
tunately it is the one that can be most easily adjusted as it only
The textbook approach to managing the risk in a portfolio of requires a trade in the underlying asset. Ever since the birth of
options involves specifying a valuation model and then calculat- exchange-traded options markets in 1973, delta hedging has played
ing partial derivatives of the option prices with respect to the un- a major role in the management of portfolios of options. Option
derlying stochastic variables. The most popular valuation models traders adjust delta frequently, making it close to zero, by trading
are those based on the assumptions made by Black and Scholes the underlying asset.
(1973) and Merton (1973). When hedge parameters are calculated Even though the Black–Scholes–Merton model assumes volatil-
from these models, the usual market practice is to set each option’s ity is constant, market participants usually calculate a “practitioner
volatility parameter equal to its implied volatility. This is some- Black–Scholes vega” to measure and manage their volatility expo-
times referred to as using the “practitioner Black-Scholes model.” sure. This vega is the partial derivative of the option price with
The “practitioner Black-Scholes delta” for example is the partial respect to implied volatility with all other variables, including the
derivative of the option price with respect to the underlying as- asset price, kept constant.1 This approach, although not based on
set price with other variables, including the implied volatility, kept an internally consistent model, has the advantage of simplicity. The
constant. price of an option at any given time is, to a good approximation,
a deterministic function of the underlying asset price and the im-
plied volatility.2 A Taylor series expansion shows that the risks be-
ing taken can be assessed by monitoring the impact of changes in
R
We thank Peter Carr, Peter Christoffersen, Tom Coleman, Emanuel Derman, these two variables.
Bruno Dupire, Andrew Lesniewski, Andrei Lyashenko, Tom McCurdy, Massimo As is well known, there is a negative relationship between
Morini, Michael Pykhtin, Lorenzo Ravagli, Managing Editor Geert Bekaert, and two an equity price and its volatility. This was first shown by Black
anonymous reviewers, as well as seminar participants at University of Toronto, the
(1976) and Christie (1982) who used physical volatility estimates.
Fields Institute, the 2015 RiskMinds International conference, a Bloomberg seminar
in 2016, a Global Risk Institute seminar in 2016, and the Derivatives and Volatil-
ity conference at NYU Stern in 2017 for helpful comments. We are grateful to the
Global Risk Institute in Financial Services for funding. Earlier versions of the paper 1
In a portfolio of options dependent on a particular asset, the options typically
were circulated with titles “Optimal Delta Hedging” and “Optimal Delta Hedging for have different implied volatilities. The usual practice when vega is calculated is to
Equity Options.” calculate the portfolio vega as the sum of vegas of the individual options. This is
∗
Corresponding author. equivalent to considering the impact of a parallel shift in the volatility surface.
2
E-mail addresses: [email protected] (J. Hull), [email protected] This is exactly true if we ignore uncertainties relating to interest rates and div-
(A. White). idends.

http://dx.doi.org/10.1016/j.jbankfin.2017.05.006
0378-4266/© 2017 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license. (http://creativecommons.org/licenses/by-nc-nd/4.0/)
 J. Hull, A. White / Journal of Banking and Finance 82 (2017) 180–190 181

Other authors have shown that it is true when implied volatility Scholes vega times the partial derivative of the expected implied
estimates are used. One explanation for the negative relation is volatility with respect to the asset price. Improving delta there-
leverage. As the equity price moves up (down), leverage decreases fore requires an assumption about the partial derivative of the ex-
(increases) and as a result volatility decreases (increases). In an al- pected implied volatility with respect to the asset price. Crépey
ternative hypothesis, known as the volatility feedback effect, the (2004) and Vähämaa (2004) test setting the partial derivative equal
causality is the other way round. When there is an increase (de- to (or close to) the (negative) slope of the volatility smile, as sug-
crease) in volatility, the required rate of return increases (de- gested by the local volatility model.4 Alexander et al. (2012) build
creases) causing the stock price to decline (increase). The two com- on the research of Derman (1999) and test eight different models
peting explanations have been explored by a number of authors for the partial derivative, including a number of regime-switching
including French et al. (1987), Campbell and Hentschel (1992), models.
Bekaert and Wu (20 0 0), Bollerslev et al. (20 06), Hens and Steude This paper extends previous research by determining empiri-
(2009), and Hasanhodzic and Lo (2013). On balance, the empirical cally a model for the partial derivative of the expected implied
evidence appears to favor the volatility feedback effect. volatility with respect to asset price. We show that, when the
A number of researchers have recognized that the negative rela- underlying asset is the S&P 500, this partial derivative is to a
tionship between an equity price and its volatility means that the good approximation a quadratic function of the practitioner Black-
practitioner Black-Scholes delta does not give the position in the Scholes delta of the option divided by the product of the asset
underlying equity that minimizes the variance of the hedger’s posi- price and the square root of the time to maturity. This leads to
tion. The minimum variance (MV) delta hedge takes account of the a simple model where the MV delta is calculated from the practi-
impact of both a change in the underlying equity price and the ex- tioner Black-Scholes delta, the practitioner Black–Scholes vega, the
pected change in volatility conditional on the change in the under- asset price, and the time to option maturity. We show that the
lying equity price. Given that delta hedging is relatively straightfor- hedging gain from approximating the MV delta in this way is bet-
ward, it is important that traders get as much mileage as possible ter than that obtained using a stochastic volatility model or a lo-
from it. Switching from the practitioner Black–Scholes delta to the cal volatility model. The results have practical relevance to traders,
minimum variance delta is therefore a desirable objective. Indeed many of whom still base their decision making on output from the
it has two advantages. First, it lowers the variance of daily changes practitioner Black–Scholes model. The hedging gain from using our
in the value of the hedged position. Second, it lowers the resid- approach for options on other indices was similar to that for op-
ual vega exposure because part of vega exposure is handled by the tions on the S&P 500. The approach also led to a hedging gain for
position that is taken in the underlying asset. options on individual stocks and ETFs, but this was not as great as
A number of stochastic volatility models have been suggested in for options on indices.
the literature. These include Hull and White (1987, 1988), Heston The structure of the rest of the paper is as follows. We first dis-
(1993), and Hagan et al. (2002). A natural assumption might be cuss the nature of the data that we use. Second, we develop the
that using a stochastic volatility model automatically improves theory that allows us to parameterize the evolution of the implied
delta. In fact, this is not the case if delta is calculated in the usual volatilities of options. The theory is then implemented and tested
way, as the partial derivative of the option price with respect to out-of-sample using options on the S&P 500. The results are com-
the asset price. To calculate the MV delta, it is necessary to use pared with those from a stochastic volatility and a local volatility
the model to determine the expected change in the option price model. Based on the results for the S&P 500 we then carry out
arising from both the change in the underlying asset and the asso- tests for options on other indices and for options on individual
ciated expected change in its volatility. stocks and ETFs.
A number of researchers have implemented stochastic volatil-
ity models and used the models’ assumptions to convert the 2. Data
usual delta to an MV delta. They have found that this produces
an improvement in delta hedging performance, particularly for We used data from OptionMetrics. This is a convenient data
out-of-the-money options. The researchers include Bakshi et al. source for our research. It provides daily prices for the underlying
(1997) who implemented three different stochastic volatility mod- asset, closing bid and offer quotes for options, and hedge param-
els using data on call options on the S&P 500 between June 1988 eters based on the practitioner Black–Scholes model. We chose to
and May 19913 ; Bakshi et al. (20 0 0), who looked at short and long- consider options on the S&P 500, S&P 100, the Dow Jones Indus-
term options on the S&P 500 between September 1993 and August trial Average of 30 stocks (DJIA), the individual stocks underlying
1995; Alexander and Nogueira (2007), who looked at call options the DJIA, and five ETFs. The assets underlying three of the ETFs
on the S&P 500 during a six month period in 2004; Alexander are commodities, gold (GLD), silver (SLV) and oil (USO). The as-
et al. (2009), who consider the hedging performance of six dif- sets underlying the other two ETFs were the Barclays U.S. 20+ year
ferent models using put and call options on the S&P 500 trading Treasury Bond Index (TLT) and the Barclays U.S. 7–10 year Treasury
in 2007; and Poulsen et al. (2009) who looked at data on S&P Bond Index (IEF). The options on the S&P 500 and the DJIA are Eu-
500 options, Eurostoxx index options, and options on the U.S. dol- ropean. Both European and American options on the S&P 100 are
lar euro exchange rate during the 20 04 to 20 08 period. Bartlett included in our data set. Options on individual stocks and those
(2006) shows how a minimum variance hedge can be used in con- on ETFs are American. The period covered by the data we used is
junction with the SABR stochastic volatility model proposed by January 2, 2004–August 31, 2015 except for the commodity ETFs
Hagen et al (2002). where data was first available in 2008.5
This paper is different from the research just mentioned in that Only option quotes for which the bid price, offer price, implied
it is not based on a stochastic volatility model. It is similar in spirit volatility, delta, gamma, vega, and theta were available were re-
to papers such as Crépey (2004), Vähämaa (2004) and Alexander tained. The option data set was sorted to produce observations for
et al. (2012). These authors note that the minimum variance delta the same option on two successive trading days. For every pair of
is the practitioner Black–Scholes delta plus the practitioner Black–
4
See for example Derman et al (1995) and Coleman et al (2001). The local
volatility model was suggested by Derman and Kani (1994) and Dupire (1994).
3 5
They also looked at puts on the S&P 500, but did not report the results as they This is a much longer period than that used by other researchers except
were similar to calls. Alexander et al (2012).
182 J. Hull, A. White / Journal of Banking and Finance 82 (2017) 180–190

observations the data was normalized so that the underlying price MV hedge as the percentage increase in the effectiveness of an
on the first of the two days was one. Options with remaining lives MV hedge over the effectiveness of the practitioner Black–Scholes
less than 14 days were removed from the data set. Call options hedge. Thus:
for which the practitioner Black-Scholes delta was less than 0.05
SSE [ f − δMV S]
or greater than 0.95, and put options for which the practitioner Gain = 1 − (3)
Black-Scholes delta was less than –0.95 or greater than –0.05 were SSE [ f − δBS S]
removed from the data set. For options on individual stocks, in ad- where SSE denotes sum of squared errors.8
dition to the filters used for options on the indices, days on which
stock splits occurred were removed.
After all the filtering there remain more than 1.3 million price 4. Analysis of S&P 500 options
quotations for both puts and calls on the S&P 500, about 0.5 mil-
lion observations for the other indices and ETFs, and about 20 0,0 0 0 In this section we examine the characteristics of the MV delta
observations for options on each individual stock in the Dow Jones for options on the S&P 500 with the objective of determining the
Industrial Index. The trading volume for puts on the S&P 500 is functional form of the MV delta. Once we have a candidate func-
much greater than that for calls.6 Puts and calls trade in approx- tional form, we will test it out of sample for both options on the
imately equal volumes for other indices. Calls trade more actively S&P 500 and options on other assets.
than puts for the individual stocks. Trading tends to be concen- We start with an implementation based on Eq. (1) applied to
trated in close-to-the-money and out-of-the-money options. One daily price changes:
notable feature is that the trading of close-to-the-money call op-
 f = δMV S + ε (4)
tions is particularly popular. The majority of trading is in options
with maturities less than 91 days. where ε is an error term. Because the mean of S and f are both
close to zero, minimizing the variance of ε in this equation, and
3. Background theory other similar equations that we will test, is functionally equivalent
to minimizing the sum of squared values. Several other variations
In the Black-Scholes model the underlying asset price follows on the model were tried such as using non-normalized data, re-
a diffusion process with constant volatility. Many alternatives to placing f with f – θ BS t, where θ BS is the practitioner Black–
Black-Scholes have been developed in an attempt to explain the Scholes theta9 and t is one trading day, or including an inter-
option prices that are observed in practice. These involve stochas- cept. None of the variations had a material effect on the results
tic volatility, jumps in the asset price or the volatility, risk aversion, we present. The results that we report are for the model in (4).
and so on. Departures from Black-Scholes tend to reduce the per- We estimated Eq. (4) for options with different moneyness and
formance of delta hedging. For example, Sepp (2012) shows that time to maturity. Moneyness was measured by δ BS . We created
this is so for a mixed-jump diffusion model and some of the pa- nine different moneyness buckets by rounding δ BS to the nearest
pers referenced earlier show that this is so for stochastic volatility tenth and seven different option maturity buckets (14–30 days, 31–
models. In this section we provide a theoretical result for deter- 60 days, 61–91 days, 92–122 days, 123–182 days, 183–365 days,
mining the minimum variance delta from the practitioner Black- and more than 365 days). For each delta and each maturity bucket,
Scholes delta. The result involves the implied volatility and is ex- the value of δ MV was estimated. In all cases δ MV – δ BS was nega-
actly true in the limit for diffusion processes while being an ap- tive. This result is consistent with the results of other researchers.
proximation in the case of other models. It means that traders of S&P 500 index options should under-hedge
Define S as a small change in an asset price and f as the call options and over-hedge put options relative to relative to the
corresponding change in the price of an option on the asset. The hedge suggested by the practitioner Black–Scholes model.10
minimum variance delta, δ MV , is the value that minimizes the vari- The bucketed results show that δ MV – δ BS is not heavily de-
ance of7 pendent on option maturity and is roughly quadratic in δ BS . It is
approximately true that11
 f − δMV S (1)
√
We show in Appendix A that it is approximately true that νBS = S T G(δBS )
   
∂ fBS ∂ fBS ∂ E σimp ∂ E σimp for some function G where T is the time to the option maturity.
δMV = + = δBS + νBS (2) From this equation, Eq. (2), and the assumption of scale invari-
∂S ∂ σimp ∂S ∂S
ance12 we obtain
where fBS is the Black-Scholes-Merton pricing function, σ imp is the ν  
implied volatility, δ BS is the practitioner Black-Scholes delta, ν BS δMV = δBS + √BS a + bδBS + cδBS
2
(5)
is the practitioner Black-Scholes vega, and E(σ imp ) is the expected
S T
value of the implied volatility as a function of S.
Other authors, in particular Alexander et al. (2012), have ex- 8
Using standard deviations rather than SSEs would produce a similar measure
plored the effectiveness of various estimates ∂ E(σ imp )/∂ S in deter- but the Gain would be numerically smaller
mining the minimum variance delta. In what follows we estimate 9
The practitioner Black-Scholes theta is the partial derivative with respect to the
this function empirically and then conduct out-of-sample tests of passage of time with the volatility set equal to the implied volatility) and time is
measured in days. If the asset price and its implied volatility do not change, the
the effectiveness of the estimated function.
option price can be expected to decline by about θ BS t in one day.
When presenting our results, we shall define the effectiveness 10
A call has a positive delta and the MV delta, δMV , is less positive than δBS ; a put
of a hedge as the percentage reduction in the sum of the squared has a negative delta and δ MV is more negative than δ BS.
√
residuals resulting from the hedge. We denote the Gain from an 11
For European options,
√ νBS = S T N (d1 )e−qT where d1 =
[ln(S/K ) + (r − q + σ /2 )T ]/σ T , K is the strike price, T is the time to matu-
2

rity, r is the risk-free rate, q is the dividend yield, and N is the cumulative normal
6
The bid-offer spread for puts on the S&P 500 is smaller than that for calls ex- distribution function.
√ However, δBS = N (d1 )e−qT so that √ d1 = N (δBS e ). As a
−1 qT

cept in the case of deep in-the-money options where the spreads are about the result, vBS = S T N  (N −1 (δBS eqT ) )e−qT . If q is zero vBS /(S T )is dependent only on
same. δ BS . When q is small this is approximately true.
7
An early application of this type of hedging analysis to futures markets is 12
A scale invariant model is one where the distribution of St /S0 is independent of
Ederington (1979) S0 . See for example Alexander and Nogueira (2007).
 J. Hull, A. White / Journal of Banking and Finance 82 (2017) 180–190 183

Fig. 1. The estimated parameters for the quadratic in Eq. (6) for puts and calls on the S&P 500 observed between 2004 and 2015. The estimations use overlapping 36-month
periods. For call options the negative of the bˆ parameter is plotted so that the same scale can be used for both charts.

where a, b, and c are constants. Applying (5) to Eq. (2) shows that volatility is a quadratic function of our measure of moneyness, δ BS .
  The same model applies across the range of deltas considered.
  a + bδBS + cδ 2
S We now consider how well our empirical hedge ratio model
E σimp = √ BS
T S works in reducing uncertainty. We estimate the MV delta using
historical data and then use that estimate to reduce the variance of
In the balance of the paper we examine the effectiveness of the the hedging error in the future. In carrying out this test we use a
approximation in Eq. (5) for δ MV . moving window where parameters are estimated over a 36-month
period and then used to determine MV hedges during the follow-
ing month. The first month for which MV hedges are estimated is
5. Out of sample tests of S&P 500 options
January 2007 and the last is August 2015. We tested moving win-
dows of length between 12- and 60-months but did not find that
To this point our work has been largely descriptive, motivated
any one of these was materially better than the others.13
by a desire to produce a simple model of how the volatility surface
The only element of our simple model that is unknown is the
for S&P 500 options evolves as a result of stock price changes. Our
quadratic function of moneyness in Eq. (5). We estimate the model
simple model is that for a particular moneyness and a particular
stock price change, the expected size of the change in the implied
volatility is inversely proportional to the square-root of the option 13
In all our reported results we consider one day changes in option prices and
life. For a particular option maturity and a particular percentage implied volatilities when estimating the MV hedge parameters. Slightly better re-
stock price change, the expected size of the change in the implied sults occur if the observation period is increased to several trading days.
184 J. Hull, A. White / Journal of Banking and Finance 82 (2017) 180–190

parameters, a, b and c, using a regression model based on Eqs.

(4) and (5).
ν S  
 f − δBS S = √BS a + bδBS + cδBS
2
+ε (6)
T S
where f is the one-day change in the option price, S is the
change in the stock price, S, T is the remaining life of the op-
tion, and δ BS and vBS are the delta and vega calculate using the
practitioner Black-Scholes model. This model is fitted to all options
in each 36-month estimation period. The estimation is done sepa-
rately for puts and calls. The estimated coefficients, aˆ, bˆ , and cˆ, are
shown in Fig. 1. Usually, the parameters of the best fit quadratic
model change slowly through time, but during the credit crisis of
2008 some extreme changes were observed.
The three coefficients estimated in a 36-month estimation pe-
riod are used to determine the hedge error resulting when the es- Fig. 2. The average R2 , the fraction of total variance of changes in implied volatility
timated model is used to hedge each option on each day in the explained by changes in the index, for all options on the S&P 500 observed between
following test month. The hedging error based on this model, ε MV , 2004 and 2015.
is The horizontal axis is ordered so that high strike prices are on the right hand end

ν S   and low strike prices are on the left.

εMV =  f − δBS S − √BS aˆ + bˆ δBS + cˆδBS
2
T S
The results for put options are somewhat different. The frac-
The hedging error based on standard Black-Scholes hedging,
tion of total variance in implied volatilities explained by changes in
εBS , is
the stock price is much smaller than that observed for call options.
εBS =  f − δBS S There is much more idiosyncratic variation in the implied volatili-
ties of put options. As a result, MV hedging of puts is less effective
Once the hedging errors have been calculated for all 104
than for calls. We note that the different hedging performance for
months the Gain (Eq. (3)) resulting from using our model to hedge
puts and calls cannot be explained by the model driving prices. For
is then calculated as
any model (including jump models), put call parity shows that the
SSE (εMV ) hedging error for a put should in theory be the same as that for
Gain = 1 −
SSE (εBS ) the corresponding call. The observed differences led us to carry out
The Gain was calculated including the residuals for all options a test of put-call parity.
and then considering only the residuals from options in a particu-
lar delta bucket. This resulted in one overall Gain and 9 bucketed 5.1. A put-call parity test
Gains for each test month.
When the residuals for all options are included, the average We used our quadratic model to test how well put-call parity
Gain is about 26% for calls and 23% for puts. The average Gain has held over the period covered by our data. We first used the
achieved for each delta bucket is shown in Table 1. This shows that put-call parity relationship to turn all call prices in our data set
for call options the Gain is largest for out-of-the-money options into synthetic put prices. We estimated aˆ, bˆ , and cˆ in our quadratic
(a Gain of about 42% for the highest strike options) and smallest ˆˆ
form using the actual put prices, and aˆˆ, b, and cˆˆ using the synthetic
(about 17%) for in-the-money options. For put options the Gains put prices for each of our three-year calibration periods. We then
are higher for low strike options (out-of-the-money) and lower for calculated the root mean square error of the difference between
high strike (in-the-money) options. We confined our hedging effec- the estimated put parameters and the put parameters calculated
tiveness test to options with maturities greater than 13 days. This from the synthetic put data under the assumption that put-call
eliminates very short term options. Including the very short matu- parity holds:
rity options slightly worsens our results due to the large gammas 
of short-term options that are close to the money. 2 2 2
MV hedging works better for calls than puts and better for out- (aˆ − aˆˆ ) + (bˆ − bˆˆ ) + (cˆ − cˆˆ)
RMSE = (8)
of-the-money options than in-the-money options. To understand 3
why this is the case we directly estimate the relationship between The results are shown in Fig. 3. These results suggest that put-
implied volatility changes and stock price changes by estimating α call parity was seriously violated before December 2008 but that
in thereafter it was approximately true. (The first observation of the
α S post-December 2008 period is December 2011.)14
σimp = √ +ε (7)
T S
6. Comparison with alternative models
The estimation is done separately for puts and calls for every
delta bucket using all options observed between 2007 and 2015.
In the previous section we tested an empirical model to deter-
The R2 for each delta bucket is shown in Fig. 2.
mine the minimum variance delta hedge. The results show that a
The R2 from the estimation in Eq. (7) for calls with δ BS in the
reasonable improvement in hedging accuracy can be achieved in
0.1 bucket is about 0.60. That is, the change in the implied volatil-
this way. However, as mentioned earlier, other researchers have
ity due to changes in the stock price explains about 60% of the to-
calculated minimum variance deltas from stochastic volatility mod-
tal variation in implied volatilities. As δ BS increases the average R2
els and local volatility models. In this section we compare the
declines due to increased idiosyncratic noise in the implied volatil-
ity data. This explains the effectiveness of MV hedging for out-of-
the money calls and the declining effectiveness of MV hedging for 14
Some violations of put-call parity are probably created by our use of mid-
in-the-money calls. market prices rather than transaction prices.
 J. Hull, A. White / Journal of Banking and Finance 82 (2017) 180–190 185

Table 1
The out-of-sample average hedging Gain (Eq. (3)) January 2007 to August 2015 for options on the S&P 500 from MV delta hedging. The SABR and Local Vol
models are calibrated daily for each option maturity and applied to determine the hedge for the next day. The Empirical model parameters, a, b and c in Eq. (5)
are estimated using all options traded in a 36 month window and then applied to determine the hedge on every day in the next month. Results are reported
for buckets based on rounding δ BS to the nearest tenth. The upper panel shows the hedging GAIN values. The lower panel shows the Newey-West adjusted
t-statistic for the difference between the empirical model GAIN and the alternative model GAINs.

GAIN

Call options Put options

δ BS empirical model SABR model Local vol δ BS EMPIRICAL model SABR model Local vol

0.1 42.1% 39.4% 42.6% –0.9 15.1% 11.2% –7.4%

0.2 35.8% 33.4% 36.2% –0.8 18.7% 19.6% 6.8%
0.3 31.1% 29.4% 30.3% –0.7 20.3% 17.7% 9.1%
0.4 28.5% 26.3% 26.7% –0.6 20.4% 16.7% 9.2%
0.5 27.1% 24.9% 25.5% –0.5 22.1% 16.7% 10.8%
0.6 25.7% 25.2% 25.2% –0.4 23.8% 17.7% 12.0%
0.7 25.4% 24.7% 25.8% –0.3 27.1% 21.7% 16.8%
0.8 24.1% 23.5% 25.4% –0.2 29.6% 25.8% 20.6%
0.9 16.6% 17.0% 16.9% –0.1 27.5% 26.9% 17.7%
All 25.7% 24.6% 25.5% All 22.5% 19.0% 10.2%

t-Statistic

Call options Put options

δ BS Empirical – SABR Empirical – local vol δ BS Empirical – SABR Empirical – local vol

0.1 3.94 –0.06 –0.9 9.71 27.11

0.2 3.75 0.79 –0.8 –2.34 12.34
0.3 3.22 4.31 –0.7 1.39 10.42
0.4 4.22 6.55 –0.6 3.29 10.25
0.5 4.74 6.65 –0.5 6.09 11.53
0.6 2.95 5.16 –0.4 7.22 13.89
0.7 3.24 2.82 –0.3 7.17 14.18
0.8 2.70 –0.45 –0.2 5.46 13.97
0.9 –0.01 0.44 –0.1 0.66 17.63
All 8.74 8.13 All 12.75 41.99

   
performance of our empirical model with these two categories of fBS F0 + F , σimp (F0 + F , σ0 + ξ ρ F /F ) − fBS F0 , σimp (F0 , σ0 )
= (10)
models. F

6.1. Stochastic volatility model The procedure for implementing this model is as follows. On
each trading day the implied volatilities of all options with a par-
The stochastic volatility model we use is a particular version of ticular maturity are determined.16 The parameters for the stochas-
the SABR model discussed by Hagen et al. (2002):15 tic volatility model (σ 0 , ξ , and ρ ) that are to be used for that par-
ticular option maturity are chosen to minimize the sum of squared
dF = σ F dz differences between the market implied volatilities and the model
dσ = ξ σ dw (9) implied volatilities.17 Once the model parameters are determined
where F is the futures stock price when the numeraire is the zero for the particular maturity, the minimum variance delta is then
coupon bond with maturity T. The dz and dw are Wiener processes determined for each option with that maturity using Eq. (10). This
with constant correlation ρ and ξ is a constant volatility of volatil- procedure is repeated for every option maturity observed on each
ity parameter. In this model the expected change in the volatility trading day.
given a particular change in the futures price is To align the tests of the stochastic volatility model with the
tests of our empirical model we calibrated the model for every
dF
E ( d σ |d F ) = ξ ρ option maturity every day from the start of 2007 to August 2015.
F Puts and calls were considered separately. Since there are about 13
Hagan et al. (2002) and Rebonato et al. (2011) show that under different maturities observed on each trading the SABR model re-
the model defined by Eq. (9), a good analytic approximation to the quires about 78 model parameters to be estimated on each trading
implied volatility for a European option can be produced. day. In total, about 29,0 0 0 optimizations are carried out and about
Define fBS (F , σ ) as the value of an option given by the Black– 87,0 0 0 model parameters are estimated. The estimated parame-
Scholes–Merton assumptions when the futures stock price is F and ters are reasonable and provide a good fit to the observed implied
the volatility is σ . An option, an estimate of the minimum variance volatilities. The average initial volatility, σ 0 , is about 19% which is
delta given by the model is then approximately equal to the average at-the-money option implied
volatility, the average volatility of the volatility, ξ , is about 1.2, and
E ( f |F )
δSV =
F
16
In practice the SABR model is used as a model for the behavior of all options
with a particular maturity. When calibrated to all options of all maturities we find
15 that it provides poor results. This is not surprising as the model is not designed to
As pointed out by Poulsen et al (2009), similar results are obtained for different
stochastic volatility models. In the general SABR model dF = σ F β dz. Setting β = 1 fit the term structure of implied volatilities.
17
ensures scale invariance which is a reasonable property for equities and equity in- For a particular maturity to be included in our sample on any day we require
dices. The model we choose is equivalent to a version of the model in Hull and that there be options with more than 10 different strike prices and that the root
White (1987). mean square error in fitting the implied volatilities be smaller than 1%.
186 J. Hull, A. White / Journal of Banking and Finance 82 (2017) 180–190

Fig. 3. The root mean squared difference between estimated parameters for the quadratic in Eq. (6) for put options, and the parameter values for put option prices that are
calculated from call option prices under the assumption that put-call parity holds (Eq. (8)). The estimation uses overlapping 36-month periods based on options on the S&P
500.

the average correlation is about –0.85 while the average root mean ferent from zero and for puts when δ = – 0.8 the SABR model out-
square error in fitting the implied volatility is about 0.32%. performs the empirical model. However, for options which trade
The upper panel of Table 1 compares the Gain from the SABR actively, the empirical model is clearly better than SABR.
model with the Gain from the empirical model developed in this
paper. The results are aggregated by practitioner Black-Scholes 6.2. Local volatility model
delta rounded to the nearest tenth. The table shows the stochas-
tic volatility model is worse at reducing hedging variance than our The slope of the volatility smile plays a key role determining
empirical model. The results are particularly compelling because the partial derivative of the expected implied volatility with re-
the SABR model utilizes many more parameters than our model. spect to the asset price for the local volatility model. Under the
In our empirical approach we estimate only the three coefficients local volatility model, for at-the-money options the rate of change
of the quadratic function (Eq. (6)) and update the estimates once in the implied volatility with respect to changes in the underlying
a month. There are a total of 104 calibrations and a total of 312 price is equal to the slope of the volatility smile. This is exactly
parameters are estimated. This can be contrasted with the SABR true for futures options and approximately true if the difference
model where nearly 10 0,0 0 0 parameters are estimated. Overall the between the interest rate and the dividend yield are small. This has
SABR model performs less well than our empirical model. Its per- been discussed and proved by many authors (for example, Derman
formance is better than the empirical model only for very-deep-in- et al. (1995) and Coleman et al. (2001)). In this case the MV delta,
the-model options. Eq. (2) becomes
In Appendix B we develop the procedure for calculating the t- ∂ σimp
statistic used to determine the statistical significance of the dif- δMV = δBS + νBS
∂K
ference between the Gains for two different hedging procedures.
We now apply this result to the S&P 500 options we are con-
Since the sample size is always greater than 10 0,0 0 0 this t-statistic
sidering. We assume this result is approximately true for options
can be considered to be a z-statistic. The lower panel of Table
which are not at the money. This is equivalent to the assumption
1 shows the t-statistic for the difference between the Gain aris-
that (a) the volatility smile is linear and (b) the volatility smile ex-
ing from the quadratic hedging model and the Gain from hedg-
hibits parallel shifts. These two assumptions are approximately, but
ing with the SABR model. Considering all call options together, t is
not exactly, true.
8.24 while for all puts it is 11.52. For individual deltas it is greater
We find that a quadratic gives an excellent fit to the implied
than 2.5 in all cases except for deep in-the-money call and put op-
volatility smile for a particular maturity. We therefore determined
tions. For calls when δ = 0.9 the difference is not statistically dif-
the slope of smile model for each maturity on each day by fitting a
 J. Hull, A. White / Journal of Banking and Finance 82 (2017) 180–190 187

Table 2
The average out-of-sample hedging Gain (Eq. (3)) from MV delta hedg-
ing when the model parameters, a, b and c in Eq. (6) are estimated using
options with all strikes and maturities observed in a 36 month window
and then applied to determine the hedge in the next month. Results are
reported for each delta bucket for European (XEO) and American (OEX)
options on the S&P 100 and for European options on the Dow Jones In-
dustrial Index (DJX).

Call options Put options

δ BS XEO OEX DJX δ BS XEO OEX DJX

0.1 36.2% 35.1% 23.9% –0.9 10.7% 32.0% –1.3%

0.2 32.8% 26.4% 28.7% –0.8 15.7% 28.4% –0.7%
0.3 27.7% 22.2% 31.6% –0.7 20.1% 27.4% 1.7%
0.4 26.0% 21.4% 29.5% –0.6 19.4% 25.3% 5.2%
0.5 23.9% 18.5% 29.2% –0.5 20.9% 23.8% 7.8%
0.6 24.0% 17.0% 28.3% –0.4 20.9% 23.2% 10.9%
0.7 22.5% 17.7% 28.9% –0.3 22.9% 24.8% 15.0%
0.8 21.2% 15.4% 24.7% –0.2 24.9% 26.2% 17.0% Fig. 4. The average R2 , the fraction of total variance of changes in implied volatility
0.9 16.1% 8.2% 15.7% –0.1 24.8% 25.9% 16.0% explained by changes in the stock price. The results shown are averaged across the
All 23.0% 16.7% 26.5% All 19.7% 27.1% 5.5% stocks underlying the DJIA.
The horizontal axis is ordered so that high strike prices are on the right hand end
and low strike prices are on the left.

quadratic function to the smile and using it to determine the slope

of the smile for each option. The results are shown in upper panel The results for put options are a bit more complicated. The re-
of Table 1 18 and the t-statistic for the difference between the Gain sults for XEO options were similar to those for options on the S&P
arising from the quadratic hedging model and the Gain from hedg- 500 and the results for DJX options are similar but weaker. The
ing with the local volatility model are shown in the lower panel. weaker DJX results may be caused by the fact that there are only
The performance of the local volatility model is not significantly 30 stocks in the index which means that there will be more id-
different from that of the empirical model for deep in- and out-of- iosyncratic variation in the implied volatilities.
the-money calls. In all other cases the local volatility hedging is Our results for in-the-money American (OEX) options are dif-
much worse than the empirical model hedging. This is particularly ferent from our results for all other assets in that the Gain for put
true for put options. options is large for both in- and out-of-the-money options. Over-
all, the conclusion that can be drawn from Table 2 is that our rule
of thumb for hedging works as well for American options as for
7. Results for other stock indices European options.

We now return to a consideration of the empirical model and 8. Results for single stocks and ETFs
test how well it works for other stock indices. Specifically, we con-
sider European (ticker XEO) and American (ticker OEX) options on We repeated the out-of-sample hedging tests based on the
the S&P 100, and European options on the Dow Jones Industrial quadratic model in Eq. (6) for each of the thirty individual stocks
Index (ticker DJX). We implement out-of-sample tests similar to underlying the DJX and each of the five ETFs. The average hedg-
those done on the S&P 500. The two contracts on the S&P 100 are ing variance reduction, averaging across the 30 stocks, the 3 com-
the same except for exercise terms. They therefore allow us to ex- modities and the 2 interest rate products, found in these tests is
plore the degree to which hedging differs for American options. reported in Table 3.
The out of sample test was based on estimating the three pa- The average hedging gain for call options on single stocks are
rameters of the quadratic function in Eq. (6) using options of all similar to but rather smaller than those for options on the Dow
strikes and maturities. The model parameters were estimated using Jones Industrial Average. For put options the results are very poor.
a 36-month estimation period and the three estimated parameters MV hedging contributes nothing or has a negative effect for puts.
were then used to delta hedge for a one-month testing period. The To understand why this is the case we carried out the regression
Gain (Eq. (3)) resulting from using our model to hedge in the test in Eq. (7) for puts and calls for every delta bucket for each of
periods is then calculated. The Gain achieved for puts and calls in the 30 stocks. The average R2 across the thirty stocks is shown in
each delta bucket is shown in Table 2. Fig. 4.
The results for call options for all indices are essentially the The R2 exhibits the same pattern observed in Fig. 2 for options
same as those found for options on the S&P 500. It is tempting to on the S&P 500 but is somewhat smaller than that for the index
think that the results for the American style (OEX) call options are options indicating that the idiosyncratic noise is larger for individ-
the same as those for the European style options because American ual stocks. The increased idiosyncratic noise reduces the MV hedg-
style call options are almost never exercised early and hence are ing effectiveness by inserting a wedge between parameters esti-
effectively European. However, for more than 80% of the sample mated in one period and the parameters that would produce the
tested the S&P 100 dividend yield is more than 1.5% higher than most effective MV hedge in the following period. The results for
the interest rate.19 In these circumstances the probability of early puts are quite different. The fraction of the variance of changes in
exercise is high. As a result, it appears that the American feature the implied volatility explained by stock price changes, R2 , is es-
of the option does not affect the hedging effectiveness of our rule sentially zero. As a result we can expect no improvement from MV
of thumb. hedging which is what we see.
The results for the ETFs are divided into results for options
18
on commodities (gold, silver, and oil) and options on interest-rate
We experimented with other implementations of the slope-of-smile model but
did not obtain better results. products (20+ year Treasury Bonds and 7–10 year Treasury bonds).
19
The relevant interest rates were almost invariably at least 1.5% lower than the The results for options on commodities are similar to those for
dividend yield between January 2009 and August 2015. individual stocks while the results for interest-rate products are
188 J. Hull, A. White / Journal of Banking and Finance 82 (2017) 180–190

Table 3
The average out-of-sample hedging Gain (Eq. (3)) observed across the 30 stocks underlying the
DJX, the three commodities (gold, silver and oil), and two interest-rate products (20 + year
bonds and 7–10 year bonds). The model parameters, a, b and c in Eq. (6) are estimated using
options with all strikes and maturities observed in a 36 month window and then applied to
determine the hedge in the next month. Gain results are reported for each delta bucket based
on rounding δ BS to the nearest tenth.

Call options Put options

δ BS stocks Commodities Int. rates δ BS stocks Commodities Int. rates

0.1 32.2% 28.9% 5.2% –0.9 0.9% 3.3% 0.0%

0.2 26.8% 20.8% 3.6% –0.8 1.7% 5.1% –0.6%
0.3 23.2% 15.2% 2.1% –0.7 3.7% 5.1% –1.1%
0.4 19.1% 12.5% 2.0% –0.6 5.7% 5.3% –1.1%
0.5 15.0% 9.8% 1.3% –0.5 5.2% 4.0% –0.5%
0.6 11.4% 7.3% 1.5% –0.4 3.5% 1.9% 1.6%
0.7 8.3% 4.2% 1.6% –0.3 1.8% –0.9% 3.1%
0.8 5.4% 1.6% 0.2% –0.2 0.7% –3.0% 5.8%
0.9 2.5% 0.6% –1.2% –0.1 2.6% –4.1% 6.2%
All 10.3% 7.7% 1.4% All 2.5% 2.5% 0.2%

much weaker. As in most of the stock index results, MV hedging each other and (b) that in the case of American options puts are
provides a much bigger Gain for call options on commodities than less likely to be exercised early than call options for most of our
put options and the gain is greatest for out-of-the-money options. sample period. It appears that the reason for the discrepancy be-
The negative correlation between price and implied volatility for tween calls and puts is a result of a very high level of idiosyncratic
commodities and interest rate products cannot be explained by a noise in the prices of put options.
leverage and therefore lends support for the volatility feedback hy- The most striking result is the ubiquity of the negative relation
pothesis. between asset price and implied volatilities for call option prices.
When asset prices rise, implied volatilities decline resulting in an
9. Conclusions MV delta that is less than the practitioner Black-Scholes delta. For
options on equities and equity indices this might be explained by
Delta is by far the most important Greek letter. It plays a key a leverage argument. As equity prices rise the firm becomes less
role in the management of portfolios of options. Option traders levered and equity volatility declines. However, this argument does
take steps to ensure that they are close to delta neutral at least not seem to apply to commodity or bond prices. For these assets it
once a day and derivatives dealers usually specify delta limits for seems likely that we have to rely on the volatility feedback effect
their traders. This paper has investigated empirically the difference in which an increase in volatility raises the required rate of return
between the practitioner Black–Scholes delta and the minimum resulting in a stock price decline.
variance delta. The negative relation between price and volatility
for equities means that the minimum variance delta is always less
than the practitioner Black–Scholes delta. Traders should under- Appendix A
hedge equity call options and over-hedge equity put options rel-
ative to the practitioner Black–Scholes delta. The implied volatility is defined as the volatility which when
The main contribution of this paper is to show that a good es- inserted into the Black-Scholes option pricing function results in a
timate of the minimum variance delta can be obtained from the model price that equals the market price. Suppose that we observe
practitioner Black–Scholes delta and an empirical estimate of the the price of an option price, f, when the stock price is S. The im-
historical relationship between implied volatilities and asset prices. plied volatility is implicitly defined by20
 
We show that the expected movement in implied volatility for an f = fBS S, σimp (A1)
option on a stock index can be approximated as a quadratic func-
tion in the option’s practitioner Black–Scholes delta divided by the where fBS is the Black-Scholes-Merton pricing function and σ imp is
square root of time. This leads to a formula for converting the the implied volatility. The Black–Scholes-Merton pricing function is
practitioner Black–Scholes delta to the minimum variance delta. continuous and continuously differentiable.
When the formula is tested out of sample, we obtain good results First let us consider a two-factor world in which the stock price
for both European and American call options on stock indices. For and the volatility of the stock price obey diffusion processes. In
options on the S&P 500 we find that our model gives better re- this world the option price also changes continuously which allows
sults that either a stochastic volatility model or a model based on us to expand Eq. (A1) in a Taylor series
the slope of the smile.  
∂f ∂f
Call options on individual stocks and ETFs exhibit the same gen- f = S + σ + O S 2
eral behavior as call options on stock indices, but the effectiveness ∂S ∂ σimp imp
of MV hedging is greatly reduced because there is more noise in or equivalently
the relationship between volatility changes and price changes. For
nearly all the assets we considered, the results for put options are  f = δBS S + vBS σimp + ε (A2)
much worse than those for call options. In the case of put options
on individual stocks and ETFs, the results are particularly disap-
pointing in that virtually none of the variation in changes in im- 20
Note that σ imp , is a “catch-all” parameter capturing the difference between the
plied volatility is explained by changes in stock prices. The rela- Black–Scholes pricing function and how option prices are determined in the real
world. It leads to an exact pricing function that is always correct in the real-world.
tively poor performance of MV hedging for put options is a puzzle All the changes we consider in f, S, and σ imp are changes in the real world. The
because (a) in the case of the European options considered put-call risk-neutral world, although used to derive any particular option pricing model, is
parity means that puts and calls can be regarded as substitutes for not relevant to the derivation of Eq (2) or our empirical work in Section 4.
 J. Hull, A. White / Journal of Banking and Finance 82 (2017) 180–190 189

√
where δ BS and vBS are the delta and vega as defined by the prac- from zero is nμ ˆ /σˆ . If the individual observations are not inde-
titioner Black–Scholes model and ε denotes the residual higher or- pendent but exhibit serial correlation, a Newey-West adjustment
der terms in the series. Subtracting δ MV S from both sides we factor must be applied to the estimated variance. Given the way in
have which the data is aggregated, the procedure used to determine the
adjustment factor warrants explanation.
 f − δMV S = (δBS − δMV )S + vBS σimp + ε (A3)
There are approximately 54,0 0 0 unique put and call options in
Conditioning on S and taking expectations we obtain the S&P 500 data set. Each of these options has a unique expiry
  date and strike price and may fall into different aggregation groups
E σimp E (ε )
δMV = δBS + vBS + at different times. For each of these unique options, the time se-
S S ries of the Diff statistic was calculated and the Newey–West ad-
In the case of diffusion processes as S approaches zero the justment factors for lags between 1 and 50 were calculated. The
last term is infinitesimal so average adjustment factor, averaging across all the option series,
  was then calculated for each lag. The magnitude of the average
E σimp
δMV = δBS + vBS (A4) adjustment factor plateaued at about 30 lags and so we used 30
S lags to determine the adjustment factor used for each option se-
which leads to Eq. (2). ries. (Using 20 or 40 lags did not materially change the results.)
Now consider the more general case in which there are many The estimate of the adjusted sample variance is
factors driving asset prices and the processes may not be dif-
 
fusion processes. In this case the option price may change dis-
n  2
continuously and the Taylor series expansion omits the effects σˆ =
2
wi xi − yi − μ
ˆ (n − 1 )
of other state variables. As a result, the residual term in Eq. i=1
(A2), ε , now includes the effects of the omitted state variables
and, if the option price changes discontinuously, may be large. where wi is the Newey–West adjustment factor for the option se-
In this case Eq. (2) is only approximately true since it omits the ries on which the ith observation is based.
term E(ε )/S.

Appendix B References

This appendix describes the calculation of the t-statistic used Alexander, C., Kaeck, A., Nogueira, L.M., 2009. Model risk adjusted hedge ratios. J.
Futures Markets 29 (11), 1021–1049.
to determine the statistical significance of the difference between
Alexander, C., Nogueira, L.M., 2007. Model-free hedge ratios and scale invariant
the Gain values for two different hedging methods. Consider a set models. J. Bank. Finance 31, 1839–1861.
of n observations of one-day changes in option prices. These price Alexander, C., Rubinov, A., Kalepky, M., Leontsinis, S., 2012. Regime-dependent
smile-adjusted delta hedging. J. Futures Markets 32 (3), 203–229.
changes can be hedged using the practitioner Black–Scholes delta,
Bakshi, G., Cao, C., Chen, Z., 1997. Empirical performance of alternative option pric-
a minimum variance delta based on our quadratic function, or a ing models. J. Finance 52 (5), 2003–2049.
minimum variance delta based on the SABR model. Let xi be the Bakshi, G., Cao, C., Chen, Z., 20 0 0. Pricing and hedging long-term options. J. Econo-
squared residual from hedging the ith one-day change in option metrics 94, 277–318.
Bartlett, B, 2006. Hedging under SABR model. Wilmott Mag. (July / August) 2–4.
price using the quadratic function, yi be the squared residual from Bekaert, G., Wu, G., 20 0 0. Asymmetric volatility and risk in equity markets. Rev.
hedging with the SABR model, and zi be the squared residual from Financ. Stud. 13, 1–42.
Black–Scholes hedging. Black, F., 1976. Studies of stock price volatility changes. In: Proceedings of
the Business and Economics Section of the American Statistical Association,
The Gain from quadratic hedging is pp. 177–181.
Black, F., Scholes, M., 1973. The pricing of options and corporate liabilities. J. Polit.

n
n
Econ. 81 (May-June), 637–659.
GainQuad = 1 − xi / zi
Bollerslev, Tim, Julia, Litvinova, George, Tauchen, 2006. Leverage and volatility feed-
i=1 i=1 back effects in high-frequency data. J. Financ. Econometrics 4 (3), 353–384.
Campbell, J.Y., Hentschel, L., 1992. No news is good news: an asymmetric model of
The Gain from SABR-hedging is
changing volatility in stock returns. J. Financ. Econ. 31, 281–331.

n
n Christie, A.A., 1982. The stochastic behavior of common stock variances: value,
GainSABR = 1 − yi / zi leverage and interest rate effects. J. Financ. Econ. 10 (December 4), 407–432.
Coleman, T, Kim, Y., Li, Y., Verma, A., 2001. Dynamic hedging with a deterministic
i=1 i=1 local volatility model. J. Risk 4 (1), 63–89.

n Crépey, S., 2004. Delta-hedging vega risk. Quant. Finance 4 (October), 559–579.
Defining i=1 zi = Z, the difference in the Gains is
Derman, E., 1999. Volatility regimes. Risk 14 (2), 55–59.
Derman, E., Kani, I., 1994. Riding on a Smile. Risk 7 (February), 32–39.
1
n
Diff = ( xi − yi ) Derman, E., Kani, I., Zou, J.Z., 1995. Goldman Sachs Selected Quantitative Strategies
Z Reports December.
i=1 Dupire, B., 1994. Pricing with a smile. Risk 7 (February), 18–20.
Ederington, L.H., 1979. The hedging performance of the new futures market. J. Fi-
Suppose the individual difference observations, xi – yi , are
nance 34 (March), 157–170.
draws from a distribution with mean μ and standard deviation σ . French, K.R., Schwert, G.W., Stambaugh, R.F., 1987. Expected stock returns and
These parameters can be estimated from the observed values of volatility. J. Financ. Econ. 19, 3–30.
the differences: Hagan, P.S., Kumar, D., Lesniewski, A.S., Woodward, D.E., 2002. Managing smile risk.
Wilmott Mag. (September) 84–108.

n Hasanhodzic, J. and A. Lo, “Black’s leverage effect is not due to leverage,” Working
μˆ = (xi − yi )/n Paper, MIT, 2013.
i=1
Hens, T., Steude, S.C., 2009. The leverage effect without leverage. Finance Res. Lett.
  6, 83–94 (2009).
n 
2 Heston, S.L., 1993. A closed form solution for options with stochastic volatility with
σˆ =
2
xi − yi − μ
ˆ (n − 1 ) applications to bonds and currency options. Rev. Financ. Stud. 6 (2), 327–343.
Hull, J.C., White, A., 1987. The pricing of options on assets with stochastic volatili-
i=1
ties. J. Finance 42 (June), 281–300.
If the individual observations are independent then the esti- Hull, J.C., White, A., 1988. An analysis of the bias in option pricing caused by a
√ stochastic volatility. Adv. Futures Options Res. 3, 27–61.
mated mean and standard deviation of Diff are nμ ˆ /Z and nσˆ /Z, Merton, R.C., 1973. Theory of rational option pricing. Bell J. Econ. Manage. Sci. 4
and the t-statistic to test whether Diff is significantly different (Spring), 141–181.
190 J. Hull, A. White / Journal of Banking and Finance 82 (2017) 180–190

Poulsen, R., K., Schenk-Hoppé, R., Ewald, C.-O, 2009. Risk minimization in stochas- Sepp, A., 2012. An approximate distribution of delta-hedging errors in a jump-dif-
tic volatility models: model risk and empirical performance. Quant. Finance 9 fusion model with discrete trading and transaction costs. Quant. Finance 12 (7),
(September 6), 693–704. 1119–1141.
Rebonato, R., McKay, K., White, R., 2011. The SABR/LIBOR Market Model: Pricing, Vähämaa, S., 2004. Delta hedging with the smile. Financ. Markets Portfolio Manage.
Calibration and Hedging for Complex Interest-Rate Derivatives. John Wiley & 18 (3), 241–255.
Sons.