The forecasting power of short-term options∗

† ‡ § ¶
Arthur Bőők Juan F. Imbet Martin Reinke Carlo Sala

March 27, 2022

21:43

Abstract

We propose robust option-implied measures of conditional volatility, skewness and kur-

tosis based upon quantiles and expectiles inferred from weekly options on the S&P 500. All
quantities are by construction forward-looking and estimated non-parametrically through a
novel robust and arbitrage-free natural smoothing spline technique that produces quick to
estimate volatility smiles. We find that the option-implied robust indicators exhibit short-,
medium- and long-term predictive ability for the U.S. equity risk premium, market volatility,
skewness and kurtosis, both in- and out-of-sample, and outperform equal indicators inferred
from historical returns.
Keywords: Volatility smile, Quantiles, Expectiles, Weekly options, Forecasting.

JEL classification: G10, G13, G14, G17.

∗
We are grateful for helpful comments and suggestions from seminar participants at the FMARC21,
the XXII Workshop on Quantitative Finance and the Munich Finance Day 2022. There is an accom-
panying web-app for the paper which illustrates the estimation results. The app can be accessed here:
https://share.streamlit.io/mreinke1/app-quantile-expectile/main/app-quantile-expectile.py
†
Department of Financial Management and Control, Universitat Ramon Llull, ESADE, Avenida de Torreblanca
59, 08172 Sant Cugat, Barcelona, Spain; E-mail: arthuremilalexis.book@alumni.esade.edu.
‡
Université Paris-Dauphine, PSL Research University, DRM, Pl. du Maréchal de Lattre de Tassigny, 75016
Paris, France; E-mail: juan.imbet@dauphine.psl.eu.
§
Institute for Finance and Banking, LMU Munich School of Management, Ludwigstrasse 28, 80539 Munich,
Germany; E-mail: reinke@lmu.de.
¶
Department of Financial Management and Control, Universitat Ramon Llull, ESADE, Avenida de Torreblanca
59, 08172 Sant Cugat, Barcelona, Spain; E-mail: carlo.sala@esade.edu. Financial support from the AGAUR -
SGR 2017-640 grant is gratefully acknowledged.

Electronic copy available at: https://ssrn.com/abstract=3622433

1 Introduction
Option contracts are forward-looking financial assets that embed the investors’ expectations for
different future states of the world (strike prices) at different time horizons (time-to-maturity).
These expectations naturally contain the investors’ risk preferences, which determine the in-
vestors’ required expected rate of return for any asset pricing framework. As a consequence,
functions of option market data might contain valuable information for forecasting future
returns. Using weekly options written on the S&P 500, we derive short-term (weekly) option-
implied quantiles and expectiles and check their forecasting power through different conditional
robust indicators. More precisely, we infer option-implied conditional quantiles and expectiles
non-parametrically through a novel arbitrage-free natural smoothing spline technique that pro-
duces robust and quick to estimate volatility smiles. From them, we calculate different robust
option-implied measures of conditional dispersion (volatility), asymmetry (skewness), and flat-
ness (kurtosis) and find that these indicators have short-, medium- and long-term forecasting
power, both in- and out-of-sample, that is not present in the same indicators once estimated
using realized returns in the underlying asset.
Assessing investors’ beliefs from option-market data is not a trivial task. Estimation requires
absence of arbitrage in the option prices used, and correction for the noise that illiquid options
could carry in the estimation. As documented by Hentschel (2003), most implied volatility
estimates are noisy and prone to bias. We tackle this issue by proposing a novel approach to
estimate arbitrage-free volatility smiles. Our estimation contains two steps: First, we perform
a cubic interpolation over the implied volatility/delta space that gives a dense pre-estimate of
the volatility smile. Second, we solve a quadratic program enabling us to compute the linear
and quadratic terms of a natural spline that yields arbitrage-free points over the full volatility
smile.
We benchmark our estimation approach (hencefort BIRS1 ) on two popular methods to infer
the risk-neutral distribution (RND) from option prices. The Positive Convolution Approxi-
mation (PCA) of Bondarenko (2003), and the Fast and Stable (FS) approach of Jackwerth
(2004). We find that the BIRS approach exhibits better performance in recovering implied
volatility smiles and observed option prices from real and simulated data. The FS is a curve
fitting technique on the implied volatility, while the PCA uses convolutions to estimate the
RND. Our approach is also fully non-parametric, allowing us a fair comparison across models.
Using real data we test which approach better recovers the observed implied volatility smile
and observed option prices. Using simulated data we test which technique performs better in
recovering the option-implied cumulative distribution function. In both situations, the BIRS
1
The name BIRS is due to the surnames of the authors, in alphabetical order.

Electronic copy available at: https://ssrn.com/abstract=3622433

approach exhibits a better performance.
Using weekly option prices, we study the predictive ability of the robust option-implied
quantities. Our results show that all robust option-implied indicators have in- and out-of-
sample forecasting power on the first two realized moments of S&P returns. Option-implied
interquantile and interexpectile indicators perform well in-sample, with 7 days-ahead R2 around
2.5%, and increasing with the time-horizon of the estimates. Out-of-sample, the option-implied
2
conditional quantile and expectile asymmetry indicators produce an ROOS (Campbell and
Thompson (2007)) using stock based indicators as a benchmark of above 1% at one week,
2
and above 6% at six months respectively. We also find that the ROOS in predicting volatility
increases up to 41% in one week ahead regressions. Both in terms of expected returns and
volatility our indicator are statistically robust and economically significant. Economically in
fact, it makes sense to find higher (lower) forecasting power as the time-horizons increases
(decreases). This is due to the empirically well-known difficulty in forecasting the expected
returns at short time horizons, due to the different behavioral irrationalities that could make
the price diverge from the fundamental (estimated) one. Moreover, since stock-based measures
are valid only if the markets are ergodic and stationary, the use of stock returns make almost
impossible to forecast expected return at short-time horizons, due to the impossibility of the
model to react quickly to changing market scenarios. We overcome this issue using weekly
option market data and, as expected, we are able to forecast expected stock returns also at
very short time horizons (one week). Similarly the difficulty in forecasting the market volatility
at long one is due to the well-know clustering property of volatility.
The two steps carried out in the BIRS estimation are complementary. On one hand, the
implied volatility smile produced in the first step provides a good fit on the original data,
but does not ensure absence of arbitrage. On the other hand, the quadratic program removes
static arbitrage, but requires an accurate and dense pre-estimation to reduce the error of the
quadratic approximation between the spline knots. Combining these two elements, we obtain
a robust and arbitrage-free implied volatility smile, which allows us to extract option-implied
quantiles and expectiles from the risk-neutral distribution (RND). The non-parametric and
forward-looking nature of our robust coefficients allow us to overcome empirical drawbacks such
as model and parameter uncertainty and statistical issues that arise when relying on historical
data (Linn et al. (2017), Barone-Adesi et al. (2020)).
While historically observed price-paths do not disclose the underlying process from which
they are generated, option prices enable the re-creation of the expected price distributions in
the form of implied densities. By construction, options are forward-looking financial assets,
and their liquidity in major markets has become high enough to provide timely estimates with
non-parametric models. Most of these estimates follow from the seminal papers of Breeden

Electronic copy available at: https://ssrn.com/abstract=3622433

and Litzenberger (1978) and Banz and Miller (1978), which show how the expected future
states of the underlying price can be inferred by differentiating the option prices with respect
to their strike prices. These quantities have been found to provide forward-looking information
of predictive value for future realizations of index returns (see, among others, the review of
Figlewski (2018) or the survey of Christoffersen et al. (2011) and references therein). Despite
the growing literature on the use of option market data for predictions, not much has been said
on the forecasting power of short-term option-implied quantiles and expectiles.
This paper develops and tests a novel approach to build conditional robust quantities
implied by option prices. These robust quantities - conditional variability indicators, conditional
asymmetry indicators and conditional flatness indicators - are all derived non-parametrically,
with the goal to preserve the subtleties of the shapes of the empirical data. More precisely,
inspired by Barone Adesi (2016) and Bellini et al. (2018) we derive and estimate non-parametric
option-implied quantiles and expectiles respectively. Barone Adesi (2016) show how to use
option-implied quantiles to infer option-implied risk measures. The empirical analysis on the
S&P 500 of Barone-Adesi et al. (2019) and on the oil market of Barone-Adesi et al. (2019),
show that both the option-implied value at risk (VaR) and the option-implied conditional value
at risk (CVaR) are good alternatives to classical risk measures based on historical returns. In
particular, as also confirmed by Molino and Sala (2020) the option-based risk measures perform
well when mostly needed - that is in periods of high volatility, when the statistical properties of
the underlying deviate the most from the past, and cannot be captured using a large amount
of historical returns.
Bellini et al. (2018) also focus on risk management, and show how, using Italian data, option-
implied expectiles can be used to construct option-implied indexes of variability comparable to
the VIX index, while demanding less data. Bellini et al. (2021) extend the analysis of Bellini
et al. (2018) on the US market and test some properties of the S&P 500 index option-implied
expectiles, confirming its ability in producing sensible risk measures without the need of a
huge amount of data. While both Barone Adesi (2016) and Bellini et al. (2018) mostly provide
theoretical insights on the option-implied quantiles and expectiles, they are silent on their out-
of-sample forecasting power of option-implied quantiles and expectiles and related quantities.
Moreover, while both paper lay the ground to the theoretical foundation on option-implied
quantiles and expectiles, they do not consider the importance of a solid volatility smile for the
estimation of their quantities.
To test the out-of-sample forecasting power of robust implied measure we first infer the
option-implied quantiles and expectiles from options, and then we calculate the robust dis-
tribution measures. As a robust measures, we propose option-implied robust coefficient of
conditional variability (Barone Adesi (2016) and Bellini et al. (2018)) as a robust indicator for

Electronic copy available at: https://ssrn.com/abstract=3622433

the second moment, option-implied robust coefficients of conditional asymmetry (Hinkley (1975)
and Ghysels et al. (2016)) as a robust indicator for the third moment, and option-implied robust
coefficients of conditional flatness (Ruppert (1987)) as a robust indicator for the fourth moment.
Robust coefficients of conditional variability are in the form of inter-quantile and inter-expectile
ranges. Being defined as the difference between two quantiles or expectiles, the inter-quantile
and inter-expectile ranges are alternative measures of market variability. It is worth noticing
that while inter-quantile and inter-expectile ranges share common properties, the theoretical
properties of the inter-expectile range (in particular their property of being consistent with the
convex order, Bellini et al. (2016)) makes (at least theoretically) inter-expectile a possibly su-
perior alternative to more classical risk measures. Robust coefficients of conditional asymmetry
are bounded between -1 and 1 and revolve around the median and indicate negative (positive)
asymmetry whenever they are below (above) 0, while being different than the third moment
of returns. As emphasized in Ghysels et al. (2016), these asymmetry indicators are robust to
outliers and can be computed at various time horizons. Differently from Ghysels et al. (2016),
our measures are short-term option-implied coefficients of asymmetry, instead of long-term
stock-based coefficients of asymmetry. As such, our measures are based upon forward-looking
information, rather than constructing the forecast on historical return data. In terms of the
time-horizon, it is worth noticing that while the analysis in this paper has a short-term focus
to study the role of weekly options, the proposed estimator can be applied, with no loss of
generality, to any forecasting horizons.2 Finally, the robust coefficients of conditional flatness
are computed following the original paper of Ruppert (1987) and using the estimated option-
implied quantiles and expectiles. Robust quantile and expectile kurtosis indicators are defined
as the ratio between two ranges of the respective distribution, and measure how the tails extend.
Once more, while the usual approach to estimate the indicator proposed by Ruppert (1987) is
to use historical stock market data, here we use option-market data.3 All option-implied robust
quantities are then used for different in- and out-of-sample forecasting exercises at different
time horizons and to predict different quantities.
In terms of out-of-sample forecasting power Metaxoglou and Smith (2017) show that their
monthly State Prices of Conditional Quantiles (SPOQ) have a significative forecasting power,
above all at long time horizons (18 to 24 months). Similarly also Conrad et al. (2012) infer ex
ante higher moments of the underlying individual securities’ RND and find that they strongly
related to future returns. Differently from Metaxoglou and Smith (2017) and Conrad et al.
(2012), we focus at shorter time horizons and develop a robust econometric approach based
upon an arbitrage-free volatility smile to infer the option-implied measures. The choice of
2
Indeed, this flexibility applies also to Ghysels et al. (2016).
3
As far as we know, only Ammann and Feser (2019) propose a similar indicator departing from option market
data.

Electronic copy available at: https://ssrn.com/abstract=3622433

shortening the time horizon and using weekly options is of key importance for the paper. First
and foremost, as aforementioned, forecasting returns at short time horizons is of great challenge.
Secondly, while it is reasonable to assume that the option-implied information in its entirety
might not be fully mapped into the physical measure, this bias is a decreasing function of the
time horizon in consideration.4 The use of short-term options thus allows us to forecast returns
at short time horizon, reducing the mapping problem as much as possible. For this purpose, we
use weekly options written on the S&P 500. Weekly options are comparatively new financial
assets that change the classical paradigm existing in the option literature of discarding short
term options often due to the lack of liquidity. The liquidity of short term options has increased
remarkably in the recent years and, as explored in Section 5, accounts now for around the 40
to 50% of the entire total liquidity of the S&P 500 Index options. This allows us to perform
for the first time a short-term analysis of the forecasting power of option-implied quantiles and
expectiles. Having said that, it is worth stressing again how all the econometric models and
tests proposed in this paper can be extended to any time horizon and with no loss of generality.

The remainder of the paper is organized as follows. Section 2 describes the econometric ap-
proach used to estimate the arbitrage-free volatility simile, which will be then used as a starting
point for all our subsequent estimations. In Section 3 we test the validity of the proposed econo-
metric model by comparing the BIRS approach with the PCA and FS models. Section 4 reviews
the statistical properties of quantiles and expectiles and presents the econometric approach used
to estimate the option-implied quantile and expectile curves and the relative indicators that we
will use for the different forecasting exercises. Section 5 describes the dataset used in the paper
with a particular focus on the role of weekly options for the analysis. Section 6 presents the in-
and out-of-sample predictability results of our option-implied measures. Section 7 concludes.

2 Robust and arbitrage-free implied volatility smile - the BIRS

model
In this section we describe how we fit the arbitrage-free implied volatility smile that will then
be used as the input for the estimation of all option-implied based quantities of the paper.
4
The pricing kernel, defined as the ratio of the risk-neutral over the physical measure, converges to 1 as the time
to maturity decreases:
St = ert T q · 1 = eRt T p · 1 (1)
where St is the current expected payoff of a primitive contingent claim under the risk-neutral measure, q, and
under the physical measure, p, respectively. In a complete and arbitrage-free economy the risk-neutral price
grows at the current risk-free rate, rt , whereas the physical price grows at the current risk-adjusted risk-free
rate, Rt . It goes by consequence that as the time to maturity T approaches zero, both quantities converge to
the same state price density, St ; conversely, when T increases their divergence increases at the rate (R − r)T .

Electronic copy available at: https://ssrn.com/abstract=3622433

After cleaning the dataset (Section 2.1), the estimation approach is made of two main parts,
among them linked one to another. Specifically, we apply a quadratic program for estimating
the linear and quadratic terms of piecewise polynomials (Section 2.3), describing the price of
call options as a function of their strikes. This method requires a dense pre-estimate . The
pre-estimate is obtained through a cubic spline interpolation on the implied volatility - delta
space (Section 2.2). The cubic interpolation requires a set of unique strike prices, and the
quality of the interpolation is determined by the starting point; we achieve an economically
sound starting point by cleaning the dataset following the option literature.

2.1 Data cleaning

In this paper we work with so called End-of-week (EOW) options, which are weekly options
with Friday expiration and seven days to maturity (more details in Section 5). As it is common
in the option literature, we clean the dataset to remove stale and/or irrational option prices,
and we only work with the most liquid assets. To avoid stale prices, we remove all options with
zero volume and zero open interest. To avoid possibly irrational prices, we discard all options
with a zero bid price and take the mid-prices, defined as the average of the bid and ask prices
as our option price in order to compute the specific implied volatility of the option. Still to
avoid possibly mispriced options, we discard all options with a very high implied volatility (>
100%). Finally, to work with the most liquid assets, we only select OTM options and we convert
OTM puts into ITM call options via the put-call party. Table 5 and table 6 provide summary
statistics of the data filter and an overview of the available number of options after the cleaning.
This cleaned dataset is then used to estimate the volatility smile.

2.2 Cubic smoothing spline

Once the raw data are cleaned, we interpolate to obtain a pre-estimate of the volatility surface.
For this, we follow and modify the approach of Bliss and Panigirtzoglou (2002) which themselves
combine and extend the approaches of Shimko (1993), Malz (1997a), Malz (1997b) and Campa
et al. (1998). Specifically, we interpolate the implied volatility curve over the implied volatil-
ity/delta space through a cubic smoothing spline, where the smoothing spline is the function
f solving:
n
X Z
minθ λ 2
wi (σi − σ̂i (δi , θ)) + (1 − λ) f 00 (θ)2 dt, (2)
i=1

where θ is a set of parameters, λ a smoothing parameter, wi the weights of the spline and σi are
the implied volatility data. Differently from Bliss and Panigirtzoglou (2002), we optimally set

Electronic copy available at: https://ssrn.com/abstract=3622433

the smoothing parameter λ at each iteration and we fix wi = 1.5 The role of λ is to determine the
goodness-of-fit of the fitted spline and its smoothness trade-off, where smoothness is determined
by the integrated squared second derivative of the implied volatility function. For λ = 0, the
function is the variational, or natural, cubic spline interpolant. For λ = 1, the function is
the least-squares straight-line fit to the data. Instead of fixing λ, we search for the optimal
balance between having a smooth curve and being close to the given data. A favorable range
of values for λ is usually close to 1/(1 + h3 /6), where h is the average spacing of the data
sites. The function chooses a default value for λ within this range. The calculation of the
smoothing spline is the resolution of a linear system whose coefficient matrix has the form
λ · A + (1 − λ) · B, with the matrices A and B depending on the data sites x. The default value
of λ makes λ ∗ Tr(A) = (1 − λ) ∗ Tr(B). The role of wi is to determine how much weight to
give to the ith option’s squared fitted implied volatility error. While the central part of the
distribution would benefit the most by setting the wi equal to greek letter vega, here we set it
equal to one because more interested equally in all points of the distribution. The choice of using
the option delta as the independent variable comes from Malz (1997a) and guarantees more
stability in the interpolation.6 Figure 1 depicts the fit of the implied volatility smile estimated
using the presented approach. The figure represent 9 days, picked at random over the implied
volatility/strike space in our analysis, where the grey dots are the market prices of the option
and the continuous black line the result of the proposed interpolation.
5
Bliss and Panigirtzoglou (2002) set λ equal to 0.99 or 0.9999 and wi = ν where ν is the Black and Scholes
greek letter defined as the first derivative of the option price with respect to the option volatility. While their
choice was justified to obtain the best fit possible for the option-implied probability density function, here the
focus is on achieving the closest possible empirical fit of the volatility smile.
6
Alternatively the implied volatility function can also be expressed as a function of strike or of moneyness.

Electronic copy available at: https://ssrn.com/abstract=3622433

 2012-10-26 2018-07-06 2011-11-18
0.6 Interpolated 0.6 Interpolated 0.6 Interpolated
Market observed Market observed Market observed
0.4 0.4 0.4

0.2 0.2 0.2

0.0 0.0 0.0

0.7 0.8 0.9 1.0 1.1 0.7 0.8 0.9 1.0 1.1 0.7 0.8 0.9 1.0 1.1
2014-06-20 2012-08-17 2017-09-22
0.6 Interpolated 0.6 Interpolated 0.6 Interpolated
Implied volatility

Market observed Market observed Market observed

0.4 0.4 0.4

0.2 0.2 0.2

0.0 0.0 0.0

0.7 0.8 0.9 1.0 1.1 0.7 0.8 0.9 1.0 1.1 0.7 0.8 0.9 1.0 1.1
2017-03-31 2017-06-30 2019-05-10
0.6 Interpolated 0.6 Interpolated 0.6 Interpolated
Market observed Market observed Market observed
0.4 0.4 0.4

0.2 0.2 0.2

0.0 0.0 0.0

0.7 0.8 0.9 1.0 1.1 0.7 0.8 0.9 1.0 1.1 0.7 0.8 0.9 1.0 1.1
Moneyness K/F

Figure 1: Implied volatility smile: obtained solving Equation 2 with wi = 1 and optimally
choosing λ ∈ [0, 1] at each iteration with the aim of finding the best balance between smoothness
and closeness to the data.

Having generated a set of fixed-maturity implied volatilities across a grid of deltas we use
an option pricing model, e.g.: the Black and Scholes (1973) pricing model, to convert delta and
the implied volatilities into European option prices over the prices/strike space. It is worth
noticing that the delta and price conversion through the Black and Scholes model is just a
convenient choice that does not impose log-normality or does not presume that the pricing
model correctly prices options.

2.3 Quadratic program to remove the remaining arbitrages

The presented pre-smoother might still be contaminated by arbitrages, above all into the tails
of the distribution. For example, given the sample of this paper, 17.2% of the option prices are
still contaminated by arbitrages, which in our case lead to option-implied quantiles that are
not strictly increasing with respect to the strike, as depicted in Figure 2. The details on the
elements depicted in Figure 2 and the analysis on the importance of having no arbitrages in
the tails are presented in Section 4.

Electronic copy available at: https://ssrn.com/abstract=3622433

Figure 2: Option-implied quantiles surface: The figure shows the time series of the option-
implied quantiles estimated as described in Section 4. The figure depicts the time-line on the
x-axis, the option-implied quantiles on the y-axis and the strike prices on the z-axis.
.

As demonstrated by Carr and Madan (2005), it is only when all call spreads, butterfly spreads
and calendar spreads are removed that it is possible to obtain a set of quoted option prices that is
free of any static arbitrages across strike and maturities on a single underlier. To fix any possibly
remaining arbitrage on the interpolated smile, we apply the Fengler (2009) quadratic program
to the cleaned and interpolated smoother.7 As a main advantage the proposed approach does
not change anything that does not have arbitrages, and removes any still possibly contaminated
price presents into the smile. As detailed in Appendix A, at each point t we solve the following
quadratic program:

1
minx − y T x + xT Bx (3)
2
T
subject to A x = 0 (4)

Please note that while the original Fengler (2009) approach is to fit the implied volatility surface,
7
As documented in Fengler (2009) and Green and Silverman (1994), the choice of the initial estimator is flexible
i.e.: any two-dimensional non-parametric smoother such as a local polynomial estimator or a thin plate spline
are valid candidates.

10

Electronic copy available at: https://ssrn.com/abstract=3622433

in this paper we do not deal with the dimension of time, fitting only the implied volatility smile.
More precisely, focusing on short-term options, we fix T = 7 such that the final output of the
quadratic program is applied to the call option price/strike plane. The quadratic program in
Equation 3 is convex, thus solvable within polynomial time (see the quadratic optimization
part of Floudas and Visweswaran (1995) and naturally casts a cubic smoothing spline in it
(Green and Silverman (1994)). The convexity property follows from the the strict positive-
definiteness of B and guarantees the uniqueness of the solution. Moreover, shape constraining
the spline smoother guarantees the optimal rate of convergence in shape-restricted Sobolev
classes (Mammen and Thomas-Agnan (2002)). A snapshot of our estimation procedure is in
Figure 3 that depicts the S&P 500 options time series of weekly implied-volatility smile for the
period 2011-2019.

Figure 3: Implied-volatility smile: for the S&P 500 index weekly options estimated with the
approach presented in Section 2. The figure depicts the strike prices on the x axis, the implied-
volatility on the y axis while the older-to-more-recent data of the time series are depicted with
a lighter-to-darker color.

As a final way to test the robustness of the estimation approach, and its ability of the model
to accommodate different market scenarios, Figure 4 depicts the implied volatility smile for low
and high volatility days of the time series, respectively. Specifically, having the implied-volatility
smile on the y-axis and the strike prices on the x-axis, the top (bottom) panel of Figure 4
represents a day with low (high)8 volatility, which translates in an implied-volatility smirk
(smile).
8
The low (high) volatility day is March 21, 2014 (January 15, 2016) where the VIX index closed at 15% (27.02%).

11

Electronic copy available at: https://ssrn.com/abstract=3622433

 High VIX day

0.6 350
Interpolated Interpolated
0.5 Market observed 300 Market observed
250
0.4
Implied volatility

Option price
200
0.3
150
0.2
100
0.1 50
0.0 0
0.80 0.85 0.90 0.95 1.00 1.05 1.10 0.80 0.85 0.90 0.95 1.00 1.05 1.10
Moneyness K/F Moneyness K/F
Low VIX day
0.6 350
Interpolated Interpolated
0.5 Market observed 300 Market observed
250
0.4
Implied volatility

Option price
200
0.3
150
0.2
100
0.1 50
0.0 0
0.80 0.85 0.90 0.95 1.00 1.05 1.10 0.80 0.85 0.90 0.95 1.00 1.05 1.10
Moneyness K/F Moneyness K/F

Figure 4: Implied-volatility smile: for two single days, both estimated with the approach
presented in Section 2. Upper Panel: depicts the estimated implied volatility and the obtained
option prices for a date in the sample on which the VIX was above its historical level. Lower
panel: illustrates the estimation results for a date when the VIX was below its historical average.
Both panels of the figure depict the moneyness (K/F ) on the x axis.

Finally, converting the estimated implied-volatility smile into prices produces a smooth and
arbitrage-free time series of option prices that can now be used to infer option-implied quantiles
and expectiles without the risk of incurring in completely biased final results.

3 Alternative estimation approaches

In this section we first quickly recall the two estimation approaches that are used to validate
and compare our estimation approach with the existing literature (Section 3.1), namely the
PCA of Bondarenko (2003) and the FS of Jackwerth (2004). Then, we compare the PCA
and the FS models with the BIRS approach proposed in this paper. To do it, we check their
capability of recovering the implied volatility and option market prices using observed real data
(Section 3.2).

12

Electronic copy available at: https://ssrn.com/abstract=3622433

3.1 The PCA and FS approaches

The PCA of Bondarenko (2003) is a non parametric approach that starts from a set of admissible
densities and optimally chooses the one that best fits a given cross-section of empirically
observed option prices. The set of admissible densities excludes all implausible densities that are
economically meaningless (e.g.: discontinuous functions) and is obtained through the convolution
of some arbitrary density function and a fixed kernel.
More formally, let Ld be the set of all possible probability densities, of which L1 (−∞, ∞) denotes
the set of all nonnegative functions that integrate to one. For the convolution, we start by fixing
a kernel function φ(x) ∈ Ld , which can be scaled with a bandwidth h to form a new density
φ(x)h := 1/hφ(x/h). The role of h is to control for the smoothness of the densities, thus playing
a crucial role in the final estimation. Given a fixed φ(x)h we define an approximating set of
all functions g as the set of admissible densities, Wh , which is obtained as the convolution of
φ(x)h and another positive density function u:

Wh := {g ∈ Ld |g = θh ∗ u for some u ∈ Ld } (5)

where g and f are some integrable function such that:

Z ∞
f ∗ g := f (x − y)g(y)dy (6)
−∞

Then, the RND is inferred by selecting a density fˆ(x) from the set of admissible densities that
optimally fits an empirically observed cross section of put option prices pi with a finite set of
strike prices x1 < ... < xn :
n
X
min (pi − D−2 fˆ(x))2 (7)
fˆ∈Wh i=1
Rx Ry
where D−2 g(x) := −∞ ( −∞ )g(z)dz)dy represents the second integral of function g(x).
While exposed in continuous form, the presented optimization problem is solved numerically
through a discretization of the admissible sets:
 
 ∞
X ∞
X 
Wh∆z := g ∈ Ld |g(x) = aj φh (x − zj ) aj > 0 aj = 1 (8)
 
j=−∞ j=−∞

where ∆z is the grid step used to determine an equally spaced grid so that for j = 0, ±, ....
the grid is defined over the real line as zj = ∆zj. It is worth noticing that Wh∆z ⊂ Wh , and
that the distance among the two sets is determined by the grid step ∆z. Given a discretized
admissible set, the discretized convolution is achieved on the equally spaced grid through a

13

Electronic copy available at: https://ssrn.com/abstract=3622433

mixture of Dirac delta functions and the basis density:
∞
X ∞
X
f = φh ∗ u where u(x) = aj δ(x − zj ) aj > 0 aj = 1 (9)
j=−∞ j=−∞

Throughout our analysis we follow Bondarenko (2003) and set the bandwidth h = 0.95h0 , the
grid step ∆z = 0.5h and the the basis density equal to the standard normal distribution:

1 2
θ(x) = n(x) := √ e−x /2 (10)
2π

Finally, as the optimization problem is a minimization with respect to the observed put option
prices, we convert all call prices into put prices via the put-call parity.

The FS of Jackwerth (2004) is the “fast and stable” which is related to the original approach
of Jackwerth and Rubinstein (1996) and Jackwerth (2000).9 The main objective of the “fast
and stable” approach is to find a smooth RND which also explains the observed option prices
by first estimating an optimal smooth implied volatilities curve, and then infer the RND
applying Breeden and Litzenberger (1978). The FS approach recovers the RND in three phases.
First, we collect the Black-Scholes option-implied volatilities σ̄i for all available strike prices
i = 1, . . . , I. To do it, we discretize the price grid Sj = S0 + j∆ for j = 1, . . . , J for ∆ equal
to the difference between two adjacent strike prices, so that the price grid coincides with the
strike grid. Secondly, we minimize the objective function:

J I
X
00 2 (J + 1)λ X
min (σj ) + (σi − σ̄i )2 (11)
σi I∆4
j=0 i=0

where λ is a smoothing parameters that determines the trade off between smoothness and
goodness of fit, σi and σ̄i are the implied-volatility and the observed implied-volatility associated
with strike price i = 1, . . . , I, respectively. The same holds for σj for j = 1, . . . , J. The second
derivative σj00 is numerically approximated by σj00 = (σj−1 − 2σj + σj+1 )/∆2 , where delta is
the difference between two adjacent strike prices. Variables σj , σi are selected so that the
curvature of the volatility curve is minimized and the estimated volatility curve agrees with the
observed volatilities. The role of the penalty is to regulate the trade-off between the smallest
squared second-order derivative of the implied volatility curve (smoothness), and the minimum
of the sum of the squared errors, which is the distance between the estimated and the implied
9
While there is a substantial change from the approach in Jackwerth and Rubinstein (1996) to the subsequent
version in Jackwerth (2004), the only “difference” between Jackwerth (2004) and Jackwerth (2020) is a rewriting
of the scaling factor in the fit-smoothness objective function used for the optimization.

14

Electronic copy available at: https://ssrn.com/abstract=3622433

volatilities (fit). Finally, after solving for the optimal volatility function, these quantities are
used to compute the Black and Scholes option prices and infer the RND by differentiating
twice the obtained option prices with respect to the strike prices. While we fix ∆ = 2.5, we
follow Jackwerth (2004) and calibrate λ such that the inferred RND are positive and being
smooth. Results are summarized in table 1.

mean std min 25% 50% 75% max

Trade-off term 0.000249 0.000390 0.000000 0.000008 0.000032 0.000327 0.001491

Table 1: Summary statistic trade-off term: The table provides summary statistics on the
trade-off term (J + 1)λ/(I∆4 ) as in Jackwerth (2004).

3.2 Real data experiment

As a first test we use observed real market data from a cross section of option prices to
empirically validate the BIRS approach. Formally, we check its capability of recovering the
implied volatility and option market prices using observed real data, and we compare it with
the PCA and FS approaches. For each approach, we follow the literature and compute two
common loss measures used to measure the accuracy of the estimates namely, the root mean
squared error (RMSE) and the mean absolute error (MAE):

N
" #1/2
1 X 2
RMSE = yi − ŷi
N
i=1
N
1 X
MAE = yi − ŷi
N
i=1

where i = 1, . . . , N denotes the total number of observations, yi the observed values, and ŷi the
interpolated ones.
In clockwise order, the four panels of Figure 5 depicts the entire time series of the RMSE price,
RMSE implied volatilities, MAE price, and MAE implied volatilities for the three approaches
presented, respectively. Placing on the x-axis the timeline of the analysis, and on the y-axis
the value of the loss function, the figures plot with a black continuous line the our approach
(BIRS), in gray the FS approach and in light gray the PCA approach.

15

Electronic copy available at: https://ssrn.com/abstract=3622433

 Methods
BIRS
Jackwerth (2004)
Bondarenko (2003)

RMSE Price RMSE IV

4 0.150
0.125
3
0.100
2 0.075
0.050
1
0.025
0 0.000
2011 2012 2013 2014 2015 2016 2017 2018 2019 2011 2012 2013 2014 2015 2016 2017 2018 2019

MAE Price MAE IV

4 0.150
0.125
3
0.100
2 0.075
0.050
1
0.025
0 0.000
2011 2012 2013 2014 2015 2016 2017 2018 2019 2011 2012 2013 2014 2015 2016 2017 2018 2019

Figure 5: Time series RMSE and MAE: Time series plots of the calculated RMSE and
MAE for the different methods to estimate risk-neutral densities (RNDs). The left hand side of
the graphs display the mean dollar price deviations on a daily base within the sample period.
The right hand side of the graphs display the respective deviations in terms of implied volatility.

From the figure it clearly emerges that for the entire time series on consideration and for
both prices and implied volatilities, the BIRS approach produces much lower loss measures,
once compared with the FS and PCA approaches.
While Figure 5 depicts graphically the entire time series of RMSE and MAE, Table 2 summarizes
the above findings, reporting different summary statistics of the RMSE and MAE metrics,
respectively.

16

Electronic copy available at: https://ssrn.com/abstract=3622433

 BIRS Jackwerth (2004) Bondarenko (2003)
Prices IV Prices IV Prices IV

RMSE
mean 0.0400 0.0019 0.1834 0.0056 0.5588 0.0211
std 0.0219 0.0006 0.1329 0.0034 0.4670 0.0139
min 0.0167 0.0008 0.0247 0.0012 0.0997 0.0034
25% 0.0275 0.0015 0.0923 0.0032 0.2757 0.0118
50% 0.0336 0.0017 0.1459 0.0046 0.4174 0.0172
75% 0.0464 0.0020 0.2349 0.0071 0.6499 0.0255
max 0.2295 0.0063 0.9407 0.0302 3.6041 0.0929

MAE
mean 0.0925 0.0028 0.3803 0.0082 0.7946 0.0254
std 0.0659 0.0012 0.2885 0.0048 0.5717 0.0168
min 0.0258 0.0012 0.0392 0.0017 0.1614 0.0040
25% 0.0522 0.0022 0.1786 0.0046 0.4321 0.0144
50% 0.0731 0.0025 0.2933 0.0071 0.6406 0.0204
75% 0.1131 0.0031 0.5285 0.0108 0.9493 0.0297
max 0.6244 0.0118 2.0561 0.0397 4.3781 0.1236

Table 2: RMSE and MAE: The table summarizes the root mean squared error (RMSE) and
the mean absolute error (MAE). The RMSE and MAE are calculate on a daily basis. Both
measures are calculated based on observed and interpolated option prices (Prices) and the
corresponding Black-Scholes implied volatility (IV). For both the RMSE and the MAE we
report the arithmetic mean (mean), standard deviation (std), minimum (min) and maximum
(max) values and three quartiles.

Once more, the BIRS approach presented in this paper clearly outperforms both the PCA and
the FS approaches. While the outperformance can be summarized by the lower average values
among the three approaches, the outperformance clearly holds up to the higher quartiles and
the maximum values. Focusing on the RMSE10 , the BIRS approach has a maximum value of
0.2295 (0.0063) for the prices (IV), which is lower than 0.9407 (0.0302) and above all the 3.6041
(0.0929) of the FS and PCA, respectively. For both the MAE and the RMSE the BIRS approach
also display the highest stability in estimation, proxied by the lowest standard deviation.

4 Option-implied quantiles and expectiles

In this section we first quickly recall the concepts of quantiles and expectiles (Section 4.1) and
then we explain how to infer option-implied quantiles (Section 4.2 and option-implied expectiles
(Section 4.3) from traded option prices. While quantiles are well-known statistical quantities
10
Which by construction has values that are always greater or equal to the MAE.

17

Electronic copy available at: https://ssrn.com/abstract=3622433

that have been heavily used for different applications (see, among others, the survey of Koenker
et al. (2017)), expectiles are less known in finance, perhaps due to their lower interpretabil-
ity. Despite being less intuitive than quantiles, expectiles have interesting and possibly even
superior statistical properties than quantiles, once used as risk management tool (Bellini and
Di Bernardino (2017)).

4.1 Quantiles and Expectiles

Given a random variable X on a filtered probability space (Ω, F, P), a quantile qα determines
the value of X such that the probability of the variable being less than or equal to that value
equals α. It follows that for a probability density function f (x) the quantile qα splits the
distribution in two parts, which integrals are of size α and 1 − α. Formally, if X is equipped
with a continuous and strictly monotonic distribution function FX (x) := P (X 6 x), then qα is
the unique solution of the equation:
FX (qα ) = α (12)

Any quantile functions satisfies (see e.g. Föllmer and Schied (2016)):

qα− (X) 6 qα (X) 6 qα+ (X), for each α ∈ (0, 1),

where qα− (X) and qα+ (X) are the left and right quantiles, respectively defined by:

qα− (X) = inf{t ∈ R | FX (t) 6 α}

qα+ (X) = sup{t ∈ R | FX (t) > α}.

Equivalently, the left and right quantiles of the random variable X are defined as the minimizer
of the asymmetric linear loss function:

[qα− (X), qα+ (X)] = argmin E [α(X − x)+ + (1 − α)(X − x)− ] , α ∈ (0, 1).
x∈R

In a financial context, quantiles are often used to calculate the Value-at-Risk (Jorion (2007))
and different finance applications exist linked to quantile regressions (Koenker (2005)).
First introduced by Newey and Powell (1987), expectiles are a one-parameter family of
coherent risk measure defined as the minimizers of an asymmetric quadratic loss function.
Formally, the expectile eθ of a random variable X ∈ L2 (Ω, F, P ) is defined as:

eθ (X) := arg min E θ(X − x)2+ + (1 − θ)(X − x)2− ,

 
θ ∈ (0, 1),
x∈R

18

Electronic copy available at: https://ssrn.com/abstract=3622433

where (.)+ = max (., 0), (.)− = min (., 0). Expectiles combine the concept of “expectation” and
“quantiles” and are the asymmetric generalization of the mean, being that for θ = 1/2, we have
eθ (X) = E(X). Expectiles can also be conveniently defined for any X ∈ L1 (Ω, F, P ) as the
unique solution of the first order condition:

θE[(X − eθ (X))+ ] = (1 − θ)E[(X − eθ (X))− ]. (13)

In a financial context expectiles are related to but still different from, what is generally referred
in risk management as the Conditional Value at Risk (CVaR); as such contain information
about what to expect when the random variable attains a value beyond the quantile (VaR)
(Taylor (2008)).
Statistically, quantiles and expectiles share many similar properties but differ substantially
in one aspect. While quantiles determine the value of X such that the probability of the variable
being less than or equal to that value equals a given level α, expectiles are linked to the
properties of the expectation of the random variable X, conditional on X being into the tail
of the distribution. Moreover, notice that for a given distribution function F 11 , the values of θ
and α are related by the following formula (see e.g. Yao and Tong (1996)):
R qα
αqα − −∞ xdF (x)
θ= R qα
E(X) − 2 −∞ xdF (x) − (1 − 2α)qα

For example, if X ∼ U [−a, a], then qα = 2αa − a, θ = α2 /(2α2 − 2α + 1) and for α =

1%, 5%, 10%, 25%, 50% the values of θ are 0.01%, 0.27%, 1.2%, 10%, 50%.

4.2 Option-Implied Quantiles

From the fundamental theorems of asset pricing, given a dynamically complete and arbitrage-
free finite economy, the value of a European call option corresponds to the present value of the
expected payoff under the risk-neutral measure (see, e.g. Ross (1976)):
Z ∞
−rt (T −t)
Ct,T = e (St − K, 0)+ f (St )dSt (14)
K

where (St − K, 0)+ is the European call option payoff and f (St )dSt the risk-neutral density
function. Given this economy and from the seminal papers of Breeden and Litzenberger (1978)
and Banz and Miller (1978), it is well-known that both, the cumulative distribution and proba-
11
It is worth noticing that this approach does not apply in this paper, being that both quantiles and expectiles
will be inferred non-parametrically from options.

19

Electronic copy available at: https://ssrn.com/abstract=3622433

bility density functions can be inferred from European option prices, taking the first and second
order derivatives of the option price with respect to the strike price, respectively. Formally,
using European put options:
RK
∂Pt,T ∂[e−rt (T −t) 0 (K − St , 0)+ f (St )d(St )]
= (15)
∂K ∂K
Z K
= e−rt (T −t) f (St )dSt (16)
0
= e−rt (T −t) αt,T (17)

With no loss of generality the same holds using European call options:
R∞
∂Ct,T (K) ∂[e−rt (T −t) K (St − K, 0)+ f (St )d(St )]
= (18)
∂K ∂K
Z ∞
= −e−rt (T −t) f (St )dSt (19)
K
= −e−rt (T −t) (1 − αt,T ) (20)

Since the results above are attained by integrating continuous functions, practical application
with discrete strike distances between price observations demand an approximate solution. The
literature offers many approaches to solve this problem; all of them can be categorized under
the classical distinction between parametric and non-parametric approaches. The parametric
approaches rely on a pricing model, which thus has known partial derivatives. For example,
following the Black and Scholes pricing model the above quantities are obtained through N (d2 )
and N (−d2 ), respectively, where N (·) is the cumulative distribution function of a standard
normal distribution function and d2 comes from the Black and Scholes model such that N (d2 )
defines the current risk-neutral probability of the underlying being in-the-money at expiration
N (d2 ) = P(ST > K). Although quick and elegant, the definition of a pricing function inherently
requires an assumption of the underlying price process, and is for this reason neglected in this
paper. For example, imposing the Black and Scholes pricing model would inherently impose log-
normality to our final results. Alternatively, finite differences-based models are non-parametric
approaches that enable results without defining a price-process. Following Barone-Adesi and
Elliott (2007), we propose a finite differences-based approach that is free of any first-order error
caused by the changes of the implied volatility across strike prices, and that eliminates the
first-order error that arises from the Taylor expansion of the derivative. Formally, for three
equidistant European put option prices Pt,T (Ki−1 ) < Pt,T (Ki ) < Pt,T (Ki+1 ) with strike prices

20

Electronic copy available at: https://ssrn.com/abstract=3622433

Ki−1 < Ki < Ki+1 the finite-difference equivalent of Equation 17 is defined as:
    
Put −rt (T −t) Pt,T (Ki+1 ) − Pt,T (Ki ) Pt,T (Ki ) − Pt,T (Ki−1 )
α̂t,T (Ki ) =e λ + (1 − λ) (21)
Ki+1 − Ki Ki − Ki−1

with:
(Ki+1 − Ki−1 )
λ= (22)
(Ki − Ki−1 )
such that λ = 0.5 in presence of equidistant strike prices Ki+1 − Ki = Ki − Ki−1 . With no loss
of generality the same follows for Equation 20 but using European call options. In summary,
this method has been chosen for its parsimony regarding assumptions of the shape of the price
density, despite being more data intensive than most of existing parametric approaches. We
acknowledge that in order to implement Equation 21 for the estimation of a single α̂t,T at the
desired level we require more data points than the parametric Black and Scholes approach
we still believe it cannot be categorized as a numerically intensive approach. More precisely,
the proposed approach works with just three data points and, differently from most of non-
parametric approaches present in the literature, it does not involve any simulations or complex
optimization procedures. Three things also alleviate this issue. First, working with European
options data scarcity is never a problem for our analysis (more detail in Section 5). Second, we
follow the VIX approach and, at each day, we only work with the most liquid options available.
Finally, the approach presented in Section 2 guarantees a dense set of prices which makes the
implementation of Equation 21 straightforward.
Figure 6 depicts the time series of the S&P 500 weekly option-implied quantile curves
estimated with Equation 21:

21

Electronic copy available at: https://ssrn.com/abstract=3622433

Figure 6: Option-implied quantiles surface: The figure shows the time series of the option-
implied quantiles estimated as described in Section 4. The figure depicts the time-line on the
x-axis, the option-implied quantiles/expectiles on the y-axis and the moneyness defined as K/F
on the z-axis.

The figure is in three dimensions, with the weekly value of the option-implied curve on the
y axis and the relative strike prices on the z-axis. The option-implied curve is in green (red)
whenever its value is greater (smaller) than 0.5. The option-implied curve shifts accordingly
with the evolution of the S&P 500. For example, the the overall uptrend of the market is visible,
with some abrupt fall (like in Summer 2018 and the end/beginning part of the 2015/2016).
Moreover, comparing Figure 6 with Figure 2 it is now recognizable the role of the quadratic
program implemented to the interpolated price smoother. In particular, while Figure 6 depicts
a smooth curve with no monotonicity violations, Figure 2 has many irregularities present in
different areas of the curve.
The difference between Figure 6 and Figure 2 thus partially explains from a visual viewpoint
the role of the quadratic program in fixing the option monotonicity. Monotonicity is a natural
property of European options, and its violation is both theoretically and empirically problematic.
It should be of no surprise that option monotonicity follows from the first order derivative of
∂Ct,T (K)
the option price with respect to the strike price: −e−rt (T −t) 6 ∂K 6 0. Such an inequality
has an upper bound that converges to one as T converges to t, and follows from the positivity

22

Electronic copy available at: https://ssrn.com/abstract=3622433

and integrability to one of the risk-neutral density f (St ). Possible arbitrages in the option chain
lead to lack of positivity and/or integrability to one which in turns impact on the cumulative
distribution function, and vice versa.

Dropping one dimension, Figure 7 shows the time series of the option-implied quantiles
with the price evolution of the S&P 500 superimposed. The figure depicts in blue the time
series of the S&P 500 index and the entire daily curve of the option-implied quantiles with all
values above (below) the underlying price in green (red).

Figure 7: Option-implied quantiles: The figure shows the time-series of the option-implied
quantile for the weekly options written on the S&P 500 index for the period 2011-2019 with
the S&P 500 index superimposed (in blue). The figure depicts in green (red) all option-implied
quantile values that are above (below) the daily price of the S&P 500 Index.

From the figure it is possible to appreciate even more how the time series of option-implied
quantiles closely mimics the underlying, both in calm and in turbulent periods. In particular,
during low volatility periods the dispersion of the daily option-implied quantile values is much
smaller compared to more volatile days, characterized by stronger up or down spikes whenever
the market goes up or down. This ability of following the market so closely would not be possible
using historical data and parametric estimation approaches. The use of historical data opens
up to the classical issue of how much data to use. A too big amount of data would not allow
the indicator to be sensitive to market changes, a too short one would be prone to numerical
errors, above all if one wants to estimate the indicator non-parametrically. Working with option
prices it is possible to overcome both problems. The daily option chain is in fact entirely
forward-looking and the nature of the model presented in Section 2 is fully non-parametric and
built to let the data speak as much as possible at each point estimate.

23

Electronic copy available at: https://ssrn.com/abstract=3622433

4.3 Option-Implied Expectiles

Departing from Equation (13) and still assuming a dynamically complete and arbitrage-free
finite economy, the θt,T -expectile is the unique strike price K̄ that solves equation:

θt,T E[(X − eθ (X))+ ] = (1 − θt,T )E[(X − eθ (X))− ] (23)

θt,T Ct,T (K̄) = (1 − θt,T )Pt,T (K̄). (24)

where E[(X − eθ (X))− ] and E[(X − eθ (X))+ ] are the payoff of a European call and put option
with underlying X and strike price eθ (X), respectively. To empirically extract the option-implied
expectile θ̂t,T (Ki ) at a generic level Ki , we reorder the above equation such that:

Pt,T (Ki )
θ̂t,T (Ki ) := . (25)
Pt,T (Ki ) + Ct,T (Ki )

Whenever necessary, and to exploit the higher liquidity of European put (call) options for strike
prices lower (greater) than the underlying, we resort to the put-call parity to infer possible
missing prices for a given strike price. More precisely, if for a given day and strike price only the
price of a European put or call option is available, the other price can be determined through
the put-call parity so that the option-implied expectile is estimated from solely European call
options:
Ct,T (Ki ) − St e−qt (T −t) + Ki e−rt (T −t)
θt,T (Ki ) = (26)
2Ct,T (Ki ) − St e−qt (T −t) + Ki e−rt (T −t)
or solely European put options:

Pt,T (Ki )
θt,T (Ki ) = (27)
2Pt,T (Ki ) + St e−dt (T −t) − Ki e−rt (T −t)

where dt and rt are the continuously compounded dividend and interest rates, respectively.
As for the option-implied quantiles, also this method is fully non-parametric and data driven.
Figure 8 depicts the time series of the S&P 500 weekly option-implied expectile curves estimated
with Equation 25.

24

Electronic copy available at: https://ssrn.com/abstract=3622433

Figure 8: Option-implied expectiles surface: The figure shows the time series of the option-
implied expectiles estimated as described in Section 4. The figure depicts the time-line on the
x-axis, the option-implied quantiles/expectiles on the y-axis and the moneyness defined a K/F
on the z-axis.

As expected, the option-implied expectile curve is both similar to the option-implied quantile
curve and less noisy, above all into the tails of the distribution. This is a direct consequence of the
two different estimation methods, being the one used to estimate option-implied expectiles less
numerically intensive than the one used for quantiles. Figure 9 shows the two-dimensions time
series of the option-implied expectiles with the price evolution of the S&P 500 superimposed,
and confirms both the ability of the time series to track the underlying and the lower amount
of noise, once compared with the option-implied quantile time series.

25

Electronic copy available at: https://ssrn.com/abstract=3622433

Figure 9: Option-implied expectiles: The figure shows the time-series of the option-implied
expectile for the weekly options written on the S&P 500 index for the period 2011-2019 with
the S&P 500 index superimposed (in blue). The figure depicts in green (red) all option-implied
expectile values that are above (below) the daily price of the S&P 500 Index.

4.4 Monte Carlo experiment

Having defined option-implied quantiles and expectiles, we now propose a further validation test
to compare the BIRS approach, with the PCA and FS models presented in Section 3. While in
Section 3.2 we compared the BIRS approach with the PCA and FS ones using real market data,
in this Section we compare it using simulated data through a Monte Carlo experiment. The
rationale behind the experiment is to check the capability of the models to recover a quantity,
knowing the parametric form of the quantity itself. For the experiment we follow Bondarenko
(2003) and test the validity of the proposed pricing models during days with high and low
volatility. A high (low) volatility day in our sample corresponds to a VIX value above (below)
its historical average. To infer the RND we depart from the observed weekly option prices on
the specific date and we calibrate a mixture of three log-normal distributions.

Mean Std Weight

LN1 7.14 0.04 0.33
LN2 7.16 0.02 0.26
LN3 7.17 0.02 0.41

Table 3: High VIX day log-normal parameter estimates: The table summarizes the
results of the fitted mixture of log-normal model with three log-normals. LN1, LN2 and LN3
represent the specific log-normal distribution used. The parameter represent a day in the sample
when the VIX was above its historical average.

26

Electronic copy available at: https://ssrn.com/abstract=3622433

In order to estimate the parameter of the mixtures of log-normal distributions, we rely on the
Nelder-Mead algorithms. To initialize the algorithm we set the starting value for the mean,
µinitial guess , equal to to the natural logarithm of the current spot S&P 500 level on the specific
date. The initial values for the standard deviation are set to σinitial guess = [0.03, 0.05, 0.3].
Based on the estimated parameter, we generate an artificial set of European call and put
options, corresponding to the available strike prices on the specific date. To include possible
option market microstructures, such as a the bid/ask spreads, we test the performance of each
method under two noise specifications. First, as a starting point we follow Rompolis (2010)
and add to the generated option prices a uniformly distributed noise term between -0.025 and
0.025. Secondly, to account for the different liquidity that an option can have depending on its
moneyness (which results in higher bid/ask spreads), we follow Aıt-Sahalia and Duarte (2003)
and add to the generated option prices a noisy term that vary with the option moneyness:

0.5 × (Priceask − Pricebid ) × Liquidity factor,

where:    
2 K
Liquidity factor = 1 + ∗ −1 .
0.2 Ft
where Ft represents the today forward value. Finally, we evaluate each method based on the
root mean integrated squared error (RMISE), and split this error measure up into two parts,
the root integrated squared bias (RISB), referred to as ‘bias‘, and the root integrated variance
(RIV), referred to as ‘variability‘, which allows us to examine the stability and accuracy of each
method separately (Bondarenko (2003)).
s Z 
1 2
RMISE = qR E (f (x) − f (x)) dx
d (28)
f (x)2 dx
sZ
1
RISB = qR (x)] − f (x))2 dx
(E[fd
f (x)2 dx
sZ
1
RIV = qR (E[(fd (x)])2 ]dx
(x) − E[fd
f (x)2 dx

while

RMISE2 = RISB2 + RIV2

Table 4 summarizes the results for October, 28th 2011, when the VIX was above its historical

27

Electronic copy available at: https://ssrn.com/abstract=3622433

average.

BIRS Jackwerth (2004) Bondarenko (2003)

LN 3 QAlpha EAlpha QAlpha EAlpha QAlpha EAlpha

No noise
RMISE 1.1341 0.9806 1.0069 1.1365 1.1590 1.2746 1.2930
Bias 0.8019 0.8061 0.8078 0.8205 0.8225 0.8334 0.8339
Variability 0.8019 0.5585 0.6010 0.7864 0.8165 0.9645 0.9882
Uniform noise
RMISE 1.1341 0.9807 1.0069 1.1363 1.1589 1.2746 1.2930
Bias 0.8019 0.8060 0.8078 0.8205 0.8225 0.8334 0.8339
Variability 0.8019 0.5586 0.6011 0.7860 0.8164 0.9644 0.9882
Moneyness noise
RMISE 1.1341 0.9801 1.0062 1.1357 1.1582 1.2739 1.2923
Bias 0.8019 0.8061 0.8079 0.8205 0.8226 0.8334 0.8340
Variability 0.8019 0.5576 0.5997 0.7851 0.8154 0.9635 0.9871

Table 4: High VIX: The table summarizes the root mean integrated squared error (RMISE).
‘Bias‘ refers to the root integrated squared bias and ‘Variability‘ to the root integrated variance
as calculated in equation 28. The ‘Uniform noise‘ specification refers to the case when we
added uniformely distributed noise terms between [-0.025, 0.025] to the artificial generated
option prices as in Rompolis (2010). The results for the category ‘Moneyness noise‘ specify the
case where the noise term for deeper out-of-the money options increases as in Aıt-Sahalia and
Duarte (2003).

Examining the RMISE across the different methods, we infer that the BIRS approach proposed
in this paper has a lower RMISE, once compared with the PCA and the FS, respectively, both
the uniform and the moneyness noises. It is noteworthy to see that increasing the complexity
of the noise specification does not necessarily lead to a higher RMISE for the BIRS.12

4.5 Constructing option-implied quantile and expectile based measures

As a follow-up to Figures 6, 8, 7 and 9, we further investigate if the estimated option-implied

quantiles or option-implied expectiles have some forecasting power.
For the regression analysis, the vectors of option-implied quantiles and expectiles for each
weekly option chain are compressed to descriptive statistics that highlight the attributes of the
price densities. These statistics are the option-implied - interquantile difference, interexpectile
difference, robust coefficient of asymmetry and robust kurtosis, all of which are based upon
12
Low VIX days confirm the superiority of the model, once compared with the PCA and FS approaches. Results
are presented in the On-line appendix.

28

Electronic copy available at: https://ssrn.com/abstract=3622433

both option-implied quantiles and expectiles.

The distance between upper and lower quantiles or expectiles gives a robust alternative
to the even moments, standard deviation, skewness and kurtosis. We therefore use the ranges
between the option-implied quantiles or option-implied expectiles of certain orders as a robust
descriptive of the variability of the estimated distributions. Specifically, we define the robust
option-implied interquantile range IQRt,T and the robust option-implied interexpectile range
IERt,T defined for α, θ > 1/2 by:

α 1−α
IQRt,T = q̂t,T − q̂t,T (29)
IERt,T = êθt,T − ê1−θ
t,T . (30)

and we set the confidence intervals α and θ equal equal to 75%, 90%, and 95%, being the
most commonly investigated.13 To bypass the arbitrary choice of how to set the confidence
intervals, we propose the integrated versions of the above indicators which summarize the entire
information for all α and θ:
R1 Z 1
α 1−α
0.5
IQRt,T = (q̂t,T − q̂t,T )dα (31)
0.5
R1 Z 1
1−θ
0.5
IERt,T = (êθt,T − êt,T )dθ. (32)
0.5

Following Hinkley (1975) and Ghysels et al. (2016), we propose a robust alternative to the third
moment. Formally we define the robust option-implied coefficient of conditionally asymmetry
(CA) based upon quantiles (QCA) and expectiles (ECA) as:

α − q 0.5 ] − [q 0.5 − q 1−α ]

[qt,T t,T t,T t,T
QCAt,T = α − q 1−α ]
for α ∈ (0.5, 1) (33)
[qt,T t,T
1−θ
[eθt,T − e0.5 0.5
t,T ] − [et,T − et,T ]
ECAt,T = 1−θ
for θ ∈ (0.5, 1) (34)
[eθt,T − et,T ]

Again, we evaluate the QCA and ECA setting the confidence levels at 0.75, 0.9, and 0.95 where,
for a confidence level equal to 0.75, we retrieve the Bowley (1920) statistic.
Once more, to avoid the arbitrary choice of setting the confidence sets we propose the integrated
13
Being strike prices not in a continuum, each range is estimated using the strike prices closest to each bound.

29

Electronic copy available at: https://ssrn.com/abstract=3622433

version of the QCA and ECA:
R1 α 0.5 ) − (q 0.5 − q 1−α )]dα
R1
0.5 [(qt,T − qt,T t,T t,T
QCAt,T =0.5
R1 α 1−α
for α ∈ (0.5, 1) (35)
0.5 [qt,T − qt,T ]dα
R1 θ 0.5 − q 0.5 − q 1−θ )]dθ
R1
0.5 [(qt,T − qt,T t,T t,T
0.5
ECAt,T = R1 θ 1−θ
for θ ∈ (0.5, 1) (36)
0.5 [qt,T − qt,T ]dθ

As a conditional measure of the flatness we follow Ruppert (1987) and propose a robust measure
of kurtosis based upon quantiles and expectiles:14

α − q 1−α ]
[qt,T t,T
QKurtt,T = ω − q 1−ω ]
for α ∈ (0.75, 1) and ω = 0.7 (37)
[qt,T t,T
[eθt,T − e1−θ
t,T ]
EKurtt,T = for θ ∈ (0.75, 1) and ω = 0.7 (38)
[eωt,T − e1−ω
t,T ]

where we select as starting the value of the tail ω = 0.7 and again select 0.75, 0.9 and 0.95
as the confidence interval. Again, and to avoid an arbitrary choice for the confidence, we also
calculate an alternative more generic version that consider the entire distribution:
R1 α 1−α
0.5 [qt,T − qt,T ]dα
R1
QKurtt,T 0.75
= ω − q 1−ω ]
for α ∈ (0.75, 1) and ω = 0.7 (39)
[qt,T t,T
R1 θ 1−θ
0.75 [et,T − et,T ]dθ
R1
0.75
EKurtt,T = 1−ω for θ ∈ (0.75, 1) and ω = 0.7 (40)
[eωt,T − et,T ]

The summary statistics of all calculated option-implied robust measures are collected in Table ??
while the BIRS time series are depicted in Figure 10.15
14
The same robust measure has been proposed also in Ammann and Feser (2019) as a measure for the option-
implied kurtosis. While here we propose it using both quantiles and expectiles, in their paper the authors only
propose one version using quantiles.
15
The time series of the PCA and FS approach are in the Online Appendix.

30

Electronic copy available at: https://ssrn.com/abstract=3622433

 0.30 12
IQR(0.75) 1.0 QCA(0.75) QKurt(0.75)
0.25 IER(0.75) ECA(0.75) 10 EKurt(0.75)
0.5
0.20 8
0.15 0.0 6
0.10 4
0.5
0.05 2
0.00 1.0 0
2011 2012 2013 2014 2015 2016 2017 2018 2019 2011 2012 2013 2014 2015 2016 2017 2018 2019 2011 2012 2013 2014 2015 2016 2017 2018 2019
0.30 12
IQR(0.9) 1.0 QCA(0.9) QKurt(0.9)
0.25 IER(0.9) ECA(0.9) 10 EKurt(0.9)
0.5
0.20 8
0.15 0.0 6
0.10 4
0.5
0.05 2
0.00 1.0 0
2011 2012 2013 2014 2015 2016 2017 2018 2019 2011 2012 2013 2014 2015 2016 2017 2018 2019 2011 2012 2013 2014 2015 2016 2017 2018 2019
0.30 12
IQR(0.95) 1.0 QCA(0.95) QKurt(0.95)
0.25 IER(0.95) ECA(0.95) 10 EKurt(0.95)
0.5
0.20 8
0.15 0.0 6
0.10 4
0.5
0.05 2
0.00 1.0 0
2011 2012 2013 2014 2015 2016 2017 2018 2019 2011 2012 2013 2014 2015 2016 2017 2018 2019 2011 2012 2013 2014 2015 2016 2017 2018 2019
0.30 12
IQR(Integral) 1.0 QCA(Integral) QKurt(Integral)
0.25 IER(Integral) ECA(Integral) 10 EKurt(Integral)
0.5
0.20 8
0.15 0.0 6
0.10 4
0.5
0.05 2
0.00 1.0 0
2011 2012 2013 2014 2015 2016 2017 2018 2019 2011 2012 2013 2014 2015 2016 2017 2018 2019 2011 2012 2013 2014 2015 2016 2017 2018 2019

Figure 10: Time series of the BIRS-based robust option-implied measures: The figures
illustrates the time series of the four robust option-implied measures. The figures depict vertically
the robust option-implied measures of conditional dispersion (left), asymmetry (center), and
flatness (right) and horizontally considering four different confidence intervals, 0.75, 0.9, 0.95
and Integral, respectively.

From both the tables and the figures it emerges the higher numerical stability of the expectile-
based robust measures and of the quantities that are closer to the center of the distribution.
The higher stability of the quantities based upon option-implied expectiles, with respect to
the ones based upon the option-implied quantiles is due to the different estimation approaches.
Option-implied quantiles are in fact estimated, at each point, with three contiguous prices,
while option-implied expectiles only need two points in estimation. Another expected result is
the higher stability of all quantities closer to the center of the distribution, due to the higher
presence of option market data. Indeed, the ones more into the tails of the distribution might
have higher forecasting power, in case of tail risks.
After the estimation of the robust option-implied quantities our subsequent goal of our in-
R1 R1
vestigation is to analyze if these robust indicators (IQR, IER, IQR 0.5 , IER 0.5 QCA, ECA,
R1 R1 R1 R1
QCA 0.5 , ECA 0.5 and finally QKurt, EKurt and QKurt 0.5 and EKurt 0.5 ) have some forecast-
ing power. More generally, we investigate the economic gains from exploiting option-implied

31

Electronic copy available at: https://ssrn.com/abstract=3622433

variability and asymmetries in the distribution of returns at a very short time period. It is
worth noticing that to promote stability and stationarity, the option-implied quantile curve
and the option-implied expectile
 curves are computed as their implied deviations from the for-
S CS
CS = log
ward price: Rt,t+τ t,t+τ CS is the index price at time t + 1 at a confidence
, where St,t+τ
Sf
set (CS) determined by α or τ for quantiles and expectiles, and Sf is the forward price of
the index. For example, to estimate the IQR at a given level α we consider, not the difference
in dollar value of the index at the level α and 1− α, but its return difference at level α and 1− α.

To better evaluate the performances of our option-implied indicators, the same analysis is
also repeated with the same indicators, but this time inferred from historical returns. More
precisely, we estimate again the aforementioned indicators using realized quantiles inferred from
a 30 days rolling window of past historical returns, in order to be used as a benchmark model
in computing out-of-sample R2 s (Campbell and Thompson (2007)).The null hypothesis of this
study is that the forward-looking information of risk-neutral quantities derived from options
can not predict future returns. To test this hypothesis, IQR, IER, QCA, ECA, EKurt, and
A
QKurt, are used as regressors X for predicting rt,t+τ , the τ -days ahead log-return adjusted for
the risk-free interest rate. The analysis is first performed in-sample through a simple linear
A as dependent variable:
regressions fitted to each feature, X, with rt+τ

A
r̂t,t+τ = α + β(X) + εt+τ (41)

For each factor we evaluate its in-sample performance (R2 ), slope β for a short (τ = 7 days),
medium (τ = 60 days) and long (τ = 180 days) time horizon. Moreover, following (Campbell
and Thompson (2007)) we compare the obtained results with a similar exercise using historical
data to compute the measures of asymmetry:
PT
2 t=1 (rt − r̂(t−w,t) )2
ROS =1− PT (42)
t=1 (rt − r̃(t−w,t) )2

where r̂(t−w,t) is estimated with Equation 41 using the forward-looking measures of asymmetry
between t−w and time t, where w is set at 36 weeks, and r̃ is estimated similarly using historical
data.16 The comparison of these two quantities gives rise to a benchmark-adjusted ROS
2 . For

2
positive values of ROOS , the proposed model, r̂(t−w,t) , is performing better than the benchmark
using historical data, r̃, and vice versa for negative values.
Finally, we test if our option-implied indicators have some forecasting power in predicting future
16
To lighten the notation for the out-of-sample analysis, we drop the time orientation t,t+τ . Nevertheless, the
analysis is again performed for the short, medium and long time horizon.

32

Electronic copy available at: https://ssrn.com/abstract=3622433

volatility, skewness, and kurtosis, at short, medium and long-term horizon through a univariate
regression:
c2 t,t+τ = α + β(X) + εt,t+τ
σ (43)

S[rt,t+k ] = α + β(X) + εt,t+τ

b (44)

b t,t+k ] = α + β(X) + εt,t+τ

K[r (45)

To account for potential time-series correlation caused by the overlapping observations in

computing compounded moments, we estimate Newey-West standard errors (Newey and West
(1987)) with lags equal to the number of weeks between observations. Finally we estimate the
same specifications but using a quantile-regression instead of a standard linear model in order
to estimate the median instead of the conditional average.

5 Dataset
The empirical analysis of this paper is based upon data provided by OptionMetrics. We use
the Friday weekly S&P 500 index options daily mid-closing prices (defined as the arithmetic
average of bid and ask closing prices) at 15:59 for the period January 2, 2011 to June 30,
2019. From OptionMetrics we also obtain the daily term-structure of risk-free zero-coupon
interest rates and the S&P 500 daily (continuously compounded) dividend yield. S&P 500
index weekly options are short-term European options written on the S&P 500 index, cash
settled and with a fixed time-to-maturity of seven days (five working days). Weekly options are
listed under the root ticker symbol, “SPXW” and are commonly included in SPX (traditional)
options chains which are AM settled. Officially introduced by the CBOE in October 28, 2005,
weekly options start being actively traded one year later. Due to their increasing adoption from
the finance community, the CBOE launched in the subsequent years weekly options for all
days of the week.17 Nowadays SPXW options constitutes more than 40% of the average daily
volumes of all options traded on the S&P 500 Index, with an average number of daily options
traded that went from below 10,000 per day to above 500,000 per day. One possible reason
that justifies the great interest of market practitioners, above all market markers, on weekly
options is their capability to provide a not expensive (in terms of delta-hedging) and effective
tool to hedge short-term market-wide exposures. In particular, weeklys can provide a hedge to
short-dated tail risks, without loading on the unneeded volatility risk presents in longer-term
17
As a consequence of the high interest from traders, the CBOE now proposes short-term options with even
shorter maturities (up to 2 days options) and written on many other underlyings, like equities (American style),
ETFs, ETNs, VIX and other indexes like the Dow Jones Industrial Average or the Russell 2000 Index. While
increasing, the liquidity of these products is still not high enough to work with non-parametric models. Please
see http://www.cboe.com/products/weeklys-options/available-weeklys.

33

Electronic copy available at: https://ssrn.com/abstract=3622433

options (i.e.: monthly or quarterly). Also, weeklys give traders the possibility to have a more
targeted exposure to events, e.g.: economic releases, earning announcements. In this paper we
use the Friday SPXW, due to their longer history and higher liquidity. Moreover, given that
the liquidity starts being stable from 2011, we drop from the dataset all data before 2011 such
that the sample size of our analysis spans the 2011-2019 period. As reported by the CBOE,
Fridays SPXW are typically listed on Thursday and expire on Friday of the subsequent week18
and, for this reason, are also called the end-of-the-week (EOW) weekly options.19 Applying the
data cleaning approach presented in Section 2 to all the Friday expirations, we end up with
28,609 observations. Tables 5 and 6 collect summary statistics of the variable removed with the
data cleaning and of the final dataset, respectively.
From a forecasting perspective weekly options are interesting financial assets that allow for
forward-looking estimates, still limiting the assumption that all investors are neutral with respect
to the risk. Moreover, from an academic viewpoint, weekly options allow for an interesting
change in the option literature. While in fact short-term options have (correctly) always been
discarded from most of empirical studies because of not being liquid enough, weekly options
are liquid enough to also work with non-parametric approaches.

6 Forecasting exercise
In this section we discuss our main findings related to the predictive power of the robust
option-implied measures presented in Section 4.5 with respect to risk premium, volatility, and
higher realized moments.20 Tables 8 and 9 present the results of regressing the option-implied
measures on realized returns, volatility, skewness and kurtosis. We provide 54 specifications for
estimating each one of the realized moments, varying the probability α in the quantiles and
expectiles as well as the horizon in the prediction. We vary the parameter α between 0.75 and
0.95 to study how the predictive power changes as we approach the tails of the distribution, and
we vary the forecast horizon between one week and 26 weeks to study the short and medium run
18
The same applies to all other weeklies, e.g. Monday SPXW options typically expire on Monday, and Wednesday
SPXW options typically expire on Wednesday.
19
Exception of this rules are if the exchange is closed on a Friday, and/or if the Friday expiration overlaps with
the expiration of monthly or quarterly options. In the former case all options are anticipated and so they expire
the first business day immediately prior to that Friday, in the latter case the Friday expiration of weekly options
is delayed to the next available Friday. In our case, whenever this happens, we substitute the quotation of the
weekly option with the corresponding quotation of the monthly or quarterly option that will expire in the week.
20
For reason of space and to facilitate the reading of the analysis we only present the findings related to the
approach presented in this paper, and compare it with historical-based benchmarks. Nevertheless, it is worth
noticing that the entire empirical analysis has been performed also for the PCA and FS approaches presented
in Section 3 and are available upon request to the authors. Results have been omitted because the PCA and FS
approaches produces among them overall very similar results, but always inferior to the ones generated with the
approach presented in this paper, thus confirming Sections 3.2 and 4.4.

34

Electronic copy available at: https://ssrn.com/abstract=3622433

forecast power. Finally, in Tables 10 and 11 we repeat the analysis using quantile regressions to
estimate the median value of the conditional distribution. We find that the measures IQR and
IER have better predictive power on the realized moments. Moreover, these measures are more
robust across the different specifications. We concentrate our discussion of the results mostly
on these two measures during the remaining of the section.
Table 8 shows that the robust option-implied measures of variability IQR and IER pre-
dict expected returns for horizons up to 26 weeks. Consistent with the literature on return
predictability (e.g. Cochrane (2008)), the magnitude of the point estimates, t-statistics, and
in-sample R2 s increase with the time horizon. Point estimates are statistically significant at
the 10% level for horizons of one week, at the 5% level for horizons of nine weeks, and are
significant at the 1% level for horizons of 26 weeks. Point estimates and in-sample R2 s decrease
as we approach the tails of the distribution, which suggest that the predictive power of the
measures of asymmetry is reduced when we take into account observations in the tails of the
distribution.
Our results are also economically significant. For short horizons, taking the α = 0.75
specification as benchmark, a one standard deviation (0.01) increase in the IQR measure, is
related to an increase of 0.27 % (1.67 %) in average market returns over one (26) week(s).
Using the IER measure, we find that the same increase for an α = 0.75 is related to an increase
of 0.39 % (2.39 %) in expected returns over horizons of one (26) week(s). The specifications
using one and 26 weeks as horizons provide positive out-of-sample R2 s when using historical
measures of asymmetry as a benchmark. For a one week horizon we obtain out-of-sample R2 s
between 0.28 and 1.68 for the IQR measure and between 0.84 and 1.92 for the IER measure. For
horizons of 26 weeks these out-of-sample R2 s vary between 6.81 and 10.52 for the IQR measure
and between 6.62 and 9.4 for the IER measure. Although we obtain in-sample R2 s as large as
5.42 for the regressions using nine week horizons, out-of-sample R2 s are low, suggesting that
forward-looking measures do not necessarily always beat the historical measures for predicting
expected returns. With respect to the option-implied measures QCA and ECA we find that
they can negatively predict expected returns for long horizons, although significant at the 5
and 10 percent level, the economic magnitude is small.
Inference using linear regression models can be misleading if residuals are not spherical. We
repeat our analysis, using quantile regressions in which the right hand side variables provide
variation in the estimated conditional median of the distribution, rather than the conditional
mean. Table 10 shows the results of estimating the median realized returns for the same set of
horizons discussed above. We confirm that the measures of asymmetry IQR and IER predict
median returns up to 26 weeks, and that all results are once more both economically and
statistically significant at the 1% level. Consistent with the results obtained in predicting

35

Electronic copy available at: https://ssrn.com/abstract=3622433

average returns, point estimates, t-statistics, and in sample R2 s increase with the time horizon.
The same one standard deviation increase for an α = 0.75 for IQR is related to an increase in
the median return of 0.36 % (one week), 0.9 % (nine weeks) and 1.61 % (26 weeks). For IER,
this increase is related to an increase of 0.49 % (one week), 1.22 (nine weeks) and 2.26 % (26
weeks) respectively. Moreover, we find that using a quantile specification most out-of-sample
R2 s are positive suggesting that the robust option-implied measures of asymmetry IQR and
IER better predict average and median market returns.
In the second half of Tables 8 and 11 we present the results of using linear and quantile
univariate regressions in predicting realized volatility. We find that the robust measures of
conditional variability IQR and IER positively predict realized volatility for horizons up to
26 weeks, and that these results are economically and statistically significant. Contrary to
how point estimates behave in predicting expected returns, we find that the point estimates,
t-statistics and in sample R2 s in predicting volatility decrease with the time horizon. This result
is consistent with the existence of volatility clusters (Engle (1982); Bollerslev (1986)). We find
that a one standard deviation increase in IQR for an α = 0.75 is associated with an increase
in volatility of 0.33% for a one week horizon, 0.19% for a nine week horizon, and 0.09% for a
26 week horizon. Similarly for the IER measure, these estimations are 0.5 %, 0.29% and 0.14%
equivalently. We find that point estimates decrease as we approach the tails of the distribution
in the specifications, however the in-sample R2 remains in average unchanged. Finally, we find
that for horizons up to 9 weeks, the out-of-sample R2 is large and positive, varying from 12%
up to 43%. This evidence suggests that forward-looking measures of asymmetry perform better
in predicting realized volatility for short horizons.
We repeat the analysis using quantile regressions, in which we estimate the median of the
realized volatility over different horizons. In Table 10 we present similar results as the ones
obtained using linear regressions. The robust measures of conditional variability IQR, and IER
positively predict realized volatility. Point estimates, t-statistics and R2 s decrease over longer
time horizons, suggesting volatility clusters. However, we find that for longer horizons, the
forward-looking measures perform better at estimating the median of the realized volatility
distribution as suggested by the positive and large out-of-sample R2 s in the last column.
We proceed to study the predictive power of the robust forward-looking measures in explain-
ing higher moments. Tables 9 and 11 present the results of predicting the realized skewness
and kurtosis of the market portfolio for short and medium run horizons. We find that for
horizons between nine and 26 weeks the IQR and IER measures positively predict skewness.
These results are statistically significant at the 5% level for a nine week horizon and 1% level
for a 26 week horizon. Both point estimates, t-statistics and in sample R2 s increase with the
horizon used in the regressions. For the benchmark case, and setting α = 0.75, we find that a

36

Electronic copy available at: https://ssrn.com/abstract=3622433

one standard deviation increase in IQR is associated with an increase in the expected realized
skewness of 11.6 (nine weeks), and 16.54 (26 weeks). Equivalently, for the IER measure a one
standard deviation increase is associated with an increase in the expected realized skewness of
16.59 and 24.32 respectively for nine and 26 weeks. We find that out-of-sample R2 s are large
and positive varying between 3% up to 12.8% which suggests that forward-looking measures
of asymmetry perform better in estimating expected skewness compared to their historical
benchmark. We obtain similar results when estimating the median of the realized skewness
of returns. Our results are again also economically and statistically significant for nine and
26 weeks. We find some smaller evidence that the measure of asymmetry EKurt negatively
predict the average of the realized kurtosis as we approach to the tails of the distribution. This
relation is significant at the 1% level. There is some evidence that the quantities QCA and
ECA negatively predict the median of the realized kurtosis of the conditional distribution for
short horizons but the magnitude of the coefficients in the regression shifts sign when looking at
horizons of nine weeks. This suggests a possible median reversion for longer horizons, predicting
a higher probability of extreme events at longer horizons.
Finally, we find that two of the measures calculated, IQR and QKurt negatively predict
the average kurtosis for horizons of nine weeks. These measures also predict the median of the
realized kurtosis for horizons of nine and 26 weeks. Additionally, we find that IER (EKurt)
negatively (positively) predict the median of realized kurtosis. The point estimates that we find
are negative for the IQR and the IER, which become less negative as we approach the tails.

7 Conclusion
Option market data are forward-looking but noisy and highly non-linear financial assets. A
proper empirical analysis of option-implied quantities thus requires solving these two problems
to propose meaningful (arbitrage-free) option-based indicators. In this paper we propose the
BIRS, a novel approach to fit the volatility smile that combines a cubic spline interpolation
with a quadratic program that produces arbitrage-free final estimates, and we use it to estimate
option-implied quantiles and expectiles which allow us to infer conditional robust indicators of
risk and return. To reduce as much as possible the fact that the option-market data impose
risk-neutrality we work with short term (weekly) options and investigate if the option-implied
indicators have predictive information content. The proposed quantities appear to have some
forecasting power - not shared by the equivalent quantities once inferred from backward-looking
historical returns-, both in- and out-of-sample at short, medium and long time horizon and for
both returns and volatility.

37

Electronic copy available at: https://ssrn.com/abstract=3622433

Bibliography
Aıt-Sahalia, Y. and J. Duarte (2003). Nonparametric option pricing under shape restrictions.
Journal of Econometrics 116 (1-2), 9–47.

Ammann, M. and A. Feser (2019). Robust estimation of risk-neutral moments. Journal of

Futures Markets 39 (9), 1137–1166.

Banz, R. W. and M. H. Miller (1978). Prices for state-contingent claims: Some estimates and
applications. The Journal of Business 51 (4), 653–672.

Barone Adesi, G. (2016). VaR and CVaR implied in option prices. Journal of Risk and
Financial Management 9 (1), 2.

Barone-Adesi, G. and R. J. Elliott (2007). Cutting the hedge. Computational Economics 29 (2),
151–158.

Barone-Adesi, G., M. A. Finta, C. Legnazzi, and C. Sala (2019). Wti crude oil option implied
var and cvar: An empirical application. Journal of Forecasting 38 (6), 552–563.

Barone-Adesi, G., N. Fusari, A. Mira, and C. Sala (2020). Option market trading activity and
the estimation of the pricing kernel: A bayesian approach. Journal of Econometrics 216 (2),
430–449.

Barone-Adesi, G., C. Legnazzi, and C. Sala (2019). Option-implied risk measures: An empirical
examination on the S&P 500 index. International Journal of Finance & Economics 24 (4),
1409–1428.

Bellini, F. and E. Di Bernardino (2017). Risk management with expectiles. The European
Journal of Finance 23 (6), 487–506.

Bellini, F., B. Klar, and A. Müller (2016, nov). Expectiles, omega ratios and stochastic ordering.
Methodology and Computing in Applied Probability 20 (3), 855–873.

Bellini, F., L. Mercuri, and E. Rroji (2018). Implicit expectiles and measures of implied
variability. Quantitative Finance 18 (11), 1851–1864.

Bellini, F., E. Rroji, and C. Sala (2021). Implicit quantiles and expectiles. Annals of Operations
Research On-line first.

Black, F. and M. Scholes (1973). The pricing of options and corporate liabilities. Journal of
Political Economy 81 (3), 637–654.

38

Electronic copy available at: https://ssrn.com/abstract=3622433

Bliss, R. and N. Panigirtzoglou (2002). Testing the stability of implied probability density
functions. Journal of Banking and Finance 26, 381–422.

Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of

Econometrics 31 (3), 307–327.

Bondarenko, O. (2003). Estimation of risk-neutral densities using positive convolution approx-

imation. Journal of Econometrics 116 (1), 85–112.

Bowley, A. L. (1920). Elements of statistics. Vol. 2. PS King.

Breeden, D. and R. Litzenberger (1978). Prices of State-Contingent Claims Implicit in Option

Prices. The Journal of Business 51 (4), 621–651.

Campa, J. M., P. K. Chang, and R. L. Reider (1998). Implied exchange rate distributions:
Evidence from OTC option markets. Journal of International Money and Finance 17 (1),
117–160.

Campbell, J. Y. and S. B. Thompson (2007). Predicting excess stock returns out of sample:
Can anything beat the historical average? Review of Financial Studies 21 (4), 1509–1531.

Carr, P. and D. B. Madan (2005). A note on sufficient conditions for no arbitrage. Finance
Research Letters 2 (3), 125–130.

Christoffersen, P., K. Jacobs, and B. Y. Chang (2011). Forecasting with option implied infor-
mation. Handbook of Economic Forecasting.

Cochrane, J. (2008). The dog that did not bark: A defense of return predictability. Review of
Financial Studies 21 (4), 1533–1575.

Conrad, J., R. Dittmar, and E. Ghysels (2012). Ex ante skewness and expected stock returns.
Journal of Finance 68 (1), 85–124.

Engle, R. F. (1982). Autoregressive Conditional Heteroscedasticity with Estimates of the

Variance of United Kingdom Inflation. Econometrica 50 (4), 987–1007.

Fengler, M. R. (2009). Arbitrage-free smoothing of the implied volatility surface. Quantitative

Finance 9 (4), 417–428.

Figlewski, S. (2018). Risk-neutral densities: A review. Annual Review of Financial Eco-

nomics 10, 329–359.

39

Electronic copy available at: https://ssrn.com/abstract=3622433

Floudas, C. A. and V. Visweswaran (1995). Quadratic optimization. In Handbook of Global
Optimization, pp. 217–269. Springer.

Föllmer, H. and A. Schied (2016). Stochastic Finance. An Introduction in Discrete Time

(Fourth ed.). Walter De Gruyter, Berlin.

Ghysels, E., A. Plazzi, and R. Valkanov (2016). Why invest in emerging markets? The role of
conditional return asymmetry. Journal of Finance 71 (5), 2145–2192.

Green, P. J. and B. W. Silverman (1994). Nonparametric Regression and Generalized Lin-

ear Models. Monographs on Statistics and Applied Probability, Volume 58. Chapman and
Hall/London.

Hentschel, L. (2003). Errors in implied volatility estimation. The Journal of Financial and
Quantitative Analysis 38 (4), 779–810.

Hinkley, D. V. (1975). On power transformations to symmetry. Biometrika 62 (1), 101–111.

Jackwerth, J. (2000). Recovering risk aversion from option prices and realized returns. Review
of Financial Studies 13 (2), 433–451.

Jackwerth, J. (2004). Option-implied risk-neutral distributions and risk aversion. Charlottesville,

VA: Research Foundation of AIMR.

Jackwerth, J. (2020). What do index options teach us about covid-19? The Review of Asset
Pricing Studies 10 (4), 618–634.

Jackwerth, J. and M. Rubinstein (1996). Recovering probability distributions from option

prices. Journal of Finance 51 (5), 1611–1631.

Jorion, P. (2007). Value at Risk - The New Benchmark for Managing Financial Risk (3rd ed.).
The McGraw-Hill Companies.

Koenker, R. (2005). Quantile Regression. Cambridge University Press.

Koenker, R., V. Chernozhukov, X. He, and L. Peng (2017). Handbook of Quantile Regression.
Chapman and Hall/CRC.

Linn, M., S. Shive, and T. Shumway (2017). Pricing kernel monotonicity and conditional
information. Review of Financial Studies 31 (2), 493–531.

Malz, A. M. (1997a). Estimating the probability distribution of the future exchange rate from
options prices. Journal of Derivatives 5 (2), 18–36.

40

Electronic copy available at: https://ssrn.com/abstract=3622433

Malz, A. M. (1997b). Option-implied probability distributions and currency excess returns.
Staff Reports no. 32, Federal Reserve Bank of New York .

Mammen, E. and C. Thomas-Agnan (2002). Smoothing splines and shape restrictions. Scandi-
navian Journal of Statistics 26 (2).

Metaxoglou, K. and A. Smith (2017). Forecasting Stock Returns Using Option-Implied State
Prices. Journal of Financial Econometrics 15 (3), 427–473.

Molino, A. and C. Sala (2020). Forecasting value at risk and conditional value at risk using
option market data. Journal of Forecasting 40 (7), 1190–1213.

Newey, W. and J. Powell (1987). Asymmetric least squares estimation and testing. Economet-
rica 55 (4), 819–847.

Newey, W. K. and K. D. West (1987, May). A Simple, Positive Semi-definite, Heteroskedasticity

and Autocorrelation Consistent Covariance Matrix. Econometrica 55 (3), 703–708.

Rompolis, L. S. (2010). Retrieving risk neutral densities from european option prices based on
the principle of maximum entropy. Journal of Empirical Finance 17 (5), 918–937.

Ross, S. (1976). Options and efficiency. The Quarterly Journal of Economics 90, 74–89.

Ruppert, D. (1987). What is kurtosis? an influence function approach. The American Statisti-
cian 41 (1), 1–5.

Shimko, D. (1993). The bounds of probability. Risk 6, 33–37.

Taylor, J. W. (2008). Estimating value at risk and expected shortfall using expectiles. Journal
of Financial Econometrics 6 (2), 231–252.

Yao, Q. and H. Tong (1996, Jan). Asymmetric least squares regression estimation: A nonpara-
metric approach. Journal of Nonparametric Statistics 6 (2-3), 273–292.

41

Electronic copy available at: https://ssrn.com/abstract=3622433

8 Tables

8.1 Dataset description

Data filter Before After in percent

Friday expiration 163246.00 90044.00 44.84
Positive open interest 90044.00 58990.00 34.49
Positive volume 58990.00 42544.00 27.88
Positive Bid price 42544.00 42544.00 0.00
OTM call 42544.00 33291.00 21.75
OTM put 33291.00 28696.00 13.80
Starting in 2011 28696.00 28656.00 0.14
Implied Volatility higher 100% 28656.00 28609.00 0.16

Table 5: Summary statistic data cleaning steps: The table summarizes the data filter
presented in Section 2.1

mean std min 25% 50% 75% max

# Calls available 22.01 8.73 4.00 16.00 20.00 26.00 61.00
# Puts available 61.88 23.72 14.00 40.00 63.00 80.00 123.00
Strike calls 2213.70 536.76 1125.00 1850.00 2175.00 2740.00 3130.00
Strike puts 2025.36 461.91 800.00 1685.00 2020.00 2420.00 2950.00
S&P 500 2195.45 498.15 1123.53 1880.05 2126.41 2670.14 2950.46
Risk-free rate 0.01 0.01 0.00 0.00 0.00 0.02 0.02
Dividend yield 0.02 0.00 0.01 0.02 0.02 0.02 0.03

Table 6: Summary statistics data used for estimation: The table provides an overview
of the data used to estimate option implied quantiles and expectiles between 2011 and 2019.

42

Electronic copy available at: https://ssrn.com/abstract=3622433

 mean std min 25% 50% 75% max

IQR(0.75) 0.02 0.01 0.01 0.01 0.02 0.03 0.08

IQR(0.9) 0.05 0.02 0.02 0.03 0.04 0.05 0.16
IQR(0.95) 0.06 0.03 0.02 0.04 0.05 0.07 0.23
IQR(Integral) 0.01 0.01 0.01 0.01 0.01 0.02 0.05
QCA(0.75) -0.11 0.25 -0.82 -0.27 -0.15 -0.00 0.66
QCA(0.9) -0.21 0.13 -0.52 -0.29 -0.23 -0.14 0.22
QCA(0.95) -0.26 0.11 -0.57 -0.34 -0.26 -0.19 0.04
QCA(Integral) -0.20 0.15 -0.58 -0.29 -0.23 -0.12 0.26
QKurt(0.75) 1.37 0.34 0.67 1.23 1.30 1.41 4.04
QKurt(0.9) 2.80 0.69 1.69 2.46 2.63 2.84 7.23
QKurt(0.95) 3.79 0.96 2.39 3.29 3.55 3.90 9.38
QKurt(Integral) 0.66 0.17 0.43 0.57 0.62 0.68 1.65
IER(0.75) 0.02 0.01 0.01 0.01 0.01 0.02 0.06
IER(0.9) 0.03 0.02 0.01 0.02 0.03 0.04 0.11
IER(0.95) 0.05 0.02 0.02 0.03 0.04 0.05 0.16
IER(Integral) 0.01 0.00 0.00 0.01 0.01 0.01 0.04
ECA(0.75) -0.09 0.08 -0.34 -0.14 -0.10 -0.00 0.14
ECA(0.9) -0.16 0.06 -0.37 -0.20 -0.16 -0.12 -0.00
ECA(0.95) -0.21 0.06 -0.38 -0.24 -0.21 -0.16 -0.05
ECA(Integral) -0.17 0.07 -0.39 -0.21 -0.17 -0.13 -0.00
EKurt(0.75) 1.30 0.08 1.00 1.25 1.30 1.33 1.50
EKurt(0.9) 2.66 0.17 2.17 2.55 2.66 2.77 3.20
EKurt(0.95) 3.67 0.27 3.15 3.50 3.64 3.84 4.42
EKurt(Integral) 0.65 0.05 0.55 0.61 0.64 0.68 0.81

Table 7: Summary statistic BIRS: The table provides summary statistics of the option
implied measures.

43

Electronic copy available at: https://ssrn.com/abstract=3622433

 Table 8: Linear Predictive Model

yt,t+k = a + bxt + εt,t+k

yt,t+k E[rt,t+k ] σ[rt,t+k ]
k 1w 9w 26w 1w 9w 26w
xt α b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos
0.27** 0.86** 1.67*** 0.33*** 0.19*** 0.09***
0.75 2.48 0.28 5.42 0.14 10.32 6.81 50.42 40.67 32.2 15.23 11.57 2.27
(2.05) (2.56) (3.36) (9.82) (6.08) (2.85)
Electronic copy available at: https://ssrn.com/abstract=3622433

0.13* 0.38** 0.86*** 0.17*** 0.1*** 0.05***

IQR 0.9 2.04 1.68 3.92 -1.02 10.1 10.52 51.67 41.08 33.97 13.65 12.79 -0.03
(1.84) (2.04) (3.39) (9.98) (6.44) (2.96)
0.09* 0.29** 0.64*** 0.13*** 0.08*** 0.04***
0.95 1.95 1.49 3.78 -1.47 9.58 7.35 50.74 40.22 33.78 13.57 13.04 1.63
(1.8) (2.07) (3.35) (10.01) (6.71) (3.09)
0.39** 1.12** 2.39*** 0.5*** 0.29*** 0.14***
0.75 2.33 0.84 3.96 -2.56 9.4 6.84 50.49 42.63 32.83 14.21 12.69 1.48
(1.98) (2.11) (3.23) (10.24) (5.98) (2.89)
0.18* 0.54** 1.19*** 0.25*** 0.15*** 0.07***
IER 0.9 1.96 1.92 3.65 -0.93 9.28 9.4 51.75 43.0 33.96 13.43 13.52 -0.14
(1.82) (2.01) (3.25) (10.13) (6.35) (3.02)
0.13* 0.38* 0.86*** 0.19*** 0.11*** 0.06***
0.95 1.85 1.76 3.21 -1.77 8.59 6.62 52.15 42.71 34.42 12.69 14.26 0.89
(1.77) (1.92) (3.06) (10.33) (6.35) (3.03)
0.0 0.01 0.01 0.0 0.0 0.0
44

0.75 0.01 -3.24 -0.15 -9.74 -0.05 0.23 0.0 -1.55 -0.13 -0.89 -0.28 -12.23
(0.88) (0.81) (1.33) (1.02) (0.92) (0.35)
0.01 0.03 0.03 -0.0 0.0 -0.0
QCA 0.9 0.12 -1.08 0.38 -6.48 0.15 1.15 0.1 2.17 -0.29 3.87 -0.15 5.43
(1.06) (1.15) (1.05) (-1.15) (0.15) (-0.81)
0.01 0.03 0.05 -0.01** -0.0 -0.0**
0.95 0.06 -0.35 0.34 -3.48 0.4 0.29 1.11 -3.25 -0.05 0.59 0.99 3.93
(1.0) (1.03) (1.04) (-2.04) (-0.79) (-2.08)
0.01 0.02 0.03 -0.01* -0.0 -0.0
0.75 -0.15 -2.38 -0.11 -9.76 -0.11 0.76 0.44 0.71 -0.24 0.12 0.28 -11.37
(0.91) (0.83) (0.98) (-1.91) (-0.43) (-1.33)
0.02 0.08* 0.13 -0.01** -0.0 -0.01
ECA 0.9 0.11 0.23 1.0 -5.12 1.22 2.28 1.78 3.08 -0.18 4.65 1.22 6.86
(1.37) (1.73) (1.41) (-2.41) (-0.48) (-1.53)
0.02 0.07 0.15* -0.01** -0.0 -0.01*
0.95 0.11 0.05 0.73 -2.78 2.2 2.63 1.37 -2.76 -0.29 0.39 1.49 4.78
(1.32) (1.35) (1.7) (-2.08) (-0.07) (-1.66)
0.0 -0.01 -0.0 -0.0 -0.0 -0.0
0.75 -0.28 1.57 -0.27 0.25 -0.29 2.08 -0.2 0.25 -0.27 0.58 -0.24 0.54
(0.3) (-0.31) (-0.15) (-0.73) (-0.35) (-0.61)
-0.01 -0.01 -0.02 -0.0** -0.0 -0.0
EKurt 0.9 -0.01 -0.75 0.03 -0.33 -0.04 1.72 1.19 5.23 0.12 5.97 -0.28 0.95
(-1.21) (-1.14) (-0.93) (-2.38) (-1.07) (-0.22)
-0.01* -0.02** -0.03 -0.0*** -0.0 0.0
0.95 0.28 -1.25 1.25 -1.7 1.18 -0.5 2.74 7.15 0.77 6.8 -0.29 0.92
(-1.69) (-2.07) (-1.56) (-3.36) (-1.49) (0.14)
-0.0 -0.0 -0.01 0.0 0.0 0.0
0.75 0.27 1.71 -0.24 0.16 -0.14 1.58 -0.16 0.4 -0.29 0.36 -0.29 -1.1
(-1.31) (-0.61) (-0.89) (0.85) (0.17) (0.01)
-0.0* -0.01 -0.01 -0.0 -0.0 -0.0
QKurt 0.9 0.71 0.38 0.4 -0.47 0.03 2.36 -0.29 0.31 -0.2 3.85 -0.29 1.47
(-1.7) (-1.46) (-1.49) (-0.18) (-0.67) (-0.16)
-0.0* -0.0 -0.01** -0.0 -0.0 -0.0
0.95 0.77 0.83 0.49 -1.46 0.36 1.03 -0.06 0.9 0.12 4.01 -0.27 1.44
(-1.74) (-1.52) (-2.06) (-1.04) (-1.18) (-0.29)
 Table 9: Linear Predictive Model

yt,t+k = a + bxt + εt,t+k

yt,t+k S[rt,t+k ] K[rt,t+k ]
k 1w 9w 26w 1w 9w 26w
xt α b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos
-0.52 11.6** 16.54*** -1.32 -20.73** -30.34
0.75 -0.28 -0.0 4.83 4.79 13.47 12.75 -0.13 0.51 2.51 -6.89 2.62 -1.37
(-0.23) (2.33) (3.22) (-0.78) (-2.11) (-1.56)
Electronic copy available at: https://ssrn.com/abstract=3622433

-0.47 5.68** 8.38*** -0.9 -11.96** -15.93

IQR 0.9 -0.25 0.36 4.28 6.33 12.91 11.89 -0.0 0.44 3.18 -3.19 2.7 1.15
(-0.42) (2.16) (3.08) (-1.04) (-2.43) (-1.57)
-0.43 4.22** 6.15*** -0.59 -9.19** -11.76*
0.95 -0.23 0.49 4.07 3.48 11.96 9.97 -0.08 0.18 3.25 -3.84 2.52 -0.2
(-0.51) (2.09) (3.11) (-0.9) (-2.5) (-1.65)
-1.62 16.59** 24.32*** -2.17 -34.79** -46.3
0.75 -0.23 0.13 4.37 4.48 12.97 12.8 -0.09 0.64 3.22 -6.25 2.72 -1.23
(-0.5) (2.23) (3.21) (-0.86) (-2.43) (-1.64)
-1.05 8.03** 11.85*** -1.15 -17.67** -23.54*
IER 0.9 -0.19 0.38 4.05 6.18 12.22 11.23 -0.07 0.56 3.31 -3.25 2.81 1.2
(-0.65) (2.12) (3.12) (-0.92) (-2.49) (-1.7)
-0.89 5.86** 8.68*** -0.89 -13.45** -17.57*
0.95 -0.17 0.52 3.86 3.12 11.76 9.79 -0.05 0.44 3.45 -3.8 2.81 0.21
(-0.75) (2.1) (3.1) (-0.96) (-2.54) (-1.72)
-0.07 0.04 0.17** -0.08 -0.05 -0.31
45

0.75 -0.18 -0.76 -0.27 -2.01 0.28 0.91 -0.06 -0.62 -0.29 -1.25 -0.17 -1.98
(-0.6) (0.34) (2.06) (-0.87) (-0.15) (-0.89)
-0.04 0.15 0.29 -0.14 -0.06 -1.85
QCA 0.9 -0.29 -0.54 -0.2 1.63 0.2 2.58 -0.05 0.21 -0.29 -3.74 0.99 3.76
(-0.16) (0.49) (1.05) (-0.93) (-0.08) (-1.52)
0.04 0.11 0.23 -0.25 0.64 -1.91
0.95 -0.29 1.83 -0.25 -2.57 -0.07 -2.82 0.21 2.46 -0.06 -12.49 0.7 -1.89
(0.14) (0.33) (0.64) (-1.32) (0.8) (-1.36)
0.22 -0.01 0.45* -0.14 2.03** -0.23
0.75 -0.18 -2.25 -0.29 -1.32 0.16 0.98 -0.21 0.07 0.92 0.15 -0.29 -1.65
(0.61) (-0.02) (1.9) (-0.56) (2.13) (-0.28)
0.7 0.53 0.82 -0.53 1.31 -4.31
ECA 0.9 0.31 -0.68 -0.04 1.69 0.5 2.62 0.35 1.45 -0.03 -3.81 1.09 3.76
(1.43) (0.69) (1.18) (-1.52) (0.83) (-1.47)
0.87** 0.91 0.86 -0.52* 1.43 -3.27
0.95 0.76 1.25 0.56 -4.03 0.71 -3.49 0.43 2.85 0.07 -12.21 0.63 -2.7
(1.99) (1.11) (1.03) (-1.65) (0.77) (-0.94)
0.12 -0.23 -0.42* 0.58** 0.57 2.34*
0.75 -0.27 -0.29 -0.22 0.61 0.06 2.06 1.04 0.41 -0.21 0.49 0.41 2.46
(0.31) (-0.51) (-1.69) (2.24) (0.6) (1.81)
-0.06 -0.37* -0.68*** 0.2 0.63 1.81
EKurt 0.9 -0.26 -0.25 0.73 3.3 4.26 1.82 0.49 -0.12 0.21 -1.48 1.76 -15.54
(-0.34) (-1.74) (-2.72) (1.54) (1.28) (1.64)
-0.07 -0.32* -0.56*** 0.11 0.3 1.27
0.95 -0.19 0.62 1.6 0.59 7.34 8.03 0.23 0.03 -0.01 0.12 2.21 -8.38
(-0.6) (-1.81) (-2.76) (1.29) (0.86) (1.54)
-0.02 -0.03 -0.04 -0.06 -0.4*** -0.19
0.75 -0.28 -0.33 -0.27 0.5 -0.22 2.33 0.01 -0.07 0.53 2.06 -0.2 1.82
(-0.22) (-0.39) (-0.65) (-1.02) (-2.64) (-1.12)
-0.02 -0.04 -0.06 -0.04 -0.26*** -0.02
QKurt 0.9 -0.25 0.29 -0.07 2.3 0.39 -3.94 0.23 1.26 1.16 0.84 -0.29 -19.55
(-0.43) (-0.81) (-1.44) (-1.28) (-2.81) (-0.24)
-0.01 -0.04 -0.07* -0.02 -0.17** 0.07
0.95 -0.22 0.96 0.01 -1.33 1.39 -0.64 -0.07 1.13 0.86 2.08 -0.2 -12.61
(-0.52) (-0.84) (-1.7) (-0.84) (-2.3) (0.57)
 Table 10: Quantile Predictive Model

yt,t+k = a + bxt + εt,t+k

yt,t+k E[rt,t+k ] σ[rt,t+k ]
k 1w 9w 26w 1w 9w 26w
xt α b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos
0.36*** 0.9*** 1.61*** 0.31*** 0.2*** 0.08***
0.75 2.63 -0.05 4.47 2.62 4.56 7.28 29.52 41.04 18.28 23.25 5.59 13.67
(4.38) (4.38) (5.47) (19.73) (16.06) (6.32)
Electronic copy available at: https://ssrn.com/abstract=3622433

0.17*** 0.44*** 0.86*** 0.16*** 0.1*** 0.05***

IQR 0.9 2.31 0.36 4.08 3.27 4.33 13.95 29.14 41.17 19.28 19.06 6.63 21.17
(4.0) (4.19) (5.64) (18.65) (15.38) (7.41)
0.13*** 0.32*** 0.63*** 0.12*** 0.07*** 0.04***
0.95 2.16 0.45 3.9 2.16 4.29 9.43 29.02 39.13 19.09 19.11 6.64 22.18
(3.96) (3.94) (5.49) (19.67) (14.03) (7.48)
0.49*** 1.22*** 2.26*** 0.47*** 0.28*** 0.14***
0.75 2.38 0.07 4.12 0.52 4.16 6.91 28.93 43.41 18.63 23.84 6.4 13.01
(3.9) (3.99) (5.18) (19.22) (15.23) (7.62)
0.25*** 0.6*** 1.19*** 0.23*** 0.14*** 0.07***
IER 0.9 2.08 1.6 3.9 4.12 4.19 13.1 29.24 42.6 19.53 19.07 6.96 21.18
(3.87) (3.9) (5.45) (19.21) (15.51) (7.9)
0.19*** 0.45*** 0.87*** 0.17*** 0.1*** 0.06***
0.95 1.96 1.44 3.54 2.12 3.9 8.31 29.59 41.74 20.15 18.99 7.34 21.32
(4.01) (3.94) (5.37) (19.81) (15.7) (8.2)
0.0 0.0 0.02 0.0 0.0 0.0
46

0.75 0.14 -1.98 0.01 -7.93 0.2 -0.19 0.22 0.87 0.14 -1.49 0.09 -4.97
(1.15) (0.22) (1.36) (0.65) (0.8) (0.6)
0.01 -0.0 0.01 -0.0 0.0 -0.0
QCA 0.9 0.29 -1.56 0.0 -2.8 0.04 -0.96 0.23 -0.11 0.03 -3.3 0.02 10.81
(1.27) (-0.02) (0.37) (-1.25) (0.49) (-0.39)
0.01 -0.01 0.02 -0.0* -0.0 -0.0*
0.95 0.41 -0.34 0.05 -4.27 0.11 1.54 0.91 -1.17 0.0 -2.57 0.67 10.98
(1.46) (-0.54) (0.55) (-1.9) (-0.11) (-1.94)
0.01 -0.02 0.02 -0.01** -0.0 -0.0
0.75 0.29 -2.65 0.07 -8.29 0.06 1.33 1.18 3.55 0.03 -0.46 0.02 -1.37
(0.94) (-0.57) (0.56) (-2.17) (-0.26) (-0.31)
0.03* 0.03 0.07 -0.01*** -0.0 -0.0*
ECA 0.9 0.64 -0.63 0.2 -1.36 0.4 1.87 1.74 -0.73 0.01 -2.97 0.55 14.34
(1.72) (0.74) (1.26) (-2.64) (-0.23) (-1.84)
0.02 0.03 0.09 -0.01** -0.0 -0.01**
0.95 0.49 0.66 0.14 -3.45 0.85 5.55 1.49 -0.9 0.01 -0.11 1.17 13.3
(1.13) (0.82) (1.62) (-2.58) (-0.18) (-2.42)
0.0 -0.01 -0.04 -0.0 -0.0 -0.0
0.75 0.0 1.0 0.01 -0.87 0.19 3.39 0.07 -1.14 0.12 -1.24 0.0 1.19
(0.1) (-0.23) (-0.84) (-0.72) (-0.72) (-0.07)
-0.0 -0.01 -0.03 -0.0** -0.0 0.0
EKurt 0.9 0.09 -0.77 0.13 0.42 0.58 1.03 1.45 4.47 0.18 1.3 0.03 0.45
(-0.74) (-0.44) (-1.4) (-2.58) (-1.07) (0.33)
-0.0 -0.02* -0.03** -0.0*** -0.0 0.0
0.95 0.19 -0.46 0.92 -2.32 1.12 0.34 2.88 6.7 0.2 7.19 0.14 0.36
(-0.77) (-1.95) (-2.53) (-3.33) (-0.89) (0.78)
-0.0* 0.0 -0.01 0.0 0.0 0.0
0.75 0.48 1.0 0.02 0.31 0.42 -0.6 0.08 -1.11 0.13 0.57 0.17 -1.9
(-1.67) (0.26) (-1.1) (0.37) (1.21) (0.67)
-0.0** -0.0 -0.01 0.0 0.0 0.0
QKurt 0.9 1.04 0.41 0.03 -1.24 0.54 -1.8 0.01 -0.97 0.01 4.85 0.22 -2.23
(-2.19) (-0.85) (-1.44) (0.1) (0.33) (0.86)
-0.0** -0.0 -0.01 -0.0 -0.0 0.0
0.95 1.19 1.48 0.09 -2.72 0.64 0.52 0.19 -1.5 0.07 4.31 0.32 -0.55
(-2.29) (-0.86) (-1.48) (-1.06) (-1.05) (1.0)
 Table 11: Quantile Predictive Model

yt,t+k = a + bxt + εt,t+k

yt,t+k S[rt,t+k ] K[rt,t+k ]
k 1w 9w 26w 1w 9w 26w
xt α b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos b/t(b) R2 R2oos
-0.41 8.72** 13.64*** -0.52 -17.65** -35.09***
0.75 0.01 -0.74 1.31 1.7 6.88 11.62 0.01 0.92 1.37 -5.65 3.41 -10.69
(-0.13) (2.55) (5.05) (-0.21) (-2.59) (-4.27)
Electronic copy available at: https://ssrn.com/abstract=3622433

-0.62 3.8** 7.07*** -0.68 -12.91*** -17.79***

IQR 0.9 0.03 0.67 0.94 4.88 6.38 6.7 0.04 -0.57 1.84 -2.29 3.37 -2.81
(-0.39) (2.14) (5.0) (-0.52) (-3.76) (-4.16)
-0.7 2.84** 5.59*** -0.19 -8.8*** -13.53***
0.95 0.07 0.83 0.8 0.71 6.0 4.96 0.01 -2.25 2.04 -4.43 3.32 -1.5
(-0.57) (2.1) (5.18) (-0.19) (-3.39) (-4.13)
-1.86 12.08** 20.54*** -1.51 -35.97*** -49.61***
0.75 0.04 -0.3 0.97 2.47 6.7 11.37 0.02 1.16 1.94 -6.02 3.63 -10.57
(-0.4) (2.33) (5.07) (-0.4) (-3.61) (-4.02)
-1.42 5.24** 10.55*** -1.02 -18.25*** -25.79***
IER 0.9 0.1 0.51 0.77 4.6 6.14 6.24 0.04 -0.41 2.01 -2.8 3.74 -3.21
(-0.61) (2.03) (5.12) (-0.54) (-3.66) (-4.16)
-1.04 3.94** 7.47*** -0.81 -13.38*** -18.81***
0.95 0.13 0.73 0.63 -0.11 5.86 5.18 0.06 -0.85 2.12 -4.85 3.84 -1.07
(-0.6) (2.04) (4.87) (-0.58) (-3.6) (-4.08)
-0.23 -0.02 0.11 -0.21* 0.41 -0.96**
47

0.75 0.25 -2.7 0.0 -1.03 0.26 4.07 0.92 0.74 0.2 0.1 1.04 -6.25
(-1.51) (-0.12) (0.8) (-1.79) (1.22) (-2.27)
-0.1 0.32 0.13 -0.35 0.74 -1.35*
QCA 0.9 0.05 -2.94 0.19 2.41 0.03 2.59 0.8 1.33 0.28 -2.17 0.97 0.43
(-0.34) (1.05) (0.47) (-1.59) (1.19) (-1.67)
-0.21 0.42 -0.09 -0.54** 1.26* -1.38
0.95 0.11 -0.61 0.34 -2.51 0.06 4.28 1.24 2.15 0.84 -7.4 0.8 -2.39
(-0.62) (1.2) (-0.28) (-2.1) (1.76) (-1.46)
0.06 0.67 0.42 -0.38 2.49** 1.31
0.75 0.01 -1.06 0.53 -1.58 0.4 3.6 0.41 1.71 1.18 3.33 0.07 -4.78
(0.13) (1.4) (0.98) (-1.05) (2.5) (1.01)
0.1 1.14* -0.1 -0.98* 2.91** -1.98
ECA 0.9 0.01 -1.48 0.8 -0.43 0.02 1.83 0.84 4.05 1.17 -1.15 0.33 0.19
(0.15) (1.73) (-0.16) (-1.96) (2.13) (-1.07)
0.55 1.61*** -0.11 -0.89* 2.49* -2.16
0.95 0.26 0.82 1.59 -7.34 0.01 0.8 0.59 4.88 1.02 -6.02 0.33 -5.21
(0.92) (2.7) (-0.19) (-1.88) (1.96) (-1.28)
0.44 -0.1 -0.24 0.43 -0.61 1.77
0.75 0.18 -0.65 0.02 1.52 0.05 3.88 0.14 0.6 0.07 1.11 0.26 3.31
(0.91) (-0.18) (-0.52) (1.09) (-0.58) (1.25)
-0.02 -0.39* -0.31 -0.03 0.03 1.62**
EKurt 0.9 0.01 0.64 0.69 6.66 1.16 23.76 0.02 1.6 0.0 3.59 1.2 -5.77
(-0.08) (-1.65) (-1.47) (-0.17) (0.06) (2.54)
-0.13 -0.25* -0.3** -0.04 -0.26 1.23***
0.95 0.17 2.94 1.17 0.55 2.01 21.21 0.08 0.28 0.07 0.61 1.46 -3.6
(-0.94) (-1.68) (-2.25) (-0.39) (-0.86) (3.17)
-0.07 -0.04 -0.04 -0.0 -0.3 0.04
0.75 0.22 -1.43 0.09 3.6 0.15 3.8 0.0 -2.41 0.67 3.55 0.02 1.32
(-0.67) (-0.36) (-0.4) (-0.02) (-1.28) (0.14)
-0.06 -0.06 -0.08 -0.03 -0.26** 0.14
QKurt 0.9 0.42 -0.75 0.29 4.81 0.61 23.05 0.06 2.46 1.19 5.09 0.24 -10.35
(-1.17) (-0.97) (-1.6) (-0.75) (-2.26) (0.94)
-0.05 -0.04 -0.07* -0.0 -0.19** 0.17
0.95 0.56 1.54 0.34 -0.06 0.96 17.23 0.0 1.24 1.21 1.83 0.51 -7.89
(-1.3) (-0.95) (-1.93) (-0.02) (-2.27) (1.55)
Appendix

A Appendix: Fengler (2009) quadratic program

Without going into any mathematical detail, in this section we recall the Fengler (2009) quadratic
program used to obtain a smooth and arbitrage-free pricing surface. The interested reader is
referred to Fengler (2009) for further details. We first present the main steps and then repropose
the same procedure in a summarized schematic approach.

For a sample of strikes and European call option prices {ui , yi }, ui ∈ [a, b]:

n
X Z b
minĝ 2
wi [yi − g(ui )] + λ [f 00 (v)]2 dv i = 1, . . . , n (46)
i=1 a

with strictly positive weights wi > 0 and smoothing parameter λ > 0. The minimizer ĝ repre-
sents a globally arbitrage-free European call price function and needs to be twice differentiable
Moreover, assuming yi to be a European call option with strike prices a = u0 , . . . , un = b a func-
tion g is a cubic spline if on each sub-interval (a, u1 ), (u2 , u3 ), . . . , (un , b) is a cubic polynomial
and is twice differentiable C([a, b]) such that:

n
X
g(u) = 1{[ui , ui+1 ]}si (u) (47)
i=0

where ui are the knots of the spline. Green and Silverman (1994) shows that an alternative
and convenient approach to represent the above cubic polynomial is given by the value second
derivative representation, of the natural cubic spline, which allows one to cast the optimization
in 46 as a quadratic problem. Setting gi = g(ui ) and γi = g 00 (ui ) and defining g = (g1 , . . . , gn )T
and γ = (γ2 , . . . , γn−1 )T with γ1 = γn = 0 the natural spline can be completely specified by g
and γ. Not all possible vectors of g and γ give rise to valid solutions, and the sufficient and
necessary conditions for a valid solution are formulated via the matrices Q and R such that

1
minx − y T x + xT Bx (48)
2
T
subject to A x = 0 (49)

where A = (Q, −RT ) and B is strictly positive definite.

48

Electronic copy available at: https://ssrn.com/abstract=3622433

 In summary: At each day t we solve the quadratic program:

1
minx − y T x + x0 Bx (50)
2

subject to the following constraints:

• AT x = 0

• λi > 0 (to guarantee convexity)

RT
g2 −g1
• h1 − h61 γ2 > − e− t rs ds
(to guarantee the price function to be non-increasing in price)

• − gnh−g
n−1
n−1
− hn−1
6 γn−1 > 0 (again to guarantee the price function to be non-increasing in
price)
RT
• g1 6 e− t qs ds
St (to prevent arbitrages)
RT RT
• g1 6 e− t qs ds
St − e − t rs ds
u1 (to prevent arbitrages)

• gn > 0 (to prevent zero or negative prices)

where: x − (g T , γ T )T . To prevent calendar arbitrages the same quadratic program could also
be applied to volatility surfaces by replacing the fifth constrained with:
R tj+1
− qs ds j+1
gij < e tj
gi , for i = 1, . . . , n (51)

for tj with j = 1, . . . , m. For this paper j = 1 and fixed at one week.

49

Electronic copy available at: https://ssrn.com/abstract=3622433