Supplementary Appendix to

“The Pricing of Tail Risk and the Equity Premium:

Evidence from International Option Markets”

Torben G. Andersen∗ Nicola Fusari† Viktor Todorov‡

April 2019

Contents

A Appendix 2
A.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
A.2 Alternative Robust Inference Procedures . . . . . . . . . . . . . . . . . . . . . . . . . 6
A.3 Construction of High-Frequency Measures . . . . . . . . . . . . . . . . . . . . . . . . 6
A.4 The Model-Free Option-Implied VIX Measure . . . . . . . . . . . . . . . . . . . . . . 12
A.5 Sensitivity of the Jump Factor to the Volatility Specification . . . . . . . . . . . . . 12
A.6 Parameter Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
A.7 Predictive Regressions for the Negative Jump Activity . . . . . . . . . . . . . . . . . 21
A.8 Interaction among Factors and Risk Premiums . . . . . . . . . . . . . . . . . . . . . 22
A.9 V IX and the Equity Risk Premium . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
A.10 Parametric vs Nonparametric Tail estimation . . . . . . . . . . . . . . . . . . . . . . 26

∗
Kellogg School, Northwestern University, Evanston, IL 60208, email: [email protected].
†
The Johns Hopkins University Carey Business School, Baltimore, MD 21202, email: [email protected].
‡
Kellogg School, Northwestern University, Evanston, IL 60208, email: [email protected].

1
A Appendix

A.1 Data

We exploit equity-index option data for the U.S. and a number of European indices. These are
supplemented by high-frequency return and futures data for the underlying equity indices.

A.1.1 Equity-Index Option Data

Our sample from the OptionMetrics Ivy DB Global Indices covers January 2007 – December 2014,
except for the Spanish data (IBEX), available only for May 18, 2007 - December 31, 2014. The
database collects historical prices from option markets worldwide. Detailed information regarding
the option contracts and the zero curve for the relevant currencies is also provided.
We obtain data for seven indices: U.S. (SPX), Europe (SX5E), Germany (DAX), Switzerland
(SMI), United Kingdom (UKX), Italy (MIB), Spain (IBEX). For each index, Table 1 provides the
exchange trading hours, which we use to align the observations with the underlying high-frequency
index returns, along with a number of contractual details. Given the novelty of the database, we
devote particular attention to filtering the data. For each contract at any given time, either the last
trade price or the exchange settlement price is reported. While it is impossible to distinguish the
two, the vendor notes that 98% of the data represent settlement prices and only 2% reflect trade
prices, with some variability depending of the specific exchange.
Index Name Country Exchange OptionMetrics Trading Hours Tick Size Multiplier

North America
SP USA CBOE SPX 8:30 am - 3:15 pm 0.05 100 $
Europe
ESTOXX Europe EUREX SX5E 8:50 am - 5:30 pm 0.1 10 e
DAX Germany EUREX DAX 8:50 am - 5:30 pm 0.1 5e
SMI Switzerland EUREX SMI 8:50 am - 5:20 pm 0.1 10 CHF
FTSE UK EURONEXT UKX 8:00 am - 4:30 pm 0.5 10 £
MIB Italy IDEM MIB 9:00 am - 5:40 pm 1.0 1e
IBEX Spain MEFF IBEX 9:00 am - 5:35 pm 1.0 2.5 e

Table 1: Option Contract Specifications. For each option contract we report the underlying
index, the corresponding country, the name of the exchange, the symbol in the OptionMetrics
database, and finally the trading hours, the tick size and the multiplier (as of December 2014).

2
 We create the final sample through the following steps. First, for each option maturity, we
compute the corresponding interest rate by interpolating the zero curve for the given country.
Second, we compute the implied forward price of the underlying index using put-call parity. For
this purpose we retain only cross sections with at least 5 put-call contracts with the same strike
price, and then extract the futures price exploiting the full set of pairs with identical strike prices.
Third, we apply a few filters to ensure that the prices are reliable. We only use options with a tenor
below one year, as longer maturity contracts tend to be illiquid. However, we do include very short-
maturity options in our analysis. This is due to the recent successful introduction of short-dated
options by several exchanges worldwide. These options are particularly informative regarding the
state of the return dynamics, see, e.g., Andersen et al. (2017) for details on the weekly S&P 500
options. Finally, we only retain options whose prices are at least threefold the minimum tick size.

SP500 ESTOXX DAX SMI FTSE MIB IBEX

Number of Options 623 327 477 349 232 164 232
Number of Maturities 11 9 9 6 6 6 6
Q05 Moneyness -6.24 -5.21 -5.34 -5.00 -4.05 -4.46 -2.37
Q95 Moneyness 1.74 1.80 1.80 2.13 1.53 2.02 1.34
Min. Maturity 11 9 9 20 21 15 21
Max. Maturity 329 316 316 316 316 316 316

Table 2: Summary Statistics for the Equity-Index Option Panels.The table reports sum-
mary statistics for the number of option contracts (first row), number of available maturities (second
row), the 5th and 95th moneyness percentile (third and fourth rows), and the shortest and longest
√
available maturities in the last two rows. Moneyness is defined as log(K/F )/(IVAT M τ ), where K
is the strike price, F is the underlying forward price, IVAT M is the at-the-money implied volatility,
and τ is the maturity of the option contract. The numbers represent the median of the daily values.

Table 2 provides summary statistics for the index option markets. The U.S. is the most liquid,
both in terms of the number of observations and available maturities. In Europe, the DAX is the
most active market, while the Italian features the lowest number of observations. Typically, every
market has more than 25 options available per cross section. The moneyness range, measured in
standard deviations from the underlying future price, covers, on average, the interval [−5, 2]. This
enables robust identification of the spot volatility, which loads on at-the-money options (moneyness
near zero), and the negative jump intensity, which loads on deep out-of-the-money put options

3
(moneyness < 0). In terms of tenor, the U.S., German, and Euro indices offer more observations
for short maturity options.

A.1.2 High-Frequency Equity-Index Futures Data

We obtain intra-day observations on futures written on each equity index from TickData. Table 3
reports country, associated exchange, and various contractual features. We stress that the trading
hours are not fully synchronized and are of different duration across exchanges. In particular, the
U.S. trading hours overlap with the European ones by about only two hours. However, we note
that electronic trading takes place outside the stated intervals. For example, the S&P 500 E-mini
futures trade almost 24 hours on the GLOBEX platform, while the table refers to the most active
period, when pit trading is also in progress. The table follows the conventions adopted by TickData.
Index Name Country Exchange TickData Daily Trading Hours Tick Size Multiplier

North America
SP USA CME ES 8:30 a.m. - 3:15 p.m. 0.25 50 $
Europe
ESTOXX Europe EUREX XX 8:00 a.m. -10:00 p.m. 1.0 10 e
DAX Germany EUREX DA 8:00 a.m. -10:00 p.m. 0.5 25 e
SMI Switzerland EUREX SW 8:00 a.m. -10:00 p.m. 1.0 10 CHF
FTSE UK EURONEXT FT 8:00 a.m. - 9:00 p.m. 0.5 10 £
MIB Italy IDEM II 9:00 a.m. - 5:40 p.m. 0.5 10 £
IBEX Spain MEFF IB 9:00 a.m. - 8:00 p.m. 0.5 10 £

Table 3: Futures Contract Specifications. For each index futures contract we report the
country, the option exchange, the TickData symbol for the contract, the daily trading hours, the
tick size, and the multiplier (as of December 2014).

Our (annualized) realized variation (RV) measures are based on five-minute returns, striking a
balance between the number of observations and the extent of market microstructure noise. We
compute the daily RV by the cumulative sum of intraday squared log returns. We further split this
measure into: (1) the truncated variation (TV), capturing the variation due to the diffusive returns,
and (2) the jump variation (JV), reflecting variation stemming from (large) discontinuities. We also
compute the negative and positive jump variation (NJV and PJV), separating JV into the variation
attributed to negative and positive jumps. The measures are obtained following the procedure of
Bollerslev and Todorov (2011), see Section A.3.

4
 √ SP500 ESTOXX DAX SMI FTSE MIB IBEX
√RV 17.57 26.87 24.56 19.32 20.29 22.39 22.30
TV 16.22 23.25 20.91 16.24 18.37 20.28 19.97
JV/RV 0.15 0.25 0.28 0.29 0.18 0.18 0.20
LJV/JV 0.50 0.54 0.55 0.56 0.54 0.56 0.54
Mean Log-Return (%) 6.83 0.79 8.18 3.05 4.42 -5.52 1.41
Std Log-Return (%) 22.30 24.76 23.67 19.31 20.87 27.74 26.51

Table 4: Summary Statistics for the Equity-Index Futures Series. The table reports
summary statistics for the realized return variation (RV), truncated return variation (TV), jump
return variation (JV), negative jump return variation (LJV), and for the daily log-returns of each
index. The numbers are annualized and given in percentage form, except for the ratios in rows
three and four. The various return variation measures are computed from log-returns within the
trading day, using the procedure detailed in the Appendix, then averaged and scaled by 252 to
represent the trading days in one calendar year. We report the square-root of these measures, so
they represent annualized standard deviation units. The daily mean and standard deviation in
rows five and six are computed from close-to-close index return.

Table 4 provides summary statistics for the index return series. The mean returns signify the
discrepancy in performance across the sample period, with DAX futures earning 8.2% and MIB
futures losing 5.5% per year. The RV measures display more parity, but the eurozone currencies
are consistently the most volatile, both in terms of the intraday RV measure (row one) and the
daily close-to-close measures (row six). Finally, we note that the jumps constitute a non-trivial
component of the overall trading day return variation (row three), ranging from around 15% to
close to 30%, with only a marginally larger impact stemming from negative jumps (row four).

5
A.2 Alternative Robust Inference Procedures

This section provides inference for the predictive regressions in Section 4.2.2 based on two alternative
ways to compute the long-run variance of the estimator. First, we obtain t-statistics using standard
errors computed via the Equal-Weighted Cosine estimator with fixed-b critical values proposed
by Lazarus et al. (2018), with degree of freedom equal to 0.4T 2/3 (Equation (4) in Lazarus et al.
(2018)). The critical values corresponding to the one-sided 10%, 5%, and 1%, are 1.72, 2.08, and
2.84, respectively. Second, we compute the standard errors according to the reverse regression
method in Wei and Wright (2011), which is an extension of the estimator developed by Hodrick
(1992).

A.3 Construction of High-Frequency Measures

The high-frequency futures data is obtained from TickData via the TickWrite software. We select
Time Based Bars interval with one-minute granularity holding the last value in case there is no
price change over the one-minute interval. We use the front maturity futures contract and we use
the Auto Roll method provided by the software to roll-over to the next maturity when the front-
maturity contract is near to expiration. We only consider the daily trading hours, combining pit
and electronic trading. We apply the following filters to clean the raw data:

• We keep only observations between Monday and Friday.

• We remove days with no price changes and days corresponding to US holidays.

• We remove days with number of observations less than the average number of daily observa-
tions in the sample. This filter removes half-trading days such as the day before Thanksgiving
or the day before Christmas for the US market.

The cleaned futures data is aggregated to five-minute frequency. For the construction of the high-
frequency measures we introduce the following auxiliary notation. A generic series observed at
high-frequency is denoted with Z (log futures price in our case). For ease of exposition, we ignore
the overnight periods and assume that we have equidistant observations on the grid 0, n1 , n2 , ..... We
1
denote ∆n = n and ∆ni Z = Z i − Z i−1 . With this notation, the construction of the high-frequency
n n

measures is done following these steps:

6
 SP

2 2

t-statistic

t-statistic
0 0

-2 -2
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

ESTOXX

2 2
t-statistic

t-statistic
0 0

-2 -2
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

DAX

2 2
t-statistic

t-statistic
0 0

-2 -2
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

SMI

2 2
t-statistic

t-statistic

0 0

-2 -2
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

FTSE

2 2
t-statistic

t-statistic

0 0

-2 -2
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

MIB

2 2
t-statistic

t-statistic

0 0

-2 -2
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

IBEX

2 2
t-statistic

t-statistic

0 0

-2 -2
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

Figure 1: Predictive Regressions for Excess Returns. Left Panel: t-statistics for the re-
gression slopes according to the reverse regression estimator in Hodrick (1992) as extended by
Wei and Wright (2011); Right Panel:t-statistics for the regression slopes computed using the
Equal-Weighted Cosine estimator of the long-run 7 variance with fixed-b critical values proposed
in Lazarus et al. (2018) with degree of freedom equal to 0.4T 2/3 . The critical value corresponding
to the one-sided 10% is equalt to 1.72.
 SP ESTOXX DAX SMI FTSE MIB IBEX
Panel A: Rt,t+1
V RP -0.35 -1.03 0.69 -0.40 1.29 0.30 -1.21 -1.05 0.64 0.23 0.05 -1.13 0.48 -0.89
V RP − NJ RP -0.70 -0.28 0.69 -1.26 0.27 -0.34 -0.24
NJ RP 0.84 0.31 1.12 1.15 1.01 1.26 0.33 -0.64 0.55 0.62 1.71 1.40 1.71 1.71
R2 0.00 0.01 0.02 0.00 0.01 0.00 0.02 0.02 0.01 0.00 0.03 0.03 0.02 0.02 0.02 0.01 0.00 0.00 0.01 0.01 0.00 0.00 0.04 0.03 0.00 0.00 0.05 0.04

Panel B: Rt,t+5
V RP 1.00 -1.41 0.63 -1.92 1.36 -0.95 -0.81 -2.27 0.36 -0.64 0.58 -1.16 0.45 -1.35
V RP − NJ RP 0.58 -1.48 -0.17 -1.13 -0.53 0.16 -0.42
NJ RP 1.96 1.52 2.74 2.29 2.84 2.70 2.48 0.27 2.21 1.89 2.15 1.87 1.92 1.57
R2 0.02 0.01 0.08 0.07 0.01 0.02 0.18 0.15 0.02 0.00 0.17 0.17 0.04 0.07 0.14 0.01 0.01 0.00 0.12 0.11 0.01 0.00 0.14 0.12 0.01 0.00 0.17 0.13

Panel C: Rt,t+7
V RP 2.08 -0.73 1.56 -0.44 2.01 0.20 -0.23 -1.38 1.31 0.23 1.68 0.03 0.96 -0.49
V RP − NJ RP 1.76 -0.27 1.03 -0.47 0.28 1.32 0.24
NJ RP 2.42 2.26 2.47 2.53 2.63 2.73 2.48 0.63 3.11 2.94 1.81 2.03 1.53 1.43
R2
8

0.06 0.03 0.12 0.12 0.05 0.00 0.19 0.18 0.05 0.01 0.16 0.16 0.00 0.01 0.06 0.02 0.02 0.00 0.12 0.12 0.07 0.04 0.16 0.16 0.03 0.00 0.14 0.14

Table 5: Predictive Regression. For each index, the first column reports the t-statistic and the R2 from the univariate regression
of future returns on V RP , while the following three columns report the identical statistics for related regressions of future returns
on risk premiums. t-statistics are computed using the reverse-regression method developed by Hodrick (1992) and extended
by Wei and Wright (2011). In the second column, the single regressor is the variance risk premium minus the risk premium
associated with negative jumps, the third column refers to a bivariate setting featuring both V RP and N JRP as regressors, and
in the fourth column the single regressor is the negative jump risk premium. Panel A displays the results for the future one-month
returns, Panel B for four-month returns, and Panel C future seven-month returns.
 SP ESTOXX DAX SMI FTSE MIB IBEX
Panel A: Rt,t+1
V RP -0.38 -0.98 1.07 -0.72 1.41 0.52 -3.52 -1.93 0.73 0.33 0.11 -1.55 0.70 -1.44
V RP − NJ RP -0.68 -0.37 0.94 -5.49 0.38 -0.35 -0.36
NJ RP 0.90 0.36 1.52 1.38 1.30 1.44 0.38 -0.64 0.72 0.78 2.44 1.81 2.09 2.16
R2 0.00 0.01 0.02 0.00 0.01 0.00 0.02 0.02 0.01 0.00 0.03 0.03 0.02 0.02 0.02 0.01 0.00 0.00 0.01 0.01 0.00 0.00 0.04 0.03 0.00 0.00 0.05 0.04

Panel B: Rt,t+5
V RP 0.93 -1.45 0.81 -4.20 1.71 -0.86 -1.82 -4.97 0.62 -0.50 0.55 -0.98 0.55 -2.28
V RP − NJ RP 0.56 -1.88 0.11 -3.84 -0.45 0.17 -0.42
NJ RP 2.14 1.42 3.02 2.25 2.97 2.86 2.47 0.43 2.23 1.96 2.70 2.23 2.53 1.98
R2 0.02 0.01 0.08 0.07 0.01 0.02 0.18 0.15 0.02 0.00 0.17 0.17 0.04 0.07 0.14 0.01 0.01 0.00 0.12 0.11 0.01 0.00 0.14 0.12 0.01 0.00 0.17 0.13

Panel C: Rt,t+7
V RP 1.74 -0.81 1.98 -0.78 2.58 0.42 -0.31 -2.70 1.35 0.20 1.70 0.03 1.15 -1.14
V RP − NJ RP 1.34 -0.42 1.36 -1.25 0.27 1.37 0.43
NJ RP 2.62 2.18 2.57 2.67 2.63 2.92 1.92 0.96 2.25 2.38 2.29 2.41 2.15 1.86
R2 0.06 0.03 0.12 0.12 0.05 0.00 0.19 0.18 0.05 0.01 0.16 0.16 0.00 0.01 0.06 0.02 0.02 0.00 0.12 0.12 0.07 0.04 0.16 0.16 0.03 0.00 0.14 0.14
9

Table 6: Predictive Regression. For each index, the first column reports the t-statistic and the R2 from the univariate regression
of future returns on V RP , while the following three columns report the identical statistics for related regressions of future returns
on risk premiums. t-statistics are computed using the Equal-Weighted Cosine estimator of the long-run variance with fixed-b
critical values proposed in Lazarus et al. (2018) with degree of freedom equal to 0.4T 2/3 . The critical values corresponding to the
one-sided 10%, 5%, and 1%, are 1.72, 2.08, and 2.84, respectively. In the second column, the single regressor is the variance risk
premium minus the risk premium associated with negative jumps, the third column refers to a bivariate setting featuring both
V RP and N JRP as regressors, and in the fourth column the single regressor is the negative jump risk premium. Panel A displays
the results for the future one-month returns, Panel B for four-month returns, and Panel C future seven-month returns.
1. Realized Variation:
⌊n(t+τ )⌋
X
RVt,t+τ = (∆ni Z)2 ,
i=⌊nt⌋+1

and if the interval [t, t + τ ] includes an overnight period, the squared overnight return is added
to the summation.

2. Bipower Variation:
⌊n(t+τ )⌋
π X
BVt,t+τ = |∆ni Z||∆ni−1 Z|.
2
i=⌊nt⌋+2

3. Jump Detection:

• For each series, we compute, the so-called Time-of-Day (TOD) function as defined in
Bollerslev and Todorov (2011). We recompute the TOD function each time the exchange
changed the trading hours.

• At each point in time, starting from the second day in the sample, we compute the
threshold level:

3 p
θi = √ RVt−h,t ∧ BVt−h,t × ∆0.49
n × T ODi−⌊i/n⌋ ,
h

where h stands for 24 hours that exclude the overnight period.

• The increment ∆ni Z is flagged as one with jump if |∆ni Z| > θi .

4. Truncated Variation:
⌊n(t+τ )⌋
X
T Vt,t+τ = (∆ni Z)2 1{|∆ni Z|≤θi } .
i=⌊nt⌋+1

5. Positive and Negative Jump Variation:

⌊n(t+τ )⌋ ⌊n(t+τ )⌋
X X
P JVt,t+τ = (∆ni Z)2 1{∆ni Z>θi } , N JVt,t+τ = (∆ni Z)2 1{∆ni Z<−θi } .
i=⌊nt⌋+1 i=⌊nt⌋+1

6. For the computation of the local continuous variation, Vbtn , used to penalize the option-based

10
 volatility estimate in the objective function in (6), we take into account the following:

• Vbtn = h1 T Vt,t+τ , where τ is equal to 3 hours.

• If we encounter more than 4 consecutive zero returns (this could happen in case of
market closure or ”lunch break”) then we extend the window until we reach 36 returns
containing less than 4 consecutive zero returns.

To simplify the notation, RVt , T Vt , P JVt , and N JVt refer to the above quantity computed on a
daily basis (i.e. from the opening to the closing of the market).

We then use the HAR model of Corsi (2009) to forecast the future quadratic variation and
negative jump variation under the statistical measure:1

P
ln(T Vt+1,t+h ) = a + b1 log(T Vt ) + b2 log(T V t−4,t−1 ) + b3 log(T V t−20,t−5 ) + b4 log(T V t−84,t−21 ) + ǫt (1)
P
T Vt,t+h = exp(a + b1 log(T V )t + b2 log(T Vt−4,t−1 ) + b3 log(T Vt−20,t−5 ) + b4 log(T V t−84,t−21 ) + σǫ2 /2),

where σǫ2 is the variance of the residual from the linear regression in Equation (1). Equivalently,
for the negative jump variation we have:

P
ln(N JVt+1,t+h ) = a + b1 log(T Vt ) + b2 log(T Vt−4,t−1 ) + b3 log(T Vt−20,t−5 ) + b4 log(T V t−84,t−21 ) + ǫt (2)
P
N JVt,t+h = exp(a + b1 log(T Vt ) + b2 log(T Vt−4,t−1 ) + b3 log(T Vt−20,t−5 ) + b4 log(T V t−84,t−21 ) + σǫ2 /2),

while for the positive jump variation we have

P
ln(P JVt+1,t+h ) = a + b1 log(T Vt−1 ) + b2 log(T Vt−4,t ) + b3 log(T Vt−20,t−5 ) + b4 log(P JVt−20,t−5 ) + ǫt (3)
P
P JVt+1,t+h = exp(a + b1 log(T Vt ) + b2 log(T Vt−4,t−1 ) + b3 log(T Vt−20,t−5 ) + b4 log(T V t−84,t−21 ) + σǫ2 /2).

Finally, the quadratic variation is obtained as:

P P P P
QVt+1,t+h = T Vt+1,t+h + N JVt+1,t+h + P JVt+1,t+h .

1
When estimating the HAR model we appropriately rescale each realized measure in order to match the uncondi-
tional daily variance level. Specifically we multiply each quantity by yt2 /RV t , where yt2 is the average squared daily
return (close-to-close) and RVt is the average level or the realized variance computed during the trading time.

11
A.4 The Model-Free Option-Implied VIX Measure

The measure is computed following Carr and Wu (2009). On each day, we take the two option cross-sections
with tenor closest to 30 calendar days. For each cross section, we create a fine grid of strike prices K
covering the moneyness range defined as −8 ≤ m ≤ 8 with increments of 0.1. We then interpolate the
implied volatility as a function of K. When K is lower (higher) than the lowest (highest) available strike,
we extrapolate the implied volatility outside the defined moneyness range as a constant equal to the implied
volatility at the lowest (highest) available strike. We obtain the VIX index for both maturities and linearly
interpolate to obtain the VIX index corresponding to 30 calendar days.

A.5 Sensitivity of the Jump Factor to the Volatility Specification

As discussed earlier, our model is parsimonious and, in particular, contains only a single volatility factor.
One potential concern is that the relatively constrained modeling of the volatility process may induce a bias
in the extraction of the implied jump dynamics from the option surface. For the S&P 500 sample, we have
a natural benchmark. Andersen et al. (2015) estimate an extended version of model (1) which provides an
excellent fit to the option prices as well as the volatility series, clearly outperforming more standard and
equally heavily parameterized representations.
The sample exploited here is shorter than in Andersen et al. (2015), due to the synchronization with the
data available for the European markets. On the other hand, it covers a wider cross-section since we include
the full set of short-dated (weekly) options, which were excluded from the analysis in Andersen et al. (2015).
Moreover, the current sample extends through December 2014, while the elaborate model is estimated over
January 1996–July 21, 2010, but with out-of-sample extraction of the jump intensity factor through April
23, 2013. Hence, the two series of option-implied jump factors overlap over the period January 2007 till
April 2013.
Because the left tail factor enters the specification for the jump intensity in distinct ways, a direct
comparison of the factor realizations across the two models is not meaningful. Instead, we focus on the
model-implied variation in the negative jump intensity stemming from the left tail factor. In both cases, the
effect is proportional to the concurrent value of U factor in the model. Consequently, Figure 2 depicts the
standardized jump factor series extracted from the two separate models, estimated from partially overlapping
periods and option samples. The models deliver remarkably similar time series paths for the return variation
induced by the left tail factor, with a correlation of 0.987 during the overlap, corroborating the robust
identification of the jump tail factor from either model.

12
 U Factor
6
5
4
3
Ut

2
1
0
-1
-2
1998 2000 2002 2004 2006 2008 2010 2012 2014

Figure 2: Comparison of Model-Implied Left Jump Tail Factors. The figure depicts the
daily standardized option-implied left jump tail factors from model (1) and the Andersen et al.
(2015) model. The series are extracted based on parameter estimates from weekly SPX option
prices observed over January 2007–December 2014 and January 1996-July 21, 2010, respectively,
with the latter series having been extended in Andersen et al. (2015) to cover jump intensity factors
for the out-of-sample period, July 22, 2010 - April 23, 2013, as well. Both series are normalized to
have mean zero and unit variance.

13
A.6 Parameter Estimates

A.6.1 S&P500

Panel A: Parameter Estimates

Parameter Estimate Std. Parameter Estimate Std.
ρ −1.000 0.012 c+
0 7.551 0.622
v 0.029 0.000 λ− 13.548 0.196
κ 6.216 0.086 λ+ 78.255 2.368
σ 0.605 0.008 µv 22.663 0.462
κu 0.600 0.062 µu 18.656 5.389
Panel B: Summary Statistics
RMSE 1.909
Mean negative jump intensity 0.853
Mean negative jump size −0.074
Mean positive jump size 0.013
Mean diffusive variance 0.038
Mean negative jump variance 0.009
Mean positive jump variance 0.002

Table 7: Estimation Results for the Parametric Model (1) Fit to the S&P 500 Option
Panel and High-Frequency Spot Volatility Measure. Panel A reports the parameter es-
timates obtained using weekly observations on Wednesday, or a neighboring business day in case
of a market closure on Wednesday. Panel B reports summary statistics for the daily series of
model-implied jump and variance estimates. The RMSE is given in annualized BSIV terms, while
the variances and jump intensity is given in annualized decimal units and, finally, the jump sizes
are given in decimal units.

14
A.6.2 ESTOXX

Panel A: Parameter Estimates

Parameter Estimate Std. Parameter Estimate Std.
ρ −0.999 0.014 c+
0 12.879 0.777
v 0.032 0.000 λ− 14.569 0.180
κ 7.471 0.170 λ+ 66.123 1.423
σ 0.687 0.011 µv 13.089 0.263
κu 1.511 0.050 µu 154.224 8.638
Panel B: Summary Statistics
RMSE 1.974
Mean negative jump intensity 1.632
Mean negative jump size −0.069
Mean positive jump size 0.015
Mean diffusive variance 0.054
Mean negative jump variance 0.015
Mean positive jump variance 0.006

Table 8: Estimation Results for the Parametric Model (1) Fit to the ESTOXX Option
Panel and High-Frequency Spot Volatility Measure. Panel A reports the parameter es-
timates obtained using weekly observations on Wednesday, or a neighboring business day in case
of a market closure on Wednesday. Panel B reports summary statistics for the daily series of
model-implied jump and variance estimates. The RMSE is given in annualized BSIV terms, while
the variances and jump intensity is given in annualized decimal units and, finally, the jump sizes
are given in decimal units.

15
A.6.3 DAX

Panel A: Parameter Estimates

Parameter Estimate Std. Parameter Estimate Std.
ρ −1.000 0.012 c+
0 9.666 0.471
v 0.024 0.000 λ− 18.967 0.189
κ 11.653 0.146 λ+ 68.833 1.148
σ 0.741 0.010 µv 32.642 0.690
κu 0.262 0.033 µu 44.814 6.474
Panel B: Summary Statistics
RMSE 2.050
Mean negative jump intensity 1.791
Mean negative jump size −0.053
Mean positive jump size 0.015
Mean diffusive variance 0.055
Mean negative jump variance 0.010
Mean positive jump variance 0.004

Table 9: Estimation Results for the Parametric Model (1) Fit to the DAX Option Panel
and High-Frequency Spot Volatility Measure. Panel A reports the parameter estimates
obtained using weekly observations on Wednesday, or a neighboring business day in case of a
market closure on Wednesday. Panel B reports summary statistics for the daily series of model-
implied jump and variance estimates. The RMSE is given in annualized BSIV terms, while the
variances and jump intensity is given in annualized decimal units and, finally, the jump sizes are
given in decimal units.

16
A.6.4 SMI

Panel A: Parameter Estimates

Parameter Estimate Std. Parameter Estimate Std.
ρ −0.752 0.010 c+
0 4.397 0.445
v 0.020 0.000 λ− 24.316 0.425
κ 12.394 0.319 λ+ 66.762 2.024
σ 0.712 0.013 µv 27.147 0.579
κu 1.724 0.090 µu 415.332 41.169
Panel B: Summary Statistics
RMSE 1.835
Mean negative jump intensity 2.649
Mean negative jump size −0.041
Mean positive jump size 0.015
Mean diffusive variance 0.032
Mean negative jump variance 0.009
Mean positive jump variance 0.002

Table 10: Estimation Results for the Parametric Model (1) Fit to the SMI Option Panel
and High-Frequency Spot Volatility Measure. Panel A reports the parameter estimates
obtained using weekly observations on Wednesday, or a neighboring business day in case of a
market closure on Wednesday. Panel B reports summary statistics for the daily series of model-
implied jump and variance estimates. The RMSE is given in annualized BSIV terms, while the
variances and jump intensity is given in annualized decimal units and, finally, the jump sizes are
given in decimal units.

17
A.6.5 FTSE

Panel A: Parameter Estimates

Parameter Estimate Std. Parameter Estimate Std.
ρ −1.000 0.014 c+
0 5.099 0.979
v 0.028 0.000 λ− 12.688 0.389
κ 5.381 0.138 λ+ 66.693 5.117
σ 0.546 0.008 µv 30.845 1.347
κu 1.191 0.117 µu 25.114 15.159
Panel B: Summary Statistics
RMSE 2.164
Mean negative jump intensity 0.604
Mean negative jump size −0.079
Mean positive jump size 0.015
Mean diffusive variance 0.040
Mean negative jump variance 0.008
Mean positive jump variance 0.002

Table 11: Estimation Results for the Parametric Model (1) Fit to the FTSE Option
Panel and High-Frequency Spot Volatility Measure. Panel A reports the parameter es-
timates obtained using weekly observations on Wednesday, or a neighboring business day in case
of a market closure on Wednesday. Panel B reports summary statistics for the daily series of
model-implied jump and variance estimates. The RMSE is given in annualized BSIV terms, while
the variances and jump intensity is given in annualized decimal units and, finally, the jump sizes
are given in decimal units.

18
A.6.6 MIB

Panel A: Parameter Estimates

Parameter Estimate Std. Parameter Estimate Std.
ρ −0.973 0.038 c+
0 1.564 0.250
v 0.029 0.000 λ− 10.823 0.284
κ 7.111 0.201 λ+ 24.902 1.458
σ 0.501 0.018 µv 24.804 1.026
κu 0.512 0.078 µu 20.066 6.613
Panel B: Summary Statistics
RMSE 2.867
Mean negative jump intensity 0.807
Mean negative jump size −0.092
Mean positive jump size 0.040
Mean diffusive variance 0.063
Mean negative jump variance 0.014
Mean positive jump variance 0.005

Table 12: Estimation Results for the Parametric Model (1) Fit to the MIB Option Panel
and High-Frequency Spot Volatility Measure. Panel A reports the parameter estimates
obtained using weekly observations on Wednesday, or a neighboring business day in case of a
market closure on Wednesday. Panel B reports summary statistics for the daily series of model-
implied jump and variance estimates. The RMSE is given in annualized BSIV terms, while the
variances and jump intensity is given in annualized decimal units and, finally, the jump sizes are
given in decimal units.

19
A.6.7 IBEX

Panel A: Parameter Estimates

Parameter Estimate Std. Parameter Estimate Std.
ρ −1.000 0.009 c+
0 1.500 0.350
v 0.040 0.001 λ− 17.339 1.104
κ 6.530 0.239 λ+ 23.695 2.251
σ 0.683 0.012 µv 6.147 1.835
κu 1.244 0.438 µu 180.477 119.791
Panel B: Summary Statistics
RMSE 2.085
Mean negative jump intensity 3.419
Mean negative jump size −0.058
Mean positive jump size 0.042
Mean diffusive variance 0.060
Mean negative jump variance 0.023
Mean positive jump variance 0.005

Table 13: Estimation Results for the Parametric Model (1) Fit to the IBEX Option
Panel and High-Frequency Spot Volatility Measure. Panel A reports the parameter es-
timates obtained using weekly observations on Wednesday, or a neighboring business day in case
of a market closure on Wednesday. Panel B reports summary statistics for the daily series of
model-implied jump and variance estimates. The RMSE is given in annualized BSIV terms, while
the variances and jump intensity is given in annualized decimal units and, finally, the jump sizes
are given in decimal units.

20
A.7 Predictive Regressions for the Negative Jump Activity
SP
0.3
4

0.2

t-statistic
2

0
0.1

-2
0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

ESTOXX
8
t-statistic 6 0.3

4
0.2
2

0 0.1

-2
0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

DAX
10
0.3
t-statistic

5
0.2

0.1
0

0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

SMI
0.3
8
6
0.2
t-statistic

4
2
0.1
0
-2
0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

FTSE
8
0.4
6
t-statistic

0.3
4

2 0.2

0 0.1
-2
0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

MIB
8 0.3

6
0.2
t-statistic

2
0.1
0

-2
0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

IBEX
0.3
6

4
0.2
t-statistic

0 0.1

-2
0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

Figure 3: Predictive Regressions for Negative Jumps. Left Panel: t-statistics for the re-
gression slopes; Right Panel: Regression R2 , where the full drawn line depicts the total degree of
explained variation and the dashed line signifies the part explained by the spot variance alone.

21
A.8 Interaction among Factors and Risk Premiums

Table 14 reports the sample correlation between various combinations of factors and risk premiums. The
reported correlations may be biased towards zero due to imperfect procedures for forecasting the expected
return variation measures entering the risk premium computations. Moreover, they may be subject to noise
stemming from a few influential outliers. Nonetheless, the general coherence of the findings across the full
set of indices should be informative regarding the underlying relations among these quantities. Specifically,
Table 15 verifies that the correlations for 2007-2009 and 2010-2014 generally are consistent with those for
the full sample period, although the smaller samples naturally do produce more erratic statistics.

V, N JV V RP, V V RP, N JV ⊥ N JRP, V N JRP, N JV ⊥ CV RP, V CV RP, N JV ⊥

SP 0.85 0.86 0.44 0.78 0.74 0.86 0.15

ESTOXX 0.72 0.83 0.41 0.47 0.95 0.86 -0.07
DAX 0.76 0.72 0.42 0.16 0.97 0.88 0.06
SMI 0.83 0.92 0.33 0.64 0.89 0.92 0.12
FTSE 0.82 0.65 0.02 0.33 0.96 0.72 -0.28
MIB 0.80 0.89 0.15 0.68 0.84 0.86 -0.08
IBEX 0.59 0.91 0.31 0.49 0.92 0.83 -0.29

Table 14: Factor and Risk Premium Correlations. The table reports the sample correlation of
daily observations for a combination of option factors and risk premiums over the period 2007-2014.

The first column of Table 14 confirms, as also evident from Figures (2) and (3), that the volatility
and negative jump factors are highly correlated. The remaining columns in Table 14 are organized in pairs,
reflecting the correlation of the two factors with a specific risk premium. The correlation of the spot variance
with the variance risk premium is consistently high. Clearly, whenever the environment is turbulent, the
variance risk premium tends to rise. The variance risk premium is also, but to a lesser extent, positively
correlated with the pure (orthogonalized) jump factor. That is, even if the jump factor reflects the (risk-
neutral) jump variation not contemporaneously accounted for by spot volatility, it is still intimately related
to the variance risk premium. Of course, this partially reflects the fact that the pure jump factor accounts
for a substantial part of the jump risk premium and, as such, constitutes a component of the variance risk
premium.
The next set of columns display the correlations of the option factors with the negative jump risk
premium. The correlations of the negative jump risk premium with the jump factor N JV ⊥ are of particular
interest. The numbers are uniformly high, lending support to the hypothesis that the pure jump factor is

22
the critical ingredient of this premium. The relatively low correlation reported for the S&P 500 index stems
from a single huge (negative) deviation between the V and N JV ⊥ at the onset of the financial crisis in 2008,
where our option pricing model identifies the initial spike in the option prices primarily as a volatility rather
than a jump intensity shock. Table 15 shows that the correlation reaches 0.97 for the 2010-2014 period, fully
in line with the remaining indices, which all attain a value of 0.90 or above over that subsample. In short,
the pure left jump factor, N JV ⊥ , is uniformly highly correlated with the jump risk premium – a particularly
striking feature given the difficulty in, and noise generated by, estimating the expectations regarding the
future negative jump variation. Hence, as hypothesized in the main text, N JV ⊥ may serve as a good proxy
for the negative jump risk premium.
Finally, the last two columns of Table 14 document a high correlation between the volatility factor
and the risk premium associated with the continuous return variation, while the latter is approximately
uncorrelated with the pure jump factor. This is telling. It implies that the variance risk premium only
correlates robustly with the N JV ⊥ factor, as reported in columns two and three, because the former includes
the jump risk premium. Once the negative jump risk premium is stripped from the variance risk premium,
the compensation for diffusive risk has no apparent relation to our primary indicator for negative jump risk
compensation.

23
 V, N JV V RP, V V RP, N JV ⊥ N JRP, V N JRP, N JV ⊥ CV RP, V CV RP, N JV ⊥

2007-2009

SP 0.86 0.90 0.32 0.79 0.72 0.89 0.01

ESTOXX 0.68 0.87 0.39 0.40 0.95 0.90 -0.08
DAX 0.70 0.80 0.37 0.07 0.99 0.92 -0.03
SMI 0.79 0.95 0.21 0.60 0.86 0.95 0.01
FTSE 0.80 0.64 -0.10 0.24 0.97 0.72 -0.38
MIB 0.82 0.93 0.05 0.72 0.79 0.89 -0.13
IBEX 0.49 0.93 0.09 0.35 0.85 0.94 -0.35

2010-2014

SP 0.91 0.91 0.85 0.88 0.97 0.89 0.69

ESTOXX 0.84 0.58 0.77 0.71 0.98 0.57 0.19
DAX 0.83 0.14 0.57 0.23 0.92 0.60 0.30
SMI 0.77 0.30 0.61 0.21 0.97 0.63 -0.03
FTSE 0.83 0.51 0.45 0.58 0.97 0.51 0.02
MIB 0.80 0.84 0.44 0.68 0.90 0.81 0.13
IBEX 0.76 0.86 0.77 0.68 0.98 0.53 -0.24

Table 15: Factor and Risk Premium Correlations. The table reports the sample correlation
of daily observations for a combination of option factors and risk premiums over two sub-periods,
2007-2009 and 2010-2014.

24
A.9 V IX and the Equity Risk Premium

SP
4 0.015

t-statistic
2
0.01

0
0.005
-2
0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

ESTOXX
0.06
4

t-statistic
2 0.04

0
0.02

-2
0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

DAX
4 0.06
t-statistic

2
0.04

0
0.02

-2
0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

SMI
0.05
4
0.04
t-statistic

2
0.03

0 0.02

0.01
-2
0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

FTSE
4 0.025

0.02
t-statistic

2
0.015
0 0.01

0.005
-2
0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

MIB
4
0.1
t-statistic

0.05
0

-2
0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

IBEX
0.06
4

0.04
t-statistic

0
0.02

-2
0
5 10 15 20 25 5 10 15 20 25
Weeks Weeks

Figure 4: VIX-Based Predictive Regressions for Excess Returns. Left Panel: t-statistics
for the regression slopes; Right Panel: Regression R2 .

25
A.10 Parametric vs Nonparametric Tail estimation

The negative jump variation measures extracted using our parametric model (see Section (2)) are similar to
the tail factors extracted from short-term option using the nonparametric method proposed by Bollerslev et al.
(2015)2 . Table 16 shows that the correlation between the two is as high as 80% in the case of the US, where
the moneyness coverage is wide.

SP ESTOXX DAX SMI FTSE MIB IBEX

Mean Parametric 0.006 0.010 0.006 0.005 0.006 0.010 0.017
Mean Nonparametric 0.005 0.006 0.004 0.003 0.003 0.006 0.001
Noise Parametric 0.161 0.226 0.158 0.144 0.110 0.243 0.138
Noise Nonparametric 0.722 0.517 0.724 0.452 0.548 1.338 0.816
Correlation 0.811 0.718 0.539 0.420 0.308 0.411 0.005

Table 16: Parametric vs Nonparametric Tail Estimation: The table reports the mean, stan-
dard deviation and correlation coefficient between the negative jump variation (NJV) computed
using the parameteric model in Section (2) and the tail jump variation extracted using the non-
parametric approximation in Bollerslev
qP et al. (2015). Noise refers to the square root of the mean
T 2
of the squared increments (i.e. t=2 (Xt − Xt−1 ) /(T − 1)), multiplied by 100.

2
The nonparametric tail variation is constructed using the cutoff moneyness points as in Bollerslev et al. (2015)

26
 SPX SX5E DAX SMI
0.08 0.08 0.08 0.08
Nonparametric Nonparametric Nonparametric Nonparametric
0.06 Parametric 0.06 Parametric 0.06 Parametric 0.06 Parametric

0.04 0.04 0.04 0.04

0.02 0.02 0.02 0.02

0 0 0 0
07 08 09 10 11 12 13 14 15 07 08 09 10 11 12 13 14 15 07 08 09 10 11 12 13 14 15 07 08 09 10 11 12 13 14 15

UKX MIB IBEX

0.08 0.08 0.08
Nonparametric Nonparametric Nonparametric
0.06 Parametric 0.06 Parametric 0.06 Parametric

0.04 0.04 0.04

0.02 0.02 0.02

0 0 0
07 08 09 10 11 12 13 14 15 07 08 09 10 11 12 13 14 15 07 08 09 10 11 12 13 14 15

Figure 5: Comparison of Model-Implied and Nonparametric Left Jump Variation. The

figure depicts the daily option-implied left jump variation from model (1) and the nonparametric
left jump variation from Bollerslev et al. (2015).

27
References
Andersen, T. G., N. Fusari, and V. Todorov (2015). The Risk Premia Embedded in Index Options. Journal
of Financial Economics 117, 558–584.

Andersen, T. G., N. Fusari, and V. Todorov (2017). Pricing Short-Term Market Risk: Evidence from Weekly
Options. Journal of Finance 72, 1335–1386.

Bollerslev, T. and V. Todorov (2011). Tails, Fears and Risk Premia. Journal of Finance 66, 2165–2211.

Bollerslev, T., V. Todorov, and L. Xu (2015). Tail Risk Premia and Return Predictability. Journal of
Financial Economics 118, 113–134.

Carr, P. and L. Wu (2009). Variance Risk Premia. Review of Financial Studies 22, 1311–1341.

Corsi, F. (2009). A Simple Approximate Long-Memory Model of Realized Volatility. Journal of Financial
Econometrics 7, 174–196.

Hodrick, R. (1992). Dividend Yields and Expected Stock Returns: Alternative Procedures for Inference and
Measurement. The Review of Financial Studies 5 (3), 357–386.

Lazarus, E., D. Lewis, J. Stock, and M. Watson (2018). HAR Inference: Recommendations for Practice.
working paper .

Wei, M. and J. Wright (2011). Reverse Regression and Long Horizon Forecasting. Journal of Applied
Econometrics 28 (3), 353–371.

28