RESEARCH NOTE

Risk Before Return:

The Layman’s Guide to Volatility Forecasting:
Targeting
PredictingVolatility
the Future,with Higher
One Frequency
Day at a Time Data

• Volatility is generally easier to predict than future returns.

• Volatility forecasting can work reasonably well—but measuring results is not as easy as it appears.
• Dynamically targeting volatility can lead to higher risk-adjusted returns vs. buy and hold.
• Estimation methods have evolved from the 1980s through today as access to more data increased.
• VIX can be a useful indicator but may not be ideal for targeting risk in some applications.
• Capturing both intraday and overnight moves is important for proper risk management.
• Higher frequency data has the potential to boost performance in volatility targeting strategies.
• More sophisticated methods that place more weight on more recent observations tend to outperform.
• Volatility targeting compares favorably to trend-following strategies in reducing risk.
• Adding high frequency returns can significantly improve forecast accuracy using relatively
simple methods.

Volatility is generally
Maintaining an accuratesomething
view on investors Figure
volatility is a critical task1: Cboe Volatility
modeling Index (VIX),
of volatility Dailyit gets used in pricing
and how
seek to avoid. Most investors know increases
in investing. Keeping tabs on market gyrations is not a derivatives or managing risk. But there are some basic
in volatility are most often accompanied by
perfect measure of risk—we care more about volatility on principles and even advanced techniques that can be
declines in their account balances.
the downside than up. But volatility is a useful gauge for used by a wider range of market participants that can
judging how much pain an investor is willing to tolerate help improve their knowledge.
From
or how1990 through
assets can beApril 2020, the
combined in ways to diversify a
correlation between S&P 500 daily returns.
portfolio to maximize risk-adjusted returns The purpose of this guide is two-fold. The first is to
and changes in the Cboe Volatility Index clearly explain some of the foundational principles and
Although compounding returns
(VIX) has been strongly negative—about over time may be the different methods of forecasting volatility for practical
goal, market
-0.70.1 volatility can strike quickly in a matter of use. While we attempt to keep the mix of Greek letters
days with even the most seasoned investors jarred by and inscrutable squiggles down to a minimum, the target
sudden shocks to their portfolio. Having an effective audience is the generally informed, sophisticated investor
While volatility is undesirable, assuming
means of estimating day-to-day risk can help better or market professional.1 The goal is not to “dumb it
some risk is required to earn a return.
prepare for volatility when it does arise and act smarter down”—only make it more accessible. We take the time
The name of the game is maximizing the
in response. to explain key elements and provide additional detail in
amount of return captured per unit of risk,
the footnotes that might be second nature to a quant but
commonly measured
Measuring, forecasting, byand
the interpreting
Sharpe ratio— volatility is
the average excess return (above a risk- could serve as a refresher to those with limited day to day
another matter. There are very smart people with exposure to some of these concepts and methods. 2
free rate like
advanced Treasury
degrees andbills) divided
training thatby its
specialize in theSource: Bloomberg, January 1990 – March 2020
standard deviation (volatility).

1
We count ourselves in this category. While some of us have math degrees and/or training in data science, our approach remains grounded in our
practical experience with institutional investors, technology, exchanges, and algorithmic trading. We value the contributions of academics and quants to
the process and strive to keep learning from them. Many (but certainly not all) academic appers can be a slog through dense material and writing that
is not conducive to absorption. And there is no getting around the fact the material can be difficult, requiring years of training to fully comprehend. Our
objective is to share what we have learned wading through the literature and present it to wider, but informed, audience.
12 Let’s be honest.
Bloomberg data, We have all
January Googled
1990 things
– March we didn’t know or forgot that would cause embarrassment to admit to a colleague.
2020
The second is to highlight the efficacy of using We begin with an overview of volatility’s unusual
higher frequency data to help improve accuracy and characteristics followed by a brief survey of different
responsiveness in forecasting volatility. Although measures and techniques that have been developed
this uses many more data points in the process, even over the years. Then we compare how accurately these
relatively simple methods leveraging high frequency data different measures can predict the volatility in the S&P
tend to perform as well or better than more complicated 500 over a very short period—just one day. We conclude
modeling techniques. We aim to demonstrate how by showing how choices of forecasting tools and
powerful forecasting tools can be constructed using benchmarks can impact the ranking of these different
this more granular data along with basic math at the models for use in practice.
undergraduate level.

Characteristics of Volatility
Volatility (and its squared cousin, variance, which we will refer to interchangeably throughout) has some unusual properties.

Volatility is not directly observable, even after the Volatility tends to be highly autoregressive.
fact.
Fortunately for market practitioners, volatility experienced
We can observe prices and changes that are the result over the recent past can be a useful predictor of volatility in
of volatility, but there is no single value that equates to a the future. In varying over time, volatility also tends to cluster
“true” value. As a result, we can only estimate volatility over around events and then takes some time to decay back to
specific periods of time. 3 The inability to precisely measure some normalized level. These properties lend themselves
its true value presents challenges in judging the accuracy of to some reasonably effective methods of forecasting in
an estimate, even when using multiple benchmarks. comparison to predicting returns, which can be much more
difficult.
Volatility levels are always changing.
Due to its autoregressive properties, even a simple 10-
It is apparent to any market watcher that volatility is day historical standard deviation of returns can be a useful
not constant, adding more complexity to the process of forecast of tomorrow’s volatility. If annualized volatility
modeling its impact. One of the weaknesses of the well- over this 10-day span is rather high at 32%, it is very likely
known Black-Scholes model for pricing options is its tomorrow’s volatility will continue to be high.4 But what
assumption of constant or deterministic volatility. Like are the inputs used to calculate this measure? Like many
many economic assumptions, this was done for modeling estimates and forecasts, 10-day historical volatility only
simplicity rather than any real-world expectations. Many looks at the daily, close-to-close price returns. But what
modern techniques assume volatility is a stochastic or time- about the rest of the day?
varying process (a random walk in a time series), which is
more realistic but also more difficult to model. It is also highly likely that large variations in the close-to-
close price indicating higher volatility will be mirrored in the
Estimating volatility has both continuous and moves intraday as well. But the collective swings in those
discrete elements. relatively small daily returns could be an indication of even
bigger day to day swings in the days to come.
Since most assets have set trading sessions with nights,
holidays, and weekends, there are both continuous and One of the ways to capture some of this information is by
discrete components that must be modeled somewhat incorporating the daily range (the differences between the
separately and can be tricky to reconcile. Theoretically, open, high, low, and close for the day), which is explored
volatility is a continuous process even though there may not in several of the methods in the next section. But we also
be prices available to help estimate its level. But that will not delve into using intraday returns themselves, capturing the
be reflected until price forms at the start of the next session, variation directly in a more continuous process that is closer
creating a gap or “jump” that must be incorporated into the
to the theoretical ideal.
forecast. Focusing on one component at the expense of the
other can lead to mismatches between theory and practice
while introducing significant estimation errors.

3
This is known as a latent variable, which must be inferred from other observable or measurable variables. With volatility, we use the known
values of point to point returns to help estimate variance.
4
A 32% annualized volatility equates to an expected daily move of 2% per day. Using the “Rule of 16”, an annualized figure can be converted
into an estimated daily price swing and vice versa (1% daily moves would equate to 16% annualized volatility). To directly annualize a standard
deviation, the result is multiplied by the square root of time. For daily returns, the square root of the average number of trading days in the US
16 becomes the scaling factor.

2
Tools of the Trade
Including all the methods of estimating and forecasting volatility is well beyond the scope of this analysis. But a core
range of techniques can be generally organized by their methods, sampling frequencies, and inputs ranging from the
very simple to highly sophisticated.

The academic literature often examines estimators more as proxies for the past unobserved latent variance, but many
analyze the same tools for use in forecasting. They are highly related but not identical concepts. For our purposes, we
focus on how well these measures can predict future short-term volatility by lagging them one day, which obviously
needs appropriate proxies to measure results.

Historical
A basic standard deviation of daily returns over the We also make some small changes to the standard
recent past is a logical starting point to begin measuring variance formula7:
volatility. Even those with very sophisticated volatility
models will find historical volatility a useful datapoint for
comparison. Common lookback periods include 10 days
(two weeks accounting for weekends), one month (22
trading days or 30 calendar days), or out a few months A standard sample variance is calculated by the sum of
or longer (100 or even 252 days). 5 the squares of the differences between the observations
and the sample mean divided by the number of
Historical volatility may be simple but there are still observations minus 1. However, there is one modification
some assumptions and modifications to be made. For we make that differs from standard variance in statistics:
one, we typically use log returns for many volatility we assume the average daily log return (the “drift”) will
calculations, as they are continuously compounded, be zero. This reduces it to the sum of the squares of the
instead of simple returns. 6 daily log returns divided by the number of observations
minus one (as shown in the formula above). When looking
at relatively short lookback periods such as 10 days, it

5
We tend to focus on annualized volatility from daily returns even over periods that go well beyond a year. It is certainly possible to use weekly or
monthly returns that may be suitable for longer term analysis. But in general, daily volatility computed over the prior 504 days (2 years) will be higher
than weekly volatility over the prior 104 weeks (the same two years). At lower sampling frequencies, the distribution gets closer to normal (Gaussian),
with the “fat tails” (kurtosis) becoming less pronounced.
6
At higher frequencies of sampling (daily), there is not much difference between simple and logarithmic returns. But log returns are computationally
easier to deal with as you can simply add them to get a compounded end of period return. There are also some quirky mathematical modeling benefits
that come with using log returns that are explained in detail (but plain English) here: https://tinyurl.com/2mu2a3zp.
7
While we use variance and volatility interchangeable, we present all formulas in the paper in their squared form as a daily variance. To
convert to an annualized daily volatility, simply take the square root and multiply by . Some will point out that multiplying by the
square root of time is an imperfect method and tends to either under- or over-estimate volatility—and they are right—but everyone does it
anyway. For additional details, see “Converting 1-Day Volatility to h-Day Volatility: Scaling by [Square Root of Time] is Worse than You Think”.

3
is possible for average returns to be so extraordinarily closing prices. If the opening price is unavailable, the
high or low, they become unrealistic to annualize and model allows for the prior day’s close to be used in place.
significantly distort the volatility calculation. As a result,
it is often simpler to assume zero drift when computing
historical volatility.

Easy to calculate and understand, historical volatility

does have its limitations. For one, it is entirely backward-
looking and makes no attempt to make any adjustments But absent that condition it normally ignores the
to forecast future volatility. It treats every single day in overnight return so suffers from the same systematic
the time series equally, giving more distant data points underestimation error as Parkinson.
the same ability to influence the outcome even though
conditions may have since changed. It also forces the Rogers-Satchell
selection of a lookback period, which can be an arbitrary Unlike the previous estimators, Rogers-Satchell assumes
decision with outsized impact. Lastly, the use of just a non-zero mean return with the same OHLC data points
close-to-close price information ignores intraday to construct a more sophisticated measure.
movements that may provide a more complete picture.

Range-Based Despite the addition of drift to the model, it still ignores

any of the information in the overnight jump, leading this
Another class of estimators attempts to address the measure to also under-estimate full day volatility.
limitation of only using closing prices by incorporating
the daily range into the calculation. Introduced over time Yang-Zhang
from about 1980 to 2000, the methods below build on
each other, adding elements that were missing from Finally, Yang-Zhang brings together the elements of the
prior versions to capture more real-world dynamics and prior range-based models and attempts to address all
improve accuracy. their shortcomings.10 It combines overnight volatility with
a weighted average of open to close and Rogers-Satchell
Parkinson volatility.

The earliest range-based estimator, Parkinson uses the

range defined by the difference between the low and high
of the day instead of closing prices. 8 Where:

One key limitation is the assumption of continuous Overnight Volatility

trading, as no information from the overnight gap (the
prior close to next day’s open or “jump” in academic
parlance) is included, which tends to underestimate
volatility. Approximately one-fifth of the total daily price Open to Close Volatility
variation occurs overnight—a significant portion to be
omitted in the estimate. 9

Garman-Klass

Extending Parkinson, the Garman-Klass (GK) estimator The measure incorporates the overnight jumps while
adds two more range-based measures, the opening and capturing elements of the trend (assumes non-zero mean
return), creating a more robust measure than the other

8
The squared range is divided by a constant to adjust it to be similar in magnitude to a statistical variance sampled over a number of
days.
9
Hansen and Lunde (2005).
10
There is another variant called Garman-Klass-Yang-Zhang that we omit for brevity that just adds an overnight return to Garman-Klass.

4
range-based estimators.11 We address additional aspects Figure 1 - Exponential Smoothing Following a Spike
of weighting overnight versus intraday volatility in a
subsequent section on measures using high frequency,
intraday returns instead of daily OHLC.

Exponential Weighting
While the Yang-Zhang model manages to address many
of the weaknesses in the earlier range-based models,
it still has a significant limitation: all the data points in
the historical lookback are weighted equally. Given the
tendency for volatility to spike and then take some time to
fade, placing more weight on more recent observations
yet accommodating the residual effects of an event that
is still lingering in the market’s memory can help better
model real-world conditions. Without some smoothing
of the data, a volatility event will abruptly drop off from a Despite its utility and simplicity as workhorse, EWMA also
moving window of historical data. has its limitations. For one, it only uses daily returns, leaving
it susceptible to shocks potentially concealed in large
One approach is to use exponentially weighted moving intraday swings that may be a precursor to a larger volatility
average (EWMA) volatility, which applies a decay factor event. Secondly, it forces you to somewhat arbitrarily select
to prior observations to smooth out the series.12 a lambda that can significantly impact the outcome (too
fast or too slow). However, using a combination of two or
more lambdas can help mitigate this, especially with simple
rules that use the higher of the two values (a fast and a
slow) to be on the conservative side.

The decay (represented by lambda, ) must be a value

less than one, typically set to 0.94 for daily data but can GARCH
be adjusted a little higher (for slower decay) or lower (for
GARCH stands for Generalized Autoregressive Conditional
faster decay). Using the default value of 0.94, starting at
Heteroskedasticity¸ and is a parameterized econometric
index 1 in the formula above results in a weight of (1-0.94)
modeling technique that can be used to estimate volatility.
*(0.94)0 = 6.00% for the most recent day followed by (1-
The key innovation with this model begins with the last
0.94) * (0.94)1 = 5.64% for day prior, (1-0.94) * (0.94)2 =
part of the name—heteroskedasticity. In standard linear
5.30% for the third prior day and so on.13
regression, the assumption is that the error term—the
The ease of computation and performance of the EWMA noise that will cause the dependent variable to deviate
estimator makes it a popular choice among practitioners. from its relationship with the independent one—will have
It is also used for index products such as structured notes a constant variance (homoscedasticity). With GARCH,
and fixed index annuities that seek to target volatility the assumption is the variance of the error term will be
levels with daily rebalancing.14 time-varying but is also conditional on the overall level
of variance. Cutting through a lot of the jargon, a GARCH
process is well-matched to the clustering behavior and
persistence of volatility following certain events in the

11
The 1.34 in the formula is a constant (the numerator is - 1) that is intended to minimize variation in the estimator itself. This value is the default
in the original Yang-Zhang paper.
12
Exponential smoothing was popularized in the financial industry by JP Morgan’s RiskMetrics in the mid-1990s as a part of the widely adopted Value
at Risk (VaR) framework. The formula shown above is a little easier to grasp conceptually but EWMA is also written in a reduced, recursive form that is
easier for computation: .
13
The lookback period for an exponentially weighted moving average is theoretically infinity as the weight will decay asymptotically into smaller
numbers that become meaningless to the output. For the default value of 0.94, the cumulative weight remaining drops below 1% after about 75 days.
14
Known as “risk-controlled” strategies, these products use a combination of assets and a risk-free asset such as T-Bills to dynamically target a specific
volatility level. For example, the index may consist of the S&P 500 and cash and seek to target a 5% volatility level. If the estimator predicts the S&P 500 will
have 20% volatility tomorrow, the strategy holds 25% in stocks and the rest in cash (25% = 5% target/20% estimate), rebalancing daily as volatility changes.

5
market. But with these more realistic assumptions come a special case of GARCH (1,1)—and GARCH (1,1) is a
more complicated techniques to manage those conditions. generalized form of EWMA. 21 What EWMA lacks is the
“long memory” in GARCH that assumes reversion to
GARCH was developed in 1986 by Tim Bollerslev some long-term mean level of variance over time. GARCH
following the work of Robert Engle who introduced the is not computationally difficult and open-source tools
“ARCH” part in 1982.15 The most commonly used GARCH in Python and other platforms make it easier to manage
(1,1) model can be specified as follows16: with access to price data and some programming skill. It
is widely used for tasks such as managing portfolio risk or
pricing derivatives. However, it is not without drawbacks.
One is that the basic GARCH (1,1) assumes that positive
or negative shocks have an equal impact on volatility
There is a lot to unpack in this relatively simple equation, when in reality there is a strong negative relationship
but we begin with a plain English description of what the between returns and volatility, commonly referred to as
model is doing. Paraphrasing Engle, GARCH is a weighted the leverage effect. 22 There have been a whole family
average of the long run variance, the predicted variance, of GARCH-based variants (EGARCH, FIGARCH, GJR-
and the variance of the most recent squared residual, GARCH, and others) developed to address asymmetry
which correspond to the terms in the equation above in and add other features to the process, but they come
that order.17 at a cost of higher complexity and difficulty in fitting
the model. More importantly, most GARCH models still
To fit the model, the constants and must be
estimated and have some constraints.18 Getting these generally use only daily closing prices, leaving them
parameters involves using maximum likelihood estimation, vulnerable to more sudden volatility shocks that could
a statistical technique designed to find the values with potentially be detected with higher-frequency intraday
the highest probability that the process in the model returns.
produced the observed data.19 Once fit, a variety of tests
for significance are performed to confirm the model is
correctly specified. 20 Once specified, GARCH produces
an entire time series of variance forecasts, easily updating Realized Volatility
with just the prior prediction and residual. The use of high-frequency returns in modeling volatility
has increased with the wider availability of intraday data
While somewhat complex, GARCH is a very powerful
by both practitioners and academics alike over the last
tool and tends to perform well in forecasting. While
20 years. Known as realized volatility (“RV”)23, it traces
not immediately apparent from the description of the
its intellectual roots back to a paper by Robert Merton
process, the GARCH formula places more weight on the
from 1980, who concluded that volatility could be more
more recent data, similar to the previously described
precisely estimated from realized returns that improve
exponentially weighted volatility. In fact, EWMA is

15
The “generalized” extension of the model is the third term in the equation that helps the model adapt by using its past prediction errors. Bollerslev
was Engle’s research assistant at the University of California San Diego and wrote the GARCH paper as a doctoral student. He currently teaches at
Duke University and continues his work on volatility forecasting, lately more focused on realized volatility techniques using intraday returns.
16
The “1,1” follows the general specification of the model (p,q) which defines p lags in the prior variances and q lags in the residual errors. In addition
to being a representation of the most common model, the GARCH (1,1) notation is cleaner to show for our purposes.
17
Googling “GARCH” and reading through many of the explanations yields some dreadful results. There are some very basic descriptions with no real
details and then a lot of very dense academic references but not much in the middle. However, Engle’s GARCH 101 intro (available at https://www.stern.nyu.
edu/rengle/GARCH101.PDF) is an accessible read that is clearly targeted towards his students (he is currently a professor at NYU’s Stern School of Business).
18
The parameters must sum to one but the term is actually the long run variance multiplied by another weight (gamma). All parameters must also
be greater than zero and alpha and beta must sum to a value less than 1.0.
19
This is commonly done using statistical packages or Python or R libraries. For additional information, check out this very clear and concise explainer
available at Probability concepts explained: Maximum likelihood estimation | by Jonny Brooks-Bartlett | Towards Data Science.
20
As with maximum likelihood, GARCH can be a lot simpler to specify and use with software packages and Python/R libraries that can include a
number of these tests for significance in the package.
21
To get EWMA from GARCH, the long run variance term is set to 0 and the rest reduces to two weights that equal 1.0 (lambda in EWMA, alpha and
beta in GARCH), each multiplied by the identical variance and squared return terms in the recursive form of the EWMA equation and GARCH above
(see footnote 12 for more details).
22
Black, F. (1976) Studies in Stock Price Volatility Changes of the Nominal Excess Return on Stocks. Proceedings of the American Statistical
Association, Business and Economics Statistics Section, 177-181.
23
This is no different than the commonly accepted realized volatility, in which we are referring to the variability of returns that has been historically
realized at a daily or lower frequency. Intraday realized volatility is simply a compressed version of this, measuring the variability within a single day via
smaller (potentially tick-by-tick) intervals.

6
with higher sampling frequencies. 24 The realized variance above that sums the squares of intraday returns for the
for a single day is simply the sum of squared intraday day at a fixed interval such as 5 minutes. This adds the
returns: rich information available from higher frequency data,
sampled at a low enough frequency to avoid distortion
from market microstructure noise. Secondly, it uses a
cascading series of lagged terms that are designed to
capture the actions of market participants with differing
The sampling frequency used generally ranges from time horizons in their investment process. 27 Lastly, it
5 to 30 minutes, with the “M” in the formula above uses simple linear regression (OLS) to estimate its three
corresponding to the number of bars that constitute a full coefficients, capturing some longer-term historical
day (78 for 5-minute bars in a 6.5-hour US trading day patterns across multiple market regimes. This component
from 9:30AM-4PM). 25 As with some of the earlier historical mimics, but does not completely replicate, the “long
estimators, there is no “drift” or mean return subtracted memory” feature of GARCH.
from each squared return as the assumed return from
intraday returns is even closer to zero than daily. The model in its simplest form can be specified as follows:

As a forecasting tool, RV suffers from some of the same

limitations as some of the earlier range-based estimators.
Without the overnight return, an RV-only estimator will
tend to underestimate the total variation that is important where , and represent realized variance
to a practitioner. The overnight gap plus the variability (squared terms) for a one-day, one-week (5 day), and one-
during the day are both metrics that matter for a position month (22 day) period on day t. 28 To get the multi-day
or hedge and would be measured by a real-time risk measure, the “building block” of a single day’s RV is simply
management system. averaged over the desired period (e.g. sum up the RV for 5
days and divide by 5).
In the next section, we focus on RV as a building block
of more sophisticated forecasting techniques. But for Estimating the coefficients involves initializing the model at
completeness, we also evaluate its efficacy as a forecast a point in the past using as much historical data as possible.
tool itself alongside the other estimators. However, for a more robust and practical implementation,
we use an expanding window—only letting the model use as
much information that existed up to the point of the forecast
date. This “walk-forward” approach helps avoid look-ahead
bias and reduces the potential for overfitting the model. 29
HAR-RV
HAR-RV tends to perform very well against the more
Introduced by Fulvio Corsi, the Heterogeneous complex GARCH and is easier to fit. It possesses some of
Autoregressive Realized Volatility (HAR-RV) model has a the mean-reverting properties of GARCH and places more
relatively simple structure featuring three main drivers. 26 weight on recent observations while capturing the dynamics
The first is the use of daily RV—the same measure defined of intraday price action for improved response to volatility
shocks. Like GARCH, HAR-RV has spawned a number of

24
High-quality intraday historical data was not generally available in the 1980s. But as Merton also pointed out, there are theoretical limits to sampling
at very high frequencies as the benefits of the increased precision can be swamped by the effects of market microstructure noise—the bouncing
around between the bid/offer and any discrepancies in latency. The original paper can be accessed at https://www.nber.org/system/files/working_
papers/w0444/w0444.pdf.
25
The standard in the academic literature is 5-minute returns but similar results are also generated from bars using anywhere from 10-30 minutes. Any more
frequent sampling than 5 minutes is generally thought to introduce too much noise in the process to be useful although there are a handful of studies going
down to 1 minute, especially on broad large-cap indices that are aggregates of many stocks and somewhat less susceptible to noisy microstructure effects.
26
Corsi (2009).
27
The rationale behind the Heterogeneous Markets Hypothesis (HMH) from Ulrich Muller (1993) is that volatility is a function of trading activity from
multiple agents in the market with distinct investing horizons: short term speculators (intraday), medium term traders (days), and long-term investors
(weeks or longer). A brief article by Muller himself that succinctly describes the theory is available at http://m.e-m-h.org/DMOP01.pdf.
28
While some researchers use overlapping data points in the regression (the one-day RV is a component of the 5-day RV), we find better results in an
alternative approach that only spans unique information in each term. Effectively, the lags are computed as the simple averages of the Zero-th, 1st to
4th, and 5th to 21st lags, respectively.
29
This process purposefully weakens the predictive power of the model in the earlier periods as less information is available to use in the forecast. It
involves recursively fitting the OLS regression every day, adding information to the model as it becomes available. Since the coefficients tend to be very
unstable with an insufficient amount of data in the beginning, we “burn off” a year of historical data to provide time for them to stabilize. As more time
passes, the marginal impact of a single day on the coefficients becomes very small, but will respond to true volatility shocks, which is a desired feature.

7
extensions, focusing on correcting for asymmetry (using Dow Jones Industrial Average stocks. In their paper that has
semi-variance in SHAR) or dynamically adjusting RV due been extensively referenced and replicated, they propose
to inherent measurement error (“integrated quarticity” in three ways to combine RV with overnight returns to get an
Bollerslev’s HARQ). But the biggest item to address is the equivalent variance for the entire day.31 The paper primarily
glaring weakness in an RV-only estimator—the lack of any focuses on the problem of getting the best estimator of
contribution from the overnight jump. integrated variance, but we use similar techniques to adjust
HAR-RV-based forecasts to daily equivalents.
Without addressing the jump, HAR-based models will tend
to underestimate volatility in comparison to methods like The first is to simply multiply the RV-based measure by a
EWMA or GARCH that capture this information. A HAR- static scalar representing the average contribution from
based model may be very effective in forecasting the RV the overnight return. Using Hansen and Lunde’s estimate
for a single day, but the overnight return represents risk to of one-fifth, this translates to a scalar of 1.25x (squared for
practitioners that cannot be ignored even if RV is a “better” variance).32 This static scalar can perform well, especially
representation of the integrated, unobservable variance. for broad market indices, but can be improved with a more
dynamic approach. The second method just adds the
Bollerslev added a jump term to the basic model with his overnight return to the RV measure as Bollerslev suggests.
HAR-J, simply folding in the previous day’s overnight return But Hansen and Lunde argue this method adds a very noisy
as another term in the regression. Others decompose the measure to the estimate, potentially distorting the result.
jump into separate components matching the 1-, 5-, 22-day
lags in the RV terms. More recently, Bollerslev concluded that The last method involves an optimized weighting of the
simply adding the squared overnight return to the sum in overnight and RV components that helps smooth out
calculating RV itself can be a more effective means of adding some of the noisiness in just adding the overnight return.
the information.30 However, we find an alternative approach The weighting method is complex and beyond the scope
performs even better: scaling the RV-based forecast up to of this paper to go through in any detail but has been
better approximate a measure using daily data. well-supported by subsequent research since its initial
publication.33 We combine this optimization process with
As referenced before, Hansen and Lunde (2005) estimate the HAR-RV in forming our preferred forecast detailed in the
that overnight returns contribute about one-fifth of the next section.
daily price variation based on their historical analysis of

30
Bollerslev, T, Tauchen, G, and Zhou, H (2009), “Expected Stock Returns and Variance Risk Premia”, Review of Financial Studies, 22, 4463—4492.
31
Hansen and Lunde (2005).
32
If one-fifth of whole-day variation is overnight that leaves 80% for the open to close period. Scaling up from 80% to 100% is 1.25x.
33
Hansen and Lunde described differences in the relationship between the overnight and intraday volatility across different types of assets, noting
a much larger impact from overnight returns in the limited number of technology stocks in their narrow universe of DJIA stocks. However, Todorova
and Soucek and Ahonieme and Lanne further demonstrate these differing relationships using the broader market (S&P 500) and international markets
(using Australian equities) in concluding the superiority of the optimal weighting to other methods.

What about the VIX?

Most investors are familiar with the so-called “Fear Gauge”, the Cboe Volatility Index (VIX). The VIX was designed to
measure market expectations of 30-day volatility implied by the pricing of S&P 500 index options. While forward-looking
and driven by market prices, we omit the VIX for this analysis of volatility forecasting for several reasons.

For starters, while the VIX is a very useful indicator for the level of market volatility, the index itself is not tradable.
Futures, options, and exchange-traded funds on the VIX are available but do not track the current spot index all that well—
they align to what the market expects the VIX will be on expiration. They also have a term structure, with differing market-
driven expectations of where VIX will be one month out, two months out, and so on.

More importantly, VIX is not well-suited for a one-day forecast as it measures expectations over the next 30 days. In
a prior study*, we did use VIX as an estimate for a simple volatility-targeting strategy with daily rebalancing and it
performed relatively poorly, even compared to a naïve 10-day historical volatility estimate.

To be clear, there is predictive value in the volatility implied by options pricing and some volatility models make use of this
data. But we can only cover so much material and VIX and implied volatility could span several volumes.

*Barchetto, A and Poirier, R., “Risk Before Return: Targeting Volatility with Higher Frequency Data”, available at https://tinyurl.com/6nvjhrnc.

8
Comparing Forecast Accuracy
Now that we have defined a general set of estimators and forecasting methods ranging from basic to advanced, we
compared the predictive power of each along with their pros and cons. For our test, we analyzed the ability of each
estimator to forecast volatility one day ahead using a variety of proxies as benchmarks.

Data

As a proxy for US equity market volatility, we used returns from the SPDR S&P 500 ETF (SPY), with data from January
1, 1997 through December 31, 2020. We sourced open, high, low, and closing prices from Bloomberg and intraday
15-minute bars from Cboe’s Livevol Datashop.

Variance Proxies

Since true volatility is unobservable, selecting more than one proxy is advisable. To avoid biasing conclusions towards
the more intraday-focused realized measures versus less frequent sampling using daily data, we used a range of both
as proxies for all estimators.

Table 1 - Variance Proxy Calculations

Proxy Formula Description

Squared Return Daily return, squared

The sum of squared intraday 15-minute returns (where

Realized Variance
M=26)

The difference between the high and low of the day,

Range
squared34

Adds the squared overnight return to the intraday realized

Full Day Realized Variance
variance to capture the whole day35

Loss Functions

In addition to benchmarks to measure against, the tools used to measure the errors from the predictions are equally
as important. Like with the proxies, we used a range of loss functions36 to provide more context. In the formulas in the
following table, represents the model-estimated value and the h represents the (true) variance proxy.

34
The high-low range and squared range are traditional variance proxies in the academic literature. However, the range as a proxy is subject to
some unrealistic assumptions in the process that generated the data to be properly evaluated—zero mean return, no jumps, and constant conditional
volatility during the day. Patton (2008) proposes an adjustment to the range (dividing it by a constant, ), which when squared becomes
an unbiased proxy for conditional variance. However, this does not change the ranking of our estimators in any way and adds some complexity, so we
use the more traditional range for simplicity. For additional information on adjusted range and evaluating variance proxies, please see: http://public.
econ.duke.edu/~ap172/Patton_robust_forecast_eval_11dec08.pdf.
35
This adjustment to realized variance is our preferred measure as it addresses both the intraday fluctuations as well as the opening gap. For
practitioners, both are very relevant to a live position or hedge that is being monitored day to day. A miscalculation that leads to a mark that produces
a significant hit to the P&L at the opening of trading or later in the day could lead to a tap on the shoulder—or worse.
36
A “loss function” is a way of measuring how different a group of predicted values are from the true/known values and then expressing this as a
single number. Lower values are preferred, indicating “less loss/error” from the true/known values.

9
Table 2 - Loss Function Calculations

Loss Function Formula Interpretation

Mean Squared Error (MSE) Penalizes more extreme errors

Denominated in the same units as the underlying

Root Mean Squared Error (RMSE) data, making it easier to compare across models
(like standard deviation compared to variance)

More robust to outliers, which may be an

Mean Absolute Error (MAE)
undesirable feature

An asymmetric measure that penalizes under-

Quasi-Likelihood (QLIKE)
estimation more harshly

Use of the “wrong” loss function can lead to some significant errors in ranking the efficacy of the forecasts. In prior
work on volatility forecasting, MSE and QLIKE are shown to be more robust to noise in the proxy, making them
generally preferred in terms of ranking. 37 While we focus on these two robust measures in our analysis, we include
the other loss functions to highlight strengths and weaknesses of the estimators and show how their rankings can be
distorted.

Forecast Methods

We use all the estimators presented in this paper for a comprehensive comparison. For measures that require a
historical lookback, we specified short (10 day), medium (22 day), and long (100 day) periods and calculated a set of
results for each.

Table 3 - Estimators and Specifications

Estimator Short Name Additional Information

Historical Variance HV 10, 22, 100 days

Parkinson Park 10, 22, 100 days

Garman-Klass GK 10, 22, 100 days

Rogers-Satchell RS 10, 22, 100 days

Yang-Zhang YZ 10, 22, 100 days

EWMA EWMA 0.94 lambda

GARCH (1,1) GARCH Fit using maximum likelihood via arch v4.19 for Python

Realized Variance RV Prior day, uses 15-minute intraday bars

HAR-RV HAR-RV 1, 5, 22 lags, unique spans only38

Scaled HAR-RV HAR-RV+ Multiplies HAR-RV by 1.252

A weighted sum of HAR-RV and the average squared overnight return over the
Optimized HAR-RV HAR-RV*
prior month39

37
See the previously referenced paper by Andrew Patton for more details on the robustness of forecasts, available at:
http://public.econ.duke.edu/~ap172/Patton_robust_forecast_eval_11dec08.pdf.
38
See footnote 28 on modification to the standard Corsi HAR process. Also uses the same 15-minute bars as RV.
39
This process uses the same weighting technique from Hansen and Lunde only with different estimators for the intraday and overnight components.
We substitute HAR-RV for RV in the intraday and the mean squared close to open return over the prior 22 trading days for the one day overnight. Both
are intended to improve accuracy and smooth out some of the noise in the overnight return from day to day.

10
Exhibit 1 - Ranked Forecast Performance by Estimator and Variance Proxy

Panel A: Squared Return

HAR-RV+ HAR-RV+

HAR-RV* HAR-RV*

RV YZ

HAR-RV GARCH

GARCH HAR-RV

RS EWMA

GK GK

YZ RS

Park Park

HV HV

EWMA RV

20 21 22 23 24 25 26 27 1.00 1.20 1.40 1.60 1.80

Mean Squared Error QLIKE

Panel B: Realized Volatility

HAR-RV HAR-RV

Park Park

RV GK

GK RS

RS RV

HAR-RV+ YZ

HAR-RV* HV

GARCH GARCH

EWMA HAR-RV*

HV EWMA

YZ HAR-RV+

0 2 4 6 8 0.10 0.15 0.20 0.25 0.30 0.35 0.40

Mean Squared Error QLIKE

Panel C: Range

HAR-RV+ HAR-RV+

HAR-RV* HAR-RV*

HV GARCH

GARCH EWMA

YZ YZ

EWMA HAR-RV

RV HV

RS GK

GK RS

Park Park

HAR-RV RV

0 10 20 30 40 50 0.0 0.5 1.0 1.5 2.0 2.5

Mean Squared Error QLIKE

Panel D: Full Day Realized Variance

HAR-RV* HAR-RV*

HAR-RV+ HAR-RV+

RV GARCH

GARCH YZ

HV HAR-RV

HAR-RV EWMA

RS GK

GK Park

Park RS

YZ HV

EWMA RV

0 2 4 6 8 10 12 14 0.0 0.1 0.2 0.3 0.4 0.5 0.6

Mean Squared Error QLIKE

11
Exhibit 2 - All Forecast Results and Performance Relative to EWMA

All Estimators and Loss Functions Relative to EWMA

MSE RMSE MAE QLIKE MSE RMSE MAE QLIKE

Squared Return

Historical Variance 26.59 5.16 1.664 1.646 99.1% 99.6% 99.4% 106.9%
Parkinson 26.04 5.10 1.458 1.583 97.1% 98.5% 87.1% 102.8%
Garman-Klass 25.80 5.08 1.457 1.566 96.2% 98.1% 87.0% 101.7%
Rogers-Satchell 25.60 5.06 1.465 1.568 95.5% 97.7% 87.5% 101.8%
Yang-Zhang 25.95 5.09 1.639 1.474 96.8% 98.4% 97.9% 95.7%
EWMA 26.82 5.18 1.674 1.540 100.0% 100.0% 100.0% 100.0%
GARCH 25.59 5.06 1.633 1.493 95.4% 97.7% 97.6% 97.0%
Realized Variance 22.48 4.74 1.384 1.800 83.8% 91.6% 82.7% 116.9%
HAR-RV 24.68 4.97 1.403 1.501 92.0% 95.9% 83.8% 97.5%
Scaled HAR-RV 22.13 4.70 1.561 1.446 82.5% 90.8% 93.2% 93.9%
Optimized HAR-RV 22.18 4.71 1.550 1.447 82.7% 90.9% 92.6% 94.0%

Realized Variance

Historical Variance 6.77 2.60 0.936 0.347 128.3% 113.3% 98.7% 96.9%
Parkinson 3.27 1.81 0.611 0.252 62.0% 78.7% 64.4% 70.4%
Garman-Klass 3.46 1.86 0.626 0.259 65.6% 81.0% 66.0% 72.4%
Rogers-Satchell 3.74 1.93 0.650 0.272 70.9% 84.2% 68.5% 76.0%
Yang-Zhang 7.15 2.67 0.939 0.321 135.4% 116.4% 99.0% 89.5%
EWMA 5.28 2.30 0.949 0.358 100.0% 100.0% 100.0% 100.0%
GARCH 4.56 2.13 0.878 0.351 86.3% 92.9% 92.5% 98.0%
Realized Variance 3.45 1.86 0.561 0.307 65.4% 80.9% 59.2% 85.7%
HAR-RV 2.65 1.63 0.524 0.238 50.3% 70.9% 55.2% 66.4%
Scaled HAR-RV 3.98 1.99 0.841 0.367 75.4% 86.8% 88.6% 102.5%
Optimized HAR-RV 4.54 2.13 0.838 0.357 86.0% 92.7% 88.4% 99.8%

Range

Historical Variance 37.49 6.12 2.004 1.251 91.7% 95.8% 99.5% 149.1%
Parkinson 43.87 6.62 2.127 1.405 107.3% 103.6% 105.5% 167.4%
Garman-Klass 43.46 6.59 2.121 1.378 106.3% 103.1% 105.2% 164.2%
Rogers-Satchell 42.87 6.55 2.122 1.400 104.9% 102.4% 105.3% 166.8%
Yang-Zhang 38.53 6.21 2.002 0.894 94.3% 97.1% 99.3% 106.5%
EWMA 40.87 6.39 2.016 0.839 100.0% 100.0% 100.0% 100.0%
GARCH 38.50 6.21 1.955 0.724 94.2% 97.1% 97.0% 86.3%
Realized Variance 41.32 6.43 2.203 2.167 101.1% 100.6% 109.3% 258.2%
HAR-RV 44.51 6.67 2.163 1.141 108.9% 104.4% 107.3% 135.9%
Scaled HAR-RV 36.99 6.08 1.886 0.581 90.5% 95.1% 93.6% 69.2%
Optimized HAR-RV 37.08 6.09 1.927 0.649 90.7% 95.3% 95.6% 77.4%

Full Day RV

Historical Variance 11.10 3.33 1.000 0.398 90.2% 95.0% 97.8% 121.2%
Parkinson 11.88 3.45 0.836 0.384 96.5% 98.2% 81.8% 116.9%
Garman-Klass 11.74 3.43 0.839 0.377 95.4% 97.7% 82.0% 114.9%
Rogers-Satchell 11.60 3.41 0.853 0.386 94.2% 97.1% 83.4% 117.6%
Yang-Zhang 11.94 3.46 0.990 0.306 97.0% 98.5% 96.8% 93.1%
EWMA 12.31 3.51 1.023 0.329 100.0% 100.0% 100.0% 100.0%
GARCH 10.43 3.23 0.940 0.295 84.8% 92.1% 91.9% 89.7%
Realized Variance 10.28 3.21 0.838 0.595 83.5% 91.4% 81.9% 181.2%
HAR-RV 11.33 3.37 0.790 0.318 92.1% 95.9% 77.3% 96.8%
Scaled HAR-RV 10.03 3.17 0.894 0.293 81.5% 90.3% 87.4% 89.1%
Optimized HAR-RV 9.88 3.14 0.886 0.286 80.3% 89.6% 86.7% 87.2%

12
Results Overview proxy, we also report performance of all estimators
benchmarked relative to EWMA, a standard estimator
In Exhibit 1, we ranked all estimators by their forecasting with wide application throughout the financial industry.
performance against each proxy, focusing on the two All values below one (more accurate than EWMA) are
robust loss functions, MSE and QLIKE. We limited analysis colored green whereas values greater than one (less
to the 10-day lookback period for all estimators that accurate than EWMA) are colored red.
require one as they clearly outperformed the longer term
22-day and 100-day lookbacks across all metrics by an Focusing on the results relative to EWMA highlights some
overwhelming margin. broad differences in how these estimators work. The HAR
models and GARCH are clustered fairly closely to EWMA
Despite being considered a noisy proxy, Squared Return across the range of variance proxies, owing to their
in Panel A in Exhibit 1 provides some initial insights into inclusion of jump returns and weighting of more recent
the strengths and weaknesses of each estimator. The values.
realized variance measures (RV, HAR-RV*, HAR-RV+)
The output for QLIKE and MSE (and indirectly RMSE as
performed best against MSE, but RV itself drops to the
it is just the square root of MSE) aligns with the rankings
bottom using QLIKE. Since Squared Return includes the
in Exhibit 1. But looking at the other metrics such as MAE
overnight jump, the intraday-only RV is penalized more for is quite instructive as it demonstrates the importance of
its under-estimation. EWMA performed relatively poorly using the correct loss function for the task. Estimators
against MSE but was much better using QLIKE, helped by like RV or HAR-RV that use intraday data and exclude
its inclusion of the overnight return in its calculation and the jump component look strong compared to EWMA,
higher weighting of more recent observations. GARCH, GARCH, HAR-RV*, or HAR-RV+ when measured against
and the three HAR variants are all consistently strong a Squared Return proxy and MAE. But when factoring in
against both loss functions. the risk of under-estimation, HAR-RV slips and RV falls to
the worst ranked estimator in using QLIKE to evaluate.
Since Realized Variance removes the overnight
component, it should be no surprise that RV, HAR-RV, and Across the loss functions and multiple proxies, the
the range-based estimators without a jump component realized volatility estimators that adjust for overnight
performed relatively better in Panel B. The relatively poor returns—Scaled HAR-RV (HAR-RV+) and Optimized HAR-
performance of standard industry methods GARCH and RV (HAR-RV*)—are consistently top performers. While
EWMA against QLIKE suggests that Realized Variance is measuring against the combination of intraday squared
not well-aligned with practice as a benchmark, as both are returns and the squared overnight gap might seem tailor-
widely used in real-world applications. And the addition made for these estimators, their strong performance
of an overnight component to HAR-RV+ and HAR-RV* against all the proxies further supports their case as very
is designed to make them equivalent to measures like effective forecasting tools. And while not showing up in
GARCH and EWMA. the top two for any column, EWMA and especially GARCH
also performed very well, reflective of their traditional
Panel C shows that despite their name, the range-based and continued use by practitioners. The more naïve early
estimators (Parkinson, Garman-Klass, Rogers-Satchell) estimators based on historical standard deviation or daily
except for Yang-Zhang were relatively poor predictors range are simply less effective, only scoring well against a
of the Range. The more sophisticated models performed limited combination of proxy and loss functions.
well, especially against QLIKE, with the weighting of
more recent results helping EWMA and GARCH guard
against under-estimation. HAR-RV* and HAR-RV+ also Conclusion
place more weight on recent results, but also capture
intraday variation which apparently leads to some value The wider availability of intraday data over the last
in forecasting the daily range. twenty years has encouraged more research into its use
for volatility forecasting. But while realized volatility
Lastly, using the full range of intraday and overnight is used in practice among quantitatively-oriented
variation as a proxy in Panel D, the more sophisticated asset managers, investment banks, and other financial
estimators that either directly capture or model both institutions, anecdotally EWMA and GARCH are still more
components perform well whereas the range-based prevalent in practice. For specialized volatility-targeted
(Parkinson, Garman-Klass, Rogers-Satchell) and more products such as risk control indices used in structured
naïve ones (RV, HV) performed poorly, especially RV products or fixed index annuities, EWMA-based
against QLIKE. forecasts for targeting are clearly dominant (although
different models are used to help hedge their exposure).
Exhibit 2 shows the full range of values by variance proxy
For portfolio risk management and derivatives pricing,
for each estimator against all four loss functions. The
more sophisticated variants of GARCH are widely in use,
top two most accurate estimators in each column for
which may perform significantly better than the standard
each proxy are colored in green and the least accurate
GARCH (1,1) presented here. There are also some GARCH
two in red. To the right of the performance against each
models that use realized volatility components, adding

13
the increased responsiveness of intraday data to improve HAR-RV* is very well suited for practical application in
their forecasts. But they come at the cost of increased day-to-day risk management applications.
complexity in fitting and maintaining these models. For
accuracy, responsiveness, and ease of calculation, it is As stated previously, covering all forms of volatility
difficult to beat some of the HAR-based estimators that forecasting in a relatively brief overview would be
are properly tuned for the full day’s volatility. impossible. In addition to many variants in the GARCH
family of models, there are other methods we do not
While the basic HAR-RV has been in use for over a even mention (the Heston model, Markov chains, Monte
decade, there are still ongoing efforts to improve its Carlo simulations, and other stochastic volatility models).
forecasting accuracy using the same basic structure. Market-derived inputs such as option implied volatility
Bollerslev, who initially developed GARCH, has been or VIX are another form of forward-looking measure that
primarily focused on research using realized volatility can be used to help predict future volatility.
and HAR-based models in recent years. The fine-tuning
of the continuous intraday and overnight components Regardless of the tools and methods used, we
into an estimate that best captures the desired risk believe that higher frequency data will continue to
characteristics for a given day is of critical importance. be an evolving part of volatility forecasting and risk
While the realized variance may be an unbiased proxy management for years to come. Intuitively, having more
for the latent integrated variance, it fails to account for frequent sampling of data should lead to more accurate
all the practical risks that a professional with capital at results. If asked to forecast tomorrow’s temperature
risk faces. A sharp loss at the opening of trading in a at 3pm, would you rather use the seasonal average,
volatile market that is exacerbated by an inferior risk the average over the last two weeks, or the hourly
model is a threat to one’s P&L and potentially their job. temperature over the last 48 hours? With the low cost
In our opinion, the impact of the overnight return plus of computing power and data storage compared to
the variation experienced throughout the trading session even ten years ago, we expect the future of forecasting
is most aligned with the practitioner and the best proxy will harness this data for even more sophisticated
for volatility. As a result, a risk forecast that explicitly techniques, especially where they tend to be most
accounts for these components with solid performance effective in the short-term.
across a range of variance proxies such as HAR-RV+ or

14
 About the Authors

Anthony Barchetto, CFA

[email protected]

Tony is the Founder and Chief Executive Officer of Salt Financial, a leading provider of index solutions and
risk analytics for investment advisors, fund sponsors, investment banks, and insurance carriers. Salt develops
risk-focused index strategies built for higher precision and responsiveness to changing markets using its
truVol® and truBeta® analytics.

Prior to founding Salt, Tony led corporate development for Bats Global Markets, helping guide the firm
from its merger with rival Direct Edge to initial public offering and subsequent acquisition by Cboe Global
Markets. Before Bats, Tony served in a similar capacity for global institutional trading platform Liquidnet.
Earlier in his career, he served in leadership roles in trading, sales, and product management with Citigroup
and Knight Capital Group.

Tony earned a BS in International Economics from Georgetown University and is a CFA Charterholder.

Ryan Poirier, ASA, CFA, FRM

[email protected]

Ryan is Director, Index Products & Research for Salt Financial. In his role, he is responsible for the
conceptualization and design of index-based strategies leveraging Salt Financials’ full suite of truVol® and
truBeta ® analytics. He also leads the research efforts, focusing primarily on risk forecasting and portfolio
construction.

Prior to joining Salt Financial, Ryan spent three years at S&P Dow Jones Indices as part of the Global
Research & Design team, where he grew the quantitative index offerings for banks, asset managers, and
insurance companies. He is a CFA Charterholder, FRM Charterholder, and an Associate of the Society of
Actuaries. Ryan earned a MS in Financial Engineering from New York University and a BS in Mathematics and
Finance from SUNY Plattsburgh.

Junseung (Eddy) Bae

[email protected]

Eddy is Manager, Products and Technology for Salt Financial. Eddy is responsible for driving the product
development and technology processes at Salt. Eddy has held several technical and analytical roles with
e-commerce and fintech startups.

He is a former army intelligence analyst with the Republic of Korea and holds a BS in Business with a
concentration in Computing and Data Science from New York University.

About Salt Financial

Salt Financial LLC is a leading provider of index solutions and risk analytics, powered by the patent pending truVol®
Risk Control Engine (RCE) and proprietary truBeta ® portfolio construction tools. We leverage the rich information
contained in intraday prices to better estimate volatility to develop index-based investment products for insurance
carriers, investment banks, asset managers, and ETF issuers. Salt is committed to collaborating with industry leaders
to empower the pursuit of financial outperformance for investors worldwide.

For more information, please visit www.saltfinancial.com.

15
Academic References

Ahoniemi, K. & Lanne, M. (2010). Realized volatility and overnight returns. Bank of Finland Research Discussion Papers, 19.
Andersen, T. G., & Bollerslev, T. (1998). Answering the skeptics: Yes, standard volatility models do provide accurate
forecasts. International Economic Review, 39(4), 885-905.
Andersen, T. G., Bollerslev, T., & Diebold, F. X. (2007). Roughing it up Including Jump Components In the Measurement
Modeling and Forecasting of Return Volatility. The Review of Economic and Statistics, 89(4), 701-720.
Audrino, F., & Knaus, S. (2016). Lassoing the HAR Model: A Model Selection Perspective on Realized Volatility Dynamics.
Econometric Reviews, 35(8-10), 1485-1521.
Audrino, F., Huang, C., & Okhrin, O. (2017). Flexible HAR Model for Realized Volatility. Working Paper.
Bandi, F. M., & Russell, J. R. (2008). Microstructure Noise, Realized Variance, and Optimal Sampling. The Review of
Economic Studies, 75(2), 339-369.
Barndorff-Nielsen, O. E., & Shephard, N. (2002). Econometric analysis of realized volatility and its use in estimating
stochastic volatility models. Journal of the Royal Statistical Society, 64(2), 253-280.
Bekierman, J., & Manner, H. (2016). Improved Forecasting of Realized Variance Measures. Working Paper.
Biais, B., Glosten, L., & Spatt, C. (2005). Market Microstructure: A Survey of Microfoundations, Empirical Results, and
Policy Implications. Journal of Financial Markets, 8(2), 217-264.
Bollerslev, T., Patton, A. J., & Quaedvlieg, R. (2016). Exploiting the errors: A simple approach for improved volatility
forecasting. Journal of Econometrics, 192, 1-18.
Bollerslev, T., Tauchen, G., & Zhou, H. (2009). Expected Stocks Returns and Variance Risk Premia. Review of Financial
Studies, 22, 4463-4492.
Buccheri, G., & Corsi, F. (2019). Hark the Shark: Realized Volatility Modelling with Measurement Errors and Nonlinear
Dependencies. Working Paper.
Corsi, F. (2009). A Simple Long Memory Model of Realized Volatility. The Journal of Financial Econometrics, 7, 174-196.
Corsi, F., & Reno, R. (2009). HAR volatility modelling with heterogeneous leverage and jumps. Working Paper.
Fleming, J., Kirby, C., & Ostdiek, B. (2003). The Economic Value of Volatility Timing Using ‘Realized’ Volatility. Journal of
Financial Economics, 63(3), 473-509.
French, K. R., & Roll, R. (1986). Stock Return Variances: The Arrival of Information and the Reaction of Traders. Journal
of Financial Economics, 17, 5-26.
Hansen, P. R., & Lunde, A. (2005). A Realized Variance for the Whole Day Based on Intermittent High-Frequency Data.
Journal of Financial Econometrics, 3(4), 525-554.
Koopman, S. J., Jungbacker, B., & Hol, E. (2005). Forecasting Daily Variability of the S&P 100 Stock Index Using
Historical, Realized and Implied Volatility Measurements. Journal of Empirical Finance, 12, 445-475.
Liu, L. Y., Patton, A. J., & Sheppard, K. (2015). Does anything beat 5 minute RV A comparison of realized measures
across multiple asset classes. Journal of Econometrics, 187, 293-311.
Luong, C., & Dokuchaev, N. (2018). Forecasting of Realised Volatility with the Random Forests Algorithm. Journal of Risk
and Financial Management, 11(4).
Madhavan, A. (2000). Market Microstructure: A Survey. Journal of Financial Markets, 3(3), 205-258.
Martens, M. (2002). Measuring and Forecasting S&P 500 Index Futures Volatility Using High-Frequency Data. Journal of
Futures Markets, 22, 497-518.
McAleer, M., & Medeiros, M. (2008). A Multiple Regime Smooth Transition Heterogeneous Autoregressive Model for
Long Memory and Asymmetries. Journal of Econometrics, 147(1), 104-119.
McAleer, M., & Medeiros, M. (2011). Forecasting Realized Volatility with Linear and Nonlinear Models. Journal of
Economic Surveys, 25(1), 6-18.
Merton, R. (1980). On Estimating the Expected Return on the Market. Journal of Financial Economics, 8, 323-361.
Patton, A. J., & Sheppard, K. (2015). Good Volatility, Bad Volatility: Signed Jumps and The Persistence of Volatility. The
Review of Economic and Statistics, 97(3), 683-697.
Todorova, N. & Soucek, M. (2014). Overnight information flow and realized volatility forecasting. Finance Research
Letters, 11(4), 420-428.

16
Disclosures

Copyright© 2021 Salt Financial Indices LLC. Salt Financial Indices LLC is a division of Salt Financial LLC. “Salt Financial”
and “TRUBETA”, and “TRUVOL” are registered trademarks of Salt Financial Indices LLC. These trademarks together
with others have been licensed to Salt Financial LLC and/or designated third parties. The redistribution, reproduction
and/or photocopying of these materials in whole or in part are prohibited without written permission. This document
does not constitute an offer of services in jurisdictions where Salt Financial Indices LLC, Salt Financial LLC or their
respective affiliates (collectively “Salt Financial Indices”) do not have the necessary licenses. All information provided
by Salt Financial Indices is impersonal and not tailored to the needs of any person, entity or group of persons. Salt
Financial Indices receives compensation in connection with licensing its indices to third parties. Past performance of
an index is not a guarantee of future results.

It is not possible to invest directly in an index. Exposure to an asset class represented by an index is available through
investable instruments based on that index. Salt Financial Indices does not sponsor, endorse, sell, promote or manage
any investment fund or other investment vehicle that is offered by third parties and that seeks to provide an investment
return based on the performance of any index. Salt Financial Indices makes no assurance that investment products
based on the indices will accurately track index performance or provide positive investment returns. Salt Financial
Indices is not an investment advisor and makes no representation regarding the advisability of investing in any such
investment fund or other investment vehicle. A decision to invest in any such investment fund or other investment
vehicle should not be made in reliance on any of the statements set forth in this document. Prospective investors are
advised to make an investment in any such fund or other vehicle only after carefully considering the risks associated
with investing in such funds, as detailed in an offering memorandum or other similar document that is prepared by or
on behalf of the issuer of the investment fund or other investment product or vehicle. Salt Financial Indices is not a
tax advisor. A tax advisor should be consulted to evaluate the impact of any tax-exempt securities on portfolios and
the tax consequences of making any particular investment decision. Inclusion of a security within an index is not a
recommendation by Salt Financial Indices to buy, sell, or hold such security, nor is it intended to be investment advice
and should not be construed as such.

These materials have been prepared solely for informational purposes based upon information generally available
to the public and from sources believed to be reliable. No content contained in these materials (including index
data, ratings, credit-related analyses and data, research, valuations, model, software or other application or output
therefrom) or any part thereof (“Content”) may be modified, reverse engineered, reproduced or distributed in any
form or by any means, or stored in a database or retrieval system, without the prior written permission of Salt Financial
Indices. The Content shall not be used for any unlawful or unauthorized purposes. Salt Financial Indices and its third-
party data providers and licensors (collectively, the “Salt Financial Indices LLC Parties”) do not guarantee the accuracy,
completeness, timeliness or availability of the Content. The Salt Financial Indices LLC Parties are not responsible for
any errors or omissions, regardless of the cause, for the results obtained from the use of the Content. THE CONTENT
IS PROVIDED ON AN “AS IS” BASIS. THE SALT FINANCIAL INDICES LLC PARTIES DISCLAIM ANY AND ALL EXPRESS
OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, ANY WARRANTIES OF MERCHANTABILITY OR
FITNESS FOR A PARTICULAR PURPOSE OR USE, FREEDOM FROM BUGS, SOFTWARE ERRORS OR DEFECTS, THAT
THE CONTENT’S FUNCTIONING WILL BE UNINTERRUPTED OR THAT THE CONTENT WILL OPERATE WITH ANY
SOFTWARE OR HARDWARE CONFIGURATION. In no event shall the Salt Financial Indices LLC Parties be liable to
any party for any direct, indirect, incidental, exemplary, compensatory, punitive, special or consequential damages,
costs, expenses, legal fees, or losses (including, without limitation, lost income or lost profits and opportunity costs) in
connection with any use of the Content even if advised of the possibility of such damages.

In addition, Salt Financial Indices provides services to, or relating to, many organizations, including but not limited to
issuers of securities, investment advisers, broker-dealers, investment banks, other financial institutions and financial
intermediaries, and accordingly may receive fees or other economic benefits from those organizations, including
organizations whose securities or services they may recommend, rate, include in model portfolios, evaluate or
otherwise address. Salt Financial Indices has established policies and procedures to maintain the confidentiality of
certain nonpublic information received in connection with each analytical process it performs in connection with the
services it provides.

17