Академический Документы
Профессиональный Документы
Культура Документы
2. What's the difference between historical actual volatility and GARCH volatility?
Short Answer: Both historical actual volatilities and GARCH volatilities can be thought of as forecasts of
future actual volatilities. GARCH forecasts are typically the more accurate of the two. Historical actual
volatilities are estimated over a specified period of time and give equal weight to each fluctuation in the
sample. The ability of GARCH to train itself over historical data to find statistically optimal weights to apply
to its forecasts leads to better out-of-sample performance than the more simple-minded straight weighting
approach.
Long Answer: Actual volatilities are sample standard deviations of underlying spot returns. Each
fluctuation in the sample is given equal weight. Since FX volatilities tend to vary over time, weighting all
fluctuations equally can result in actual volatilities that may not reflect the market environment at the end
of the sample period. Reducing the length of the sample period may reduce this effect but it increases the
uncertainty in the estimate due to small sample size. GARCH estimates use longer spans of data but also
anticipate that the volatility may change over time. By attaching more importance to recent results,
GARCH accounts for volatility clustering within the sample period. The model parameters are determined
by maximum likelihood estimation, tailored to the data in question. If we view the true population volatility
as a time-varying process then it makes sense to use a model like GARCH, which attempts to capture
this time evolution explicitly.
Long Answer: It depends on the convergence properties of the model estimation. A typical rule of thumb
is that 3 or 4 years of daily data will suffice. It can be less and can be more depending on whether the
parameter estimates from the optimisation routine are well behaved. With too little data the parameter
estimates may be imprecise. It should also be noted that the observations should be sampled at regularly
ab2
spaced intervals. If some data points include 3 trading days of market movement and others includes just
one; GARCH will garble the proper weightings when giving its one-day forecast. Gaps in the data can
cause problems as well since a lagged value used in the model specification could include a gap that
should not be there.
Long Answer: Implied volatilities hope to anticipate price fluctuations over the life of the option and can
foresee some market-moving events, but not all. In fact, they are statistically more closely related to the
recent historical volatility than to the actual volatility that ultimately ensues over the remaining term. For
that reason, they correlate rather highly with GARCH forecasts since GARCH is driven by the same
fluctuations. That being said, there is information that can be imputed in an implied volatility that a
standard GARCH model cannot use. For instance, if a G7 meeting, or Fed announcement, or influential
economic figure has the potential to move markets, option prices can reflect this. GARCH models, unless
they are modified to include indicator variables like the ones just mentioned, are unable to use this
information. GARCH, however, has the advantage of trained dynamics and optimised weightings on
previous fluctuations. In fact, GARCH models often win forecast tournaments against implieds. An idea
we're pursuing is combining information from both implieds and GARCH into a kind of composite forecast.
Preliminary evidence shows that the hybrid performs very well.
Long Answer: If GARCH parameters are updated in an automated system with initial values as inputs,
it's typically best to use the previous period's estimates as a starting point in the optimisation. If
convergence fails, different starting points should be tried. If problems persist, a different historical sample
is likely to work.
Long Answer: In real life market volatility is determined by the actions of individual participants in the
market. These in turn are determined by their views of the market, trading strategies and goals. Since all
of these participants interact in a complex and dynamic way it is a hopeless task to predict their effects on
the underlying in a deterministic way. One approach to this is to ignore these complexities and model
volatility as an evolving random variable. If we do this, each volatility in time is assumed to come from a
statistical probability distribution. GARCH characterises this by explaining or predicting the return
variance, the square root of which is standard deviation or volatility.
ab2
7. Could we use an exponentially weighted sample standard deviation to forecast
volatility?
Short Answer: Indeed this would, in general, produce a better forecast of volatility than a simple equally
weighted sample standard deviation would. However, the exponential decay is not as flexible a functional
form as GARCH can be.
Long Answer: Exponentially declining weights have been used by RiskMetrics to forecast volatilities.
The parameter dictating the exponential decay is often held fixed across time and across products.
GARCH models are typically more customised to the data. They're updated frequently and differ by
product. With the richer parameterisation, GARCH models are also able to fit the observed historical
fluctuations better. A greater variety of relative weights are available through GARCH.
Long Answer: Yes, but with problems. The volatility of implied volatility is less stable than that of spot
data. This causes problems for GARCH models at the parameter estimation stage. The problems of
convergence to good parameter estimates crop up more often.
Long Answer: Even with the most efficient optimisers, the maximum likelihood function requires a time
consuming recursive run through the data to incorporate the lag effects. Initial values for the parameter
estimates are provided, the likelihood function is evaluated, search directions are determined for the next
iteration, and the process continues until convergence to the optimum is achieved.
Long Answer: GARCH, as the acronym implies, generalises the ARCH model. The effect is to simplify the
number of parameters needed to build a proper model of the conditional variance by accounting for
information not only in the lagged squared deviations as in ARCH, but in the lag(s) of the conditional
variance itself. In most cases, one lag of each is sufficient. In contrast, ARCH models often require quite a
few more lag terms to explain the same data. The parameter weightings applied to these ARCH lag terms
often must be constrained to ensure positive values and smooth decays, whereas the GARCH model is
parsimonious and well-behaved by nature.
11. What do GARCH models imply about the distribution of spot returns?
Short Answer: A model like GARCH, which features varying volatility over time, implies a more fat-tailed
distribution than the normal.
Long Answer: The usual assumption made by option pricing models like Black-Scholes is that the spot
returns are normally distributed. The empirical distributions, however, are almost always fatter in the tails,
meaning that more extreme moves are more common than the normal curve would suggest. GARCH
models are consistent with fatter tailed return distributions. Since a GARCH volatility may be 12% one
day and 10% the next and so on, varying through time, the returns drawn from those distributions of
varying widths, when aggregated, are non-normal.
ab2
12. What are the more advanced models for?
Short Answer: The more advanced models are intended to explain additional features of the data that
GARCH itself may not capture.
Long Answer: Though a complete review of all the extensions to GARCH would be infeasible, there are
a few innovations that stand out for their potential in capturing observed market behaviour. One
interesting tangent is called EGARCH. This model allows the possibility that up moves and down moves
can have differing effects on the volatility that is likely to follow. Others recognise inherent nonlinearities in
the response of vols to spot moves. Different distributional assumptions can also be employed to account
for errors in the predicted variance equation that are fatter-tailed than the normal distribution would
suggest.
Answers provided by Jonathan Roberts, Currency Management and Advisory Service
ab2
This material has been prepared by UBS AG, or an affiliate thereof ("UBS"). In certain countries UBS AG is referred to as UBS SA.
This material is for distribution only under such circumstances as may be permitted by applicable law. It has no regard to the specific investment objectives, financial situation or
particular needs of any recipient. It is published solely for informational purposes and is not to be construed as a solicitation or an offer to buy or sell any securities or related financial
instruments. No representation or warranty, either express or implied, is provided in relation to the accuracy, completeness or reliability of the information contained herein, nor is it
intended to be a complete statement or summary of the securities, markets or developments referred to in the materials. It should not be regarded by recipients as a substitute for the
exercise of their own judgement. Any opinions expressed in this material are subject to change without notice and may differ or be contrary to opinions expressed by other business
areas or groups of UBS as a result of using different assumptions and criteria. UBS is under no obligation to update or keep current the information contained herein. UBS, its
directors, officers and employees' or clients may have or have had interests or long or short positions in the securities or other financial instruments referred to herein and may at any
time make purchases and/or sales in them as principal or agent. UBS may act or have acted as market-maker in the securities or other financial instruments discussed in this material.
Furthermore, UBS may have or have had a relationship with or may provide or has provided investment banking, capital markets and/or other financial services to the relevant
companies. Neither UBS nor any of its affiliates, nor any of UBS' or any of its affiliates, directors, employees or agents accepts any liability for any loss or damage arising out of the use
of all or any part of this material.
Options, derivative products and futures are not suitable for all investors, and trading in these instruments is considered risky. Past performance is not necessarily indicative of future
results. Foreign currency rates of exchange may adversely affect the value, price or income of any security or related instrument mentioned in this presentation. Prior to entering into a
transaction you should consult with your own legal, regulatory, tax, financial and accounting advisers to the extent you deem necessary to make your own investment, hedging and
trading decisions. Any transaction between you and UBS will be subject to the detailed provisions of the term sheet, confirmation or electronic matching systems relating to that
transaction. Clients wishing to effect transactions should contact their local sales representative. Additional information will be made available upon request.
United Kingdom and rest of Europe: Except as otherwise specified herein, this material is communicated by UBS Limited, a subsidiary of UBS AG, to persons who are market
counterparties or intermediate customers (as detailed in the FSA Rules) and is only available to such persons. The information contained herein does not apply to, and should not be
relied upon by, private customers. Switzerland: These materials are distributed in Switzerland by UBS AG to persons who are institutional investors only. Italy: Should persons
receiving these materials in Italy require additional information or wish to effect transactions in the relevant securities, they should contact Giubergia UBS SIM SpA, an associate of
UBS SA, in Milan. South Africa: UBS South Africa (Pty) Ltd (incorporating J.D. Anderson & Co.) is a member of the JSE Securities Exchange SA. United States: These materials are
distributed by UBS Securities LLC or UBS Financial Services Inc., subsidiaries of UBS AG, or solely to US institutional investors by UBS AG or a subsidiary or affiliate thereof that is not
registered as a US broker-dealer (a "non-US affiliate"). Transactions resulting from materials distributed by a non-US affiliate must be effected through UBS Securities LLC or
UBS Financial Services Inc. Canada: These materials are being distributed in Canada by UBS Securities Canada Inc., a subsidiary of UBS AG and a member of the principal Canadian
stock exchanges & CIPF. Hong Kong: The materials relating to equities, corporate finance and other securities business, and related research, are being distributed in Hong Kong by
UBS Securities Asia Limited. The material relating to foreign exchange, fixed income products and other banking business, and related research, are being distributed in Hong Kong by
UBS AG, Hong Kong Branch. Singapore: These materials are distributed in Singapore by UBS Securities Singapore Pte. Ltd or UBS AG, Singapore Branch. Japan: The materials
relating to equities, fixed income products, corporate finance and other securities business, and related research, are distributed in Japan by UBS Securities Japan Ltd. The materials
relating to foreign exchange and other banking business, and related research, are distributed in Japan by UBS AG, Tokyo Branch. Australia: These materials are distributed in
Australia by UBS Advisory and Capital Markets Australia Ltd and UBS Securities Australia Ltd, licensed securities dealers. New Zealand: These materials are distributed in New Zealand
by UBS New Zealand Ltd.
2004 UBS. All rights reserved. UBS specifically prohibits the redistribution of this material and accepts no liability whatsoever for the actions of third parties in this respect.
UBS Investment Bank is the global investment banking and securities business group of UBS AG