[ T H O U G H T
L E A D E R S H I P
close/daily volatility, as the effect
has become more nuanced. Models
predicting stocks’ intraday volume
and volatility profiles (fundamental
inputs in the cost and risk mod-
els we discussed already) can be
significantly enhanced by incorpo-
rating factors related to the relative
weight of passive ownership by
index-linked instruments and flows
in and out of said instruments.
Trading volumes have shifted
dramatically toward the closing
auction – which has an effect on
market impact, both in continuous
trading and at the close, through the
changes in intraday volume and vol-
atility dynamics. With a significant
amount of intraday liquidity now
in the close, traditional approaches
to the estimation of daily market
impact need to be reconsidered to
incorporate close impact specifical-
ly and the accompanying shifts in
higher frequency volatility. When
considering trades that would last
more than one day, the overnight
dynamics that have arisen over the
past few years pose complications
that need to be examined in detail
and lie beyond the realm of standard
algorithmic approaches, as those are
tailored to intraday trading para-
digms only (figure 1).
|
G O L D M A N
S A C H S ]
Trading costs are notoriously
difficult to estimate upfront -
traditional approaches to assessing
trade costs by market cap, ADV
bucket or exchange constituency are
limiting. The explanatory power (R 2 )
of standard market impact models
ranges between 1-3% in most major
markets and provides little a-priori
guidance of cost expectations to help
inform the investment process. By
using machine learning techniques
like (semi-supervised) clustering,
we can screen dozens of relevant
microstructure features that allow
us to group together assets with
common trading characteristics and
disentangle their non-linear depen-
dencies. This has led us to develop
13 global trade clusters, ordered by
easiest to most complicated to trade,
helped inform our algo logic imple-
mentation and equally importantly,
increased the goodness of fit of our
cost models substantially. As an
example, figure 2 depicts two stocks
from the FTSE100 index that most
algorithms would be parameterised
to treat in a similar manner, yet ac-
cording to our cluster model, should
be treated quite differentially.
Stocks A and B broken by tradi-
tional and clusters view, left and
right, respectively.
Intraday risk and correlations
vary dramatically – it is a well-
known, observable fact that volatil-
ity is higher after market open and
lower the rest of the day. What has
been less researched (and even more
difficult to implement) is appropri-
ate models to account for the sig-
nificant fluctuation in the intraday
correlation patterns amongst stocks’
performance. For example, beyond
the easy to observe spike in volumes,
the mere opening of the US market
induces a doubling of volatility and a
corresponding jump in correlations
amongst all European names in the
5/10/15/30-minute interval after
14:30 GMT.
Figure 3 depicts these shifts in
intraday correlation and volatility
for S&P500 stocks, as an illustra-
tion, over the past six years. The
observations that volatility toward
the end of the day can be a quarter
of that of the beginning of the day
and correlations can conversely be
five-to-six times higher—especially
in periods of crisis—jointly warrant a
more accurate covariance structure
estimation of a set of trades than
is broadly available. Portfolio-level
modelling of risk gives rise to risk
forecasts which are materially
different (and more accurate) than
Figure 1: Overnight reversion dynamics for STOXX600 constituents (08/2017-08/2019).
Source: GS Global Markets Division, Reuters
12 // TheTRADE // Spring 2020