Professional Documents
Culture Documents
Inves(ga(on
into
Correla(ons
between
Financial
Sen(ment
and
Prices
in
Financial
Markets
Markets can remain irrational longer than you or I can remain solvent.
J,
M
Keynes
(aJributed)
Market Eciency Financial Markets are generally assumed to eciently process new informa(on that is, any nancially relevant informa(on is reected immediately in the price of assets. Typically nancially relevant informa(on is taken to mean sta(s(cal informa(on about company performance, or the wider economy. Traders are believed to ra(onally interpret this informa(on and adjust their strategies accordingly.
Markets and Online Informa6on Increasingly economists and nancial researchers have found assump(ons about ra(onality or ecient pricing fail to explain empirically observed price movements. Research has begun to turn to aJemp(ng to price soK informa(on, such as sen(ment contained in text documents, into the price of assets.
Of all the weapons in the Federal Reserve arsenal, words were the most unpredictable in their consequences.
J,
K
Galbraith
(1954),
Quoted
from
his
famous
book
on
the
wall
street
crash
-
The
Great
Crash,
1929
Words
are
Poorly
Behaved
Time
Series
Objects
Although
we
have
a
wealth
to
text
data
to
draw
upon,
language
is
dicult
to
treat
mathema(cally.
Sta(s(cal
procedures
for
correla(ng
(me
series
together
rely
on
several
assump(ons
about
the
behaviour
of
each
(me
series,
all
of
which
tend
to
be
violated
by
the
nature
of
language.
Time is a Poorly Behaved Time Series Object Time means dierent things in the context of dierent systems, stock markets have closing (mes and weekly structures that may dier from online systems we wish to draw text data from. Even if we could perform good sta(s(cs with language, how would we view (me in the context of informa(on and prices?
"The only reason for time is so that everything doesn't happen at once.
Albert
Einstein
Alterna6ve
Sta6s6cal
Models
of
Language
We
begin
by
looking
at
alterna(ve
ways
of
dealing
with
language
sta(s(cally.
Looking
at
the
rela(onships
between
dierent
sets
of
words
these
rela(onships
can
be
viewed
as
set-subset
rela(onships
as
shown
in
the
gure
1,
or
as
network
rela(onships
between
words
as
shown
in
gure
2.
Fig 1
Fig 2
Alterna6ve Sta6s6cal Models of Time We also consider how best to t the two series together, taking into account the poten(al eects of weekly structure (g 4) and also poten(al dierences in correla(ons using dierent (me scales. (g 3)
Fig 3
Fig 4
So can you Predict the Market with Online Informa6on? Well yes, sort of Our results to date show there are reasonably strong contemporaneous correla(ons between online sen(ment and the price of the S&P 500. In terms of predic(on we have ini(al evidence that traders appear to perceive pre-weekend news as more risky than news during the week. As a result pre-weekend nega(ve sen(ment is magnied, causing a post weekend bounce as prices recover. Correla(ons are reasonably modest though, we can predict about 20-27% of the variance in prices at present.
Financial Sen(ment
t = 1
t = friday
t = friday
t = friday
t = n