# Séminaires

**Autocorrelation in an unobservable global trend does it help to forecast market returns?**

**Le : **08/06/2015 11h00

**Par : **Anatoly Petesetsky, National Research University Higher School of Economics, International Laboratory of Quantitative Finance

**Lieu : **I 103

**Lien web : **

**Résumé : **The plot of daily closing values of stock indices of dierent exchanges
reveals a common trend in their behavior. Historically, the existence of a
common stochastic trend in synchronous data was discussed in the framework
of the error-correction VAR models. The description of this approach, based
on Engle, Granger, (1987); Johansen (1991) can be found in Kasa (1992).
This approach for nding a common stochastic trend is widely used in the
literature. To avoid the problem of asynchronous trading returns, long time
intervals are considered, when this problem could be neglected. To mention
some of them: Chung and Lin (1994), Jeon and Chiang (1991) use weekly
data; Siclos and Ng (2001), Phengpis, Apilado (2004) use monthly data;
Rangvid (2001) use quarterly data. Some papers, for example Choudhry,
Peng (2007), Lucey, Muckley (2011), Bentes (2015) use daily market data.
They consider markets from dierent geographical regions. In this case it
is not clear to which point in world time the common trend is attributed.
Even high-frequency intra-day one-minute returns data are used Flad, Jung
(2008). To avoid the problem of non-synchronous data they use data only
for the overlapping hours of the US and German stock market trading ses-
sions. Error correction VAR models could be formulated in the state-space
approach (see Aoki, 1987, 1988). This approach is more
exible. Kasa (1992)
points out that in this approach a common trend can contain transitory el-
ements, rather than being simple random walks. Chang, Miller and Park
(2009) use a Kalman lter model to derive a latent common stochastic trend
from daily observations on the 30 price series of the stocks that comprise
the Dow Jones Industrial Average (DJIA). They found that the extracted
common stochastic trend resembles the DJIA quite closely up to an ane
transformation.
Korhonen, Peresetsky (2013b) suggested Kalman-lter type model, which
splits daily return of each index into two independent components local and
global. Both components are unobservable. Additionally, they suppose a
martingale property of the global component. They designed the model,
which takes into account the asynchrony of the stock indices, and estimated
it with a state-space model. This very simple model produces results in good
accordance with the observed properties of stock market returns. In our
paper we introduce model which allow for the autocorrelation in the global
stochastic trend, which means that its increments are predictable. It does not
necessarily mean the predictability of market returns, since the global trend
is unobservable. The performance of the model for the forecast of market
returns is explored for three markets: Japan, UK, US.