Master Degree Thesis Position

[Amaury Lendasse]

Dear Connectionists,
 
I am looking for a Master Degree student.  The position is described in
http://www.cis.hut.fi/~lendasse/masterdegreeposition.html
 
Dr. Amaury Lendasse
 
 
*** PLEASE FORWARD TO POTENTIAL CANDIDATES ***
 
 
MASTER DEGREE THESIS
in the field of Time Series Prediction

 
We propose a Master Degree Thesis in the field of Time series prediction and
principally for the long-term prediction. These fields are briefly described
below.

This Master Degree Thesis should be done between September 2004 and December
2005.

PROFILE: 
Applicants should have a Bachelor degree (or equivalent), in fields, such as
Computer Science, Electrical Engineering, or Applied Math, or an equivalent
or similar background. Knowledge about Neural Networks, Time Series
Prediction, or System Identification is an advantage, but not a requirement.
Ability to work in team is expected. The candidate will also have the
following desirable qualifications:
- strong programming skills (especially with Matlab);
- good communication skills in English.

APPLICATION FORM: 
Please send your application including a CV via E-mail to the contact person.

CONTACT:
Dr. Amaury Lendasse
Helsinki University of Technology
Laboratory of Computer and Information Science 
P.O. Box 5400
FIN-02015 HUT
FINLAND
tel. +358-9-451 4499
fax +358-9-451 3277
cell +358-40-770 0237
Email: lendasse at cis.hut.fi
URL: http://www.cis.hut.fi/~lendasse 
 
Time series Prediction

Time series forecasting is a challenge in many fields. In finance, one
forecasts stock exchange indices or stock market indices; data processing
specialists forecast the flow of information on their networks; producers of
electricity forecast the load of the following day. The common point to
their problems is the following: how can one analyze and use the past to
predict the future? Many techniques exist: linear methods such as ARX, ARMA,
etc., and nonlinear ones such as artificial neural networks. In general,
these methods try to build a model of the process that is to be predicted.
The model is then used on the last values of the series to predict future
ones. The common difficulty to all methods is the determination of
sufficient and necessary information for a good prediction. If the
information is insufficient, the forecasting will be poor. On the contrary,
if information is useless or redundant, modeling will be difficult or even
skewed.
 
Long-Term Prediction of Time Series 

As this problem can be found in many fields, many methods have been
developed with very different approaches, from statistics to system
identification and more recently neural networks. Most of the time, the
models are linear and perform well on a rather short-term horizon, depending
on the complexity of the problem. Their efficiency on a longer term is more
questionable. This fact is due to the learning strategy used to fit the
model to the data. The goal is usually to optimize the performance at a
given term, most often the next time step. There are only a few attempts to
explicitly predict values at long term, or at least global trends. This
problem is quite hard since the uncertainty increases with the horizon of
prediction. Another issue generally shared by classical models (such as ARX,
ARMAX, .) is that they are used to predict a single value of a scalar time
series. In practice some industrial applications require the prediction of a
set of values in one single step instead of several independent values.
Forecasting a vector of values requires more complex models able to predict
several components together. If the approach is to develop several simple
models and combine them to predict a vector, one can lose the correlation
information between the vector components. Though each model may perform
well, the forecasting accuracy could be rather poor when considering the
vector of predicted values as a whole. Developing methods able to predict
several values at each step, with the same expected performance on each
value, should thus be a major concern. Let us consider now the general
problem of forecasting at long term. Despite the fact that long-term
predictions in real situations will probably never be very accurate, in some
applications there is a need to have at least some ideas about the future of
the time series. For example, answers to questions such as "Are there bounds
on the future values?" or "What can we expect in average?" or even "Are the
confidence intervals on future values large or narrow?" can give some ideas
about the time series evolution at long-term. 
 
Publications about Times Predictions can be found in 
http://www.cis.hut.fi/~lendasse/Publications.html