PhD Thesis available: On Neural Networks as Statistical Time Series Models

[Adrian Trapletti]

Dear colleagues,

I am pleased to announce that my PhD thesis, titled 'On Neural
Networks as Statistical Time Series Models' is now available for
electronic download at:

http://www.olsen.ch/people/adrian/adrian.html

Abstract: This thesis provides a rigorous mathematical analysis of the
stochastic properties for what probably are the most popular classes
of neural networks for time series analysis and forecasting:
feedforward autoregressive neural networks and recurrent
autoregressive moving average neural networks. In particular, it is
shown that the characteristic roots of the shortcuts, the standard
conditions from linear time series analysis, determine the stochastic
behaviour of both feedforward and recurrent neural network models. If,
e.g., all the characteristic roots are outside the unit circle, then
the neural network models are geometrically ergodic and asymptotically
stationary. This thesis also investigates training and testing of
neural network models. In particular, it is shown that the least
squares estimators are consistent and asymptotically normal provided
the neural network model is stationary.  Furthermore, training of
nonstationary neural network models is considered. In particular, the
hypothesis test for a unit root of Phillips and Perron is introduced
as a tool to discriminate between stationary and integrated neural
network models and a new neural network based unit root test is
constructed which can be seen as a nonlinear extension of the
augmented Dickey-Fuller test.

Long abstract: in the thesis

Best regards and merry Xmas

Adrian Trapletti

--
Adrian Trapletti, Olsen & Associates   Ltd., See-
feldstrasse 233,   CH-8008  Zrich,   Switzerland
Phone: +41 (1) 386 48 48   Fax: +41 (1) 422 22 82
E-mail: adrian at olsen.ch  WWW: http://www.olsen.ch