|
This system is the oldest version
of Neurogest, completed in January 2003.
Neurogest 1.1 is a mixture of expert networks
in which a given temporal series of data
(stock, index, currency etc...) is segmented
into a series of sub-periods. These sub-periods
are trained on separately by 20 networks
(Recurrent Backprop and Radial Basis Functios
networks).
 |
The input to the system is described
in the figure on the left. The networks
predict the direction of variation that
will occur the next business day by
being exposed to a set of inputs comprising
not only the given security that is
the object of the forecast, but also
other index/currency/interest rates
that might influence the direction of
the market. These raw input data are
further modified by applying some Techincal
Analysis tools commonly used by institutional
investors. This
data is then fed to the different networks.
The output of these networks is further
refined by three distinct voting procedures,
and ultimately
combined into a single decision. This
decision is delivered in the form of
an estimate (positive or negative) of
the direction of variation and the
strength (or confidence) of the prediction.
|
The forecast
that is published on this page is the results
of multiple systems (trained with different
parameters and in different time frames) whose
decisions are pooled into a final consensus.
For more information, feel free to contact
us. Disclaimer |
