Using NeuroShell Classifier for
Fundamental Stock Selection
There are some major companies who make their stock
selections every day using the NeuroShell Classifier. Although we helped these
companies and are non-disclosed to tell you who they are, what we taught them to
do is very easy, and anyone can do it.
The hardest part of the
process is obtaining good, clean, fundamental data on a fairly large number of
stocks. We are not talking about just price data or transformations of price
data like moving averages and stochastic indicators. That sort of analysis is
called technical analysis and is best done with our
NeuroShell Trader software.
We are talking instead about such fundamental factors as revenue growth, long
term debt, insider trading, dividend yields, etc.
Many companies provide this
data for a fee, with varying degrees of cleanliness. One of the best sources as
of this writing, according to some of our customers, is the American Association
of Individual Investors (AAII) www.aaii.com.
If you join you get access to their database, which has dozens of fields for
thousands of stocks, and the price is very reasonable.
You’ll need to collect data
on at least several hundred stocks, and that data needs to be at least 3 months
old. That’s because the neural net will need to evaluate how well the stock has
done over the last 3 months (or you can use 6, 12, whatever you want) in terms
of the fundamental data available at the beginning of that time.
You load your fundamental
data into a spreadsheet, each stock on a different row. You can have several
rows for the same stock if each row represents fundamental data on that stock at
different quarters. The columns contain the fundamental data variables (you can
use technical variables too), and you
create one column that describes how the stock performed over the next 3 months
(or 6, 12, whatever). You may need to clean up bad data, or eliminate fields
where the data is suspect or largely missing.
Next you start training nets
with the NeuroShell Classifier. This may be an iterative process, where you
choose to eliminate some variables that the Classifier tells you are not useful.
You may make more variables, using calculations on the existing ones. When you
are done, you have a model that, given recent fundamental data, gives you
probabilities that each stock will perform well in the next 3 month period (or 6
At each stage of this
process, you will have free access to our experts to help you.
Here is a sample spreadsheet
which gives you a general idea of how NeuroShell Classifier training data would
be laid out. For clarity, we have not included the numeric fundamental
data. Also, the actions on which the net is trained (Buy, Sell, Hold) are
hypothetical actions which you would enter using any criteria you wish based on
the stock's future movement some number of months after the indicated year and
quarter when the fundamental data was recorded.