(July 11th, 2005)
Some knowledge of Visual Basic as well as prior acquaintance to neural networks
are necessary to use NXL2.
NXL2 is now released (v1.132). It is basically a neural net library (API) using
an algorithm close to QuickProp (Scott Fahlman), which can easily be called from
MS Excel or MS Access. It is composed of over 40 different functions catering
for most neural net applications (predictions and classifications).
Each function is documented separately on the ForeTrade site (http://www.foretrade.com/nxl.htm),
but the best way to start with NXL2 is to run the provided Excel spreadsheet
samples, and read the VBA code included which is relatively simple.
Neural net VB code:
The general VB code structure is as the following:
Set number of inputs, outputs, patterns
This will allow NXL2 to construct data structures when a new neural net is built
Check the lib is enabled (Demo or Full mode)
A personal 12 character long code will be generated upon valid registration
Generation of a neural net architecture and involved data structures in DLL
Change some training parameters
Default parameters set while building can be changed at this stage
Load Data in memory (training data and often test data)
Train the net
Loop through training data, and check for error level on training or test data.
Release memory inside the DLL
Absolutely essential, otherwise no new net can be created, and memory will NOT
be deallocated with closing Excel. If anything else, a ReleaseNet call should be
added to the workbook close event.
One important feature to remember is that the library uses dynamic arrays so
that it can be used with virtually unlimited number of inputs and outputs (in
full mode). Memory is allocated at the time of building the neural net (BuildNet)
and must be released with ReleaseNet when the neural net is no longer needed. It
is well described in the accompanying spreadsheet code (QP2.2.xls). Failing to
release memory may cause memory leaks, which are errors often hard to trace, and
would eventually crash Excel. It is possible that a really really huge neural
model exceeds the DLL memory capacity, but it has never been reached in any of
Most variables (input, output, pattern number) passed to the library are
"1-based", and reset to 0 inside the library. This means for instance, and
inputs go from 1 to NumInputs, patterns range from 1 to NumPatterns. The only
exception is the SaveNet/LoadNet functions which call weights starting with
weight no 0. Again, this is well commented in the VBA code.
The data provided in the Excel sample is a fairly complex non-linear function,
to be mapped by the neural network. NXL2 does a very good job at modeling that
Like in most neural net applications, the required accuracy influences
significantly the processing time (network learning).
A basic "early stopping" technique is implemented. It stops learning when the
gradient of the error curve (on Test data) comes close to nil. This means that
learning can then only improve almost insignificantly.
Training is calibrated (i.e. training data is presented several times before
testing the net on unseen data), in other words, NXL2 validates the network
against test data every so many epochs, or training batches. An epoch is the
presentation of the entire training set to the neural network while training.
Calibrating generally improves learning time, however, difficult (noisy data or
sparse patterns in data set) problems may require to lower the Calibration
parameter to test the net more often.
Data is scaled in Excel using a Map function provided in the VBA code. Mapping
is often essential in neural applications, to level the impact of inputs against
one another. it is also possible to apply data squashing, or clip statistical
Saving/Loading a net, Exit Conditions:
It is possible to store/retrieve the net to/from a separate worksheet in the
same workbook for instance. It is usually good to associate SaveNet with the net
training exit condition so that the best neural net is saved. Indeed, the error
level often deteriorates (on test data) after reaching an overall minimum. With
the supplied data, the function to map is relatively simple, so the net can
almost always improve if one lets it carry on forever, if not stopped after a
number of training epochs.
The provided code stores node weights. It is left up to the user to also
store/retrieve other parameters.
You can copy/paste or program indicators using Excel or VBA functions, and use
NXL2 to predict a %change in Open or Close, and to predict the future value of
an indicator. In many trading experimentations, classifications seem to bring
about better results. In any case, the quality of the neural model depends
essentially on the quality of data preprocessing. The provided sample proves
that the neural nets do a great job on sample data mapping a pattern that is
present on every record. It is NEVER so easy in trading applications, and
whatever most neural net salespeople will tell you, this neural library does
perfectly what it is designed to do. The rest is left on your hands, i.e. find
good inputs, output(s), and preprocess them the best way you can.
The library will run in demo mode until activated for your computer. A code can
be generated (c.f. additional Excel menu). Registered clients will receive their
personal code to fully enable their neural library. In demo mode, nets are
limited to 300 training patterns, and 15 nodes (including input and output
nodes). Once activated, the NXL2 library has virtually no limits (cf. ForeTrade
site for latest price details).
Assistance in building your own neural solutions will be charged at reasonable
rates: $35/week email support (firstname.lastname@example.org). Support can however be
VERY time consuming, particularly when it comes to using an API, so this implies
the question is 'standard' (so the answer also is), and that the user has done
his homework with sample code.
Dedicated development is charged from $30 to $100/hour according to complexity
Latest information will posted on the Foretrade web site.
Thank you for your interest in NXL2