Sie sind auf Seite 1von 2

Stocks & Commodities V.

11:14 (601-602): Product Review: N-Train by John Sweeney

PRODUCT REVIEW

N-Train
Version 1.1

SCIENTIFIC CONSULTANT gee-whiz package for spectacular color


SERVICES, INC. graphics, point-and-click interface and
20 Stagecoach Road 101 preprogrammed indicators built for
Selden, NY 11784 the masses of retail traders. This is a
Phone: 516 696-3333 workaday package of programs for a
Product: Neural net development user skilled in file handling, spreadsheet input neuron numbers proliferate. If you
package import/export, spreadsheet formulations have a $5,000 to $10,000 accelerator or
Equipment requirements: IBM com- and even a little C++. This is the real a workstation, hardware can overcome
patible 386 or better with 32-bit nu- thing, rawly presented in the cold, black the slowdown. If not, N-Train speeds
meric coprocessor. DOS 3.0 or higher. screen of DOS. If a slowly blinking cur- up computation by eliminating all the
No hard disk drive required if there is sor waiting for your typed commands spreadsheet overhead and having you
enough space on floppy disks. Cannot gives you chills, if picking up hints of do the preprocessing and file transfer
run under Windows or OS/2, but a Win- problems from the way DOS displays a work. A neural network beginner would
dows version is to be released soon. file or flashes an error message isn’t your probably choose something like
Price: $747 for N-Train; $159 for cup of tea, read no further. Neuralyst to begin with and then gradu-
TradeNet module. ate to N-Train when size, speed and
MANUAL LABOR MINIMAL sophistication needs increased later on.
by John Sweeney Scientific Consultant Services (SCS) has
given more than adequate guidance in a WORKING, WORKING...
t looks as if the neural net idea plain but thorough manual replete with With N-Train, there is no waiting. The

I may not follow the fads of the


past: more and more of these
products are appearing — and
selling. Where I could once recall off the
thorough, accurate explanations and an program is very, very fast. If it’s not,
exhaustive run-through of a real world there’s even a specially modified cus-
development situation. Exemplary C++ tom-designed board you can slap into
code is provided where needed, usually to your personal computer that will accel-
top of my head two or three trading- show how to preprocess your data files, erate computations to warp speed.
oriented neural network packages, my though most but not all (given N-Train’s Setting up for a run just involves
latest count showed 12. Of these, one that ASCII file setup) of that can be done in some typing with its usual potential for
is prepared for heavier duty is N-Train, your spreadsheet. No index is provided, errors. Once you know the game and the
which supports fast, full 32-bit addressing but the table of contents and the manual’s regions you want to explore, you’ll start
as well as 64-bit numeric processing for organization make it useful and usable. using batch files to tweak all the param-
accurate adjusting of weights while build- The major omission is a table of error eters and rerun the training.
ing networks. message explanations. As a result, you The sequence is setting up the data
One nifty feature of N-Train is the get to figure out what “Se-
TradeNet module. This program sup- quence error reading user fact
plies the hook between TradeStation file!” or other such cryptic This is a workaday package for
(2.0 or higher) and N-Train. You can messages mean — or you get a user skilled in file handling,
see the neural net’s indicator directly in to call to find out. spreadsheet import/export,
TradeStation’s graphics, plus Trade- Nearly all neural packages spreadsheet formulations and
Station is able to set up the fact files (the require you to preprocess the
raw data) that the neural network will data for presentation to the even a little C++. This is the
use to come up with its solution. Though package. The more sophisti- real thing, presented in the
this does not eliminate the interface cated have built their packages cold, black screen of DOS.
issues with N-Train, it is an order of as add-ons to spreadsheets so
magnitude more efficient and effective that the data already resides where the files (for programmers, an example of C
for a trader than doing things in DOS or neural add-on can grab it, and the add-on code is supplied), specifying the scaling
working from within a spreadsheet. knows exactly where to store the results. for the data (necessary since most nets
N-Train itself is not your modern This is convenient but slow, especially as operate best in the range of -1 to +1),

Copyright © Technical Analysis, Inc.


Stocks & Commodities V. 11:14 (601-602): Product Review: N-Train by John Sweeney

that it worked well. Then again, most


users we talked to were professionals,
deeply familiar with their machines,
operating systems and competitive soft-
ware. Here are some of their comments:

• “This is a package for someone who


likes to tinker, not a beginner.”
• “It’s fast. And accurate. It’s the 32-
bit capability.”
• “It can handle huge data sets. I’ve
given it 25,000 facts all the time.”
• “I went back to BrainMaker six
months ago. I was just more used to
the interface.”

In addition to this, we received some


suggested user advice:

• “The seed you give the network has


FIGURE 1: N-TRAIN OUTPUT TO LOTUS. Taking a crack at forecasting the closing price of IBM, N- a huge impact on the result, so
Train does a fine job at first, but then the error of the forecast begins to increase, indicating the need experiment a lot with it.”
for one or two more independent variables. Graph is by Lotus 1-2-3. N-Train has no graphic capability. • “Don’t feed price-based indicators
to networks and expect tomorrow’s
price. Instead, use the indicator to
trigger the neural network’s esti-
mate of the situation.”
loading up the training and testing data, stable patch and start oscillating back and • “Don’t feed indicators to nets at
specifying the depth and width of the forth between two sets of node weights, all! The weights show they are just
network you want, setting the param- a tribute, probably, to its bit depth. I’d filtered out in favor of raw price
eters, training, checking results, training be hard put to compare N-Train’s abil- action. Instead, increase the offer-
some more, tweaking things and repeat- ity with other neural net packages be- ing of truly independent variables.”
ing the whole process. Results informa- cause network development is an art of • “Preprocessing the data and finding
tion is limited during the training run to blending the investigator’s intuition the right data seem to have more of
an impact on the effectiveness of the
correlation coefficients and mean error. about the data and skill in preprocess-
results than anything else.”
You can also ask for a text file to be ing it with the package’s capabilities
dumped that shows expected vs. actual and parameter sets. There are too many CONCLUSION
output. These can be fed to your spread- variables there for even a neural net- Once it comes to decision time, N-
sheet for error analysis. The entire pro- work to sort out! Train’s pricing is competitive with
cess requires familiarity with the envi- other comparable programs. If you’re
ronment, text editors, spreadsheets or C. USERS’ COMMENTS comfortable with the DOS interface,
Once you try to feed it numbers daily, Most users we contacted like the give N-Train the nod for its speed in
you’d better be fluent, not just familiar. program’s no-holds-barred efficiency, handling huge data piles and its 32-bit
You say you want to try this intraday? stability and reproducible results (it precision.
Not without TradeStation. saves the entire net and the seed†, so
you can come back to where you started, John Sweeney is the Technical Editor
RESULTS if need be). One user had even ported it of STOCKS & COMMODITIES.
N-Train converged quickly in all the ex- to UNIX and a workstation to work with TA
amples I gave it. Never did it hit an un- a particularly large data set and found

Copyright © Technical Analysis, Inc.

Das könnte Ihnen auch gefallen