

    (echo'd from the Neural Traders Private BBS)



    June/1994




    A Review of "HNET" Neural Network Software  V1.34 for S & P Forecasts
    =====================================================================

    Performed by a group of experienced investors with actual trades,
      not paper trades. HNET N/Network was used for stock price forecasting
      of US stocks.



    SUMMARY:  This N/N was able to forecast the 3-9 month market direction
              with 100% accuracy. Trading a basket of stocks (to simulate
              the SP500) using this network would have been profitable.
              Individual stocks are more difficult to predict because
              the N/N is dependent on the accuracy of the data provided.
              The trading signals were 70-80% accurate for individual
              stocks, with a 3-6 month exit window.
              The returns provided would have been much higher than
              just the SP. Also, the portfolio was "hedged" meaning
              that we took short signals as well as long signals.


    VENDOR:   AND AMERICA, tel (416) 829-0788.

    PRICE:    At last check, it was $395. Functional demo disk was $10.
              The demo disk is not a slide show. It's a functioning,
              working N/N with one-column output capability.

    VENDOR SUPPORT:  Excellent! Unlimited phone support for 90 days.


    TEST PROCEDURE:  We took historical data from "Value Line Investment
                     Survey" books available at the public library.

                     All the financial data was entered into a database.
                     For example, we entered columns for: Annual High,
                     Annual Low, Revenue Per Sh, Cash Flow Per Sh,
                     Earnings Per Sh, Div Decld Per Sh, Book Value
                     Per Sh, LTD ratio, CE ratio, etc.

                     The output row was based on the Value Line
                     estimates. Two rows were done for each year,
                     one for the "Annual High" and another for the
                     "Annual Low" stock price. We used the current
                     price for the (last) output row and asked
                     the N/N to determine if this was a "top" or
                     "bottom" or "neither" using a numeric value
                     as determinant.

                     We also used a spreadsheet to input columns and rows.
                     A word-processor can also be used to input data.



    DATA COMPATIBILITY: Takes ASCII space-delimited data in rows and
                        columns.


    ESTIMATED ANNUAL R.O.I. WITH THIS N/N:  You should double or triple
                                            the overall market performance
                                            using this N/N.



    ADVANTAGES: This N/N is unique because it took raw numerical data
                and processed it without "normalizing." Most other N/N's
                require normalizing and de-trending in order to work.
                Other N/N's also require a set range of data values;
                HNET took any value without choking or over-learning.

                HNET has unique slider bars for Memory, Learning Rate,
                Optimization.

                HNET has buttons for pre-processing data into Gaussian
                or Linear distribution. This function eliminates the
                need for tedious pre-processing (normally required by
                other N/N software.)

                The N/N predicted the output column and gave a
                graphical representation of "learning" of past
                data (or inability to learn the data, whichever
                may apply.) The graph function was exceptionally well
                done and practical.



    SPEED: Takes about 20 seconds to process a forecast on a 486-50 with
           math processor. Took about two minutes on a 386/33 with no
           math chip.


    DISADVANTAGES: Works under Windows (we don't like Windows, but that's
                   a preference.)


    BUGS DETECTED: Minor. The slider bar for memory does not work in a
                   particular spot -- but this bug is obvious, as the
                   value setting indicator doesn't move properly until
                   you find a working area on the slider bar.


    WISH LIST:     The manual could have had more application examples,
                   like Brainmaker 2.5 Pro. However, Brainmaker Pro costs
                   $895.


    CAN THIS BE USED BY A BEGINNER?  If you know how to use a spreadsheet,
                                     and a word processor, this is usable.
                                     However, if you don't know DOS
                                     commands, this will be too
                                     complicated.

                                     You will probably have to ask the
                                     AND AMERICA staff to help you with
                                     the design of your datafile, as their
                                     manual is no help in this area.
                                     Their staff were very helpful in
                                     this regard and did not shy away from
                                     lengthy discussions.


    OTHER COMMENTS:   The Value Line books are found in Toronto main library
                      and the Mississauga Central Library. An annual
                      subscription is about US$500 so it's far cheaper to
                      photocopy the pages for the stocks you follow.
                      We recommend that you test no more than 10 stocks.


    OTHER APPLICATIONS: This N/N would be excellent for "sample
                        identification" applications. We initially tested it
                        to detect "tops" and "bottoms" from a dummy datafile
                        and it was 100% successful.


  ==========================================================================

