Trading on the Edge/Trading On The Edge - CD-ROM Toolkit (Wayzata Technology)(2031)(1994).bin/pc/mac_file/vendor_d/neurofor/quicksta.rt

Text File | 1992-08-15 | 26KB | 566 lines

__________________________________________________________________
NeuroForecaster¿ QuickStart Instructions: (Version 2.05, August 1992)
__________________________________________________________________
Print this file for reference.  Also read the User's Guide and the
reference book "Neuro Fuzzy Computing For Business Applications" (to
be published by Prentice Hall) for more details.

__________________________________________________________________
Initial Setup Information
__________________________________________________________________
NeuroForecaster¿ is a user-friendly neural network & fuzzy logic
program capable of mimicking human thinking, learning the
movement patterns of stock, currency and property markets, GDP,
and also the rating of credit, selection of stocks, etc, by itself.
It uses only historical data for trainining, and no explicit expert
rules are needed. It is an ideal tool for building forecasting and
decision-support systems, as well as for indicator analysis.  Its
user-friendly interface allows you to build such systems easily,
without having to memorise any command.

Ñ Purpose:
(1) time series forecasting,
(2) cross-sectional classification,
(3) indicator analysis

Ñ Methods used:
It has 12  built-in neural network paradigms, including the BackProp
and RBF models, and the proprietory FastProp and Neuro-Fuzzy
models.  The reference book describes the network structures and the
theoretical background.

It accepts numerical input attributes, patterns, codes, technical
indicators and fundamental indicators, and allows them to be
combined.   However, it cannot handle descriptive information such
as news,  rumors or fiscal policies, unless such information is
accompanied by quantitative data.

Ñ Before You Start...
You need to know how to use the keyboard and the mouse, how to
choose a command from a menu, how to open a file, activate an
application, etc.

Ñ Application Examples:
The software has been tested with these real applications:
(1)  stock index forecasting (Weekly,Monthly,6 Monthly)
(2)  stock market return forecasting (Monthly, 6 Monthly)
(3)  stock price forecasting (Weekly)
(4)  stock selection
(5)  foreign exchange rate forecasting (Daily, Monthly, 6 Monthly)
(6)  GDP forecasting (Quarterly)
(7)  property price valuation
(8)  credit rating
(9)  air passenger arrival forecast (Yearly)

Some of these applications are included in this package.  To run
these applications, simply follow the 5-step procedure described
below. The software can also be applied to other types of forecasting
and classification problems, as well as indicator analysis.  The
data file structure is explained in details in the User's Guide.

Ñ System Requirements:
Macintosh¿: Mac SE and Mac II series, System 6.0.2 or later, at least
1MB free RAM, and a hard disk with at least 2MB contiguous free
space.

MultiFinder and various monitor sizes are supported, but Single
Finder mode is preferred, especially if the RAM memory is limited.
For a clear display of text, make sure that Geneva12,14 and 18 fonts
are installed in the System Folder.  For larger problems, more memory
space will be required.

Ñ Backup Procedure:
Before you do anything else, be sure to make copies of your
NeuroForecaster¿ diskette.  To do this, first write-protect the
original diskette and insert it into the floppy disk drive, double
click to open the NeuroForecaster¿ icon and drag the NeuroForecaster¿
folder to a formatted diskette on the second drive (if there is no
second drive, first drag the folder to the hard disk then to a
formatted diskettes on the floppy disk drive).  Now keep the original
original diskette in a safe place.

Ñ Installation Procedure:
NeuroForecaster¿ MUST be run from the hard disk. Insert the
NeuroForecaster¿ diskette in the disk drive, double click to open the
diskette icon and drag the entire NeuroForecaster¿ folder to the
harddisk.

Ñ What's On The NeuroForecaster¿ folder:
a. The NeuroForecaster¿ program
b. The DATA Folder which contains all the data files
c. The FORECASTER Folder which contains all the forecaster files
d. The RESULT Folder for the result files
e. The QuickStart file which contains this QuickStart procedure

__________________________________________________________________
To Start NeuroForecaster¿:
__________________________________________________________________
Double click the application "NeuroForecaster¿" to run it from the
hard disk.

Copyright Message:  If you agree with the copyright message, press
the mouse button to proceed.  When the title page appears, press the
mouse again.

__________________________________________________________________
The Quck Start 5-Step Instructions:
__________________________________________________________________
Follow the instructions associated with the 5 buttons on the main
screen:

Step 1: Load data from the "DATA" folder
Click the "DATA" button to load a data file from the "DATA" folder.
The data file must be loaded from the "DATA" folder.

Step 2: Load forecaster from the "FORECASTER" folder
Click the "FORECASTER" button to load a forecaster file. Again,
the forecaster file must be loaded from the "FORECASTER" folder.

Ñ Note:  A common mistake here is to load the forecaster file from
other folder especially the "DATA" folder.

Select "Hidden Nodes.." from the same menu if you want to view or
change the number of hidden nodes.

Step 3:  Show Ranges Using "CHART"
Click the "CHART" button to view the data just loaded.

ÑShow Actual:  You can view the actual target (the values to be
forecasted) and the accompanying indicators (the attributes or
factors that influence the target values) by selecting "Show Actual"
and "Show Indicators" from the "CHART" menu.  The indicators can
also be displayed by pressing the "COMMAND" and "I" keys
simultaneously.

ÑScreen Size:  If you have selected the wrong screen size, you will
not be able to see the full chart.  To re-size the screen, select the
correct screen size from the "SCREEN" menu from the menu bar.

ÑShow Range:  Click the two "Show Range" buttons on the left-bottom
corner to show the data ranges used for training and forecasting.

ÑClose Chart:  To proceed to the next step, click the "Chart Close"
button from the chart screen, or select "Chart Close" from the
"CHART" menu.

....more on charting functions....
ÑRanges:  Do not change the ranges for the time being, because the
forecaster just loaded has been trained for that learning range.  If
the ranges are changed, further training will be needed.

The range informatioin (Learn From, Learn To, Forecast From,
Forecast To) also appear in the left window in numerical forms.

ÑPattern:  The initial charting pattern is line chart, which is
suitable for both screen display and printing;  however, the charting
pattern can be changed by clicking the "Pattern" button on the left
side of the chart.  Click the "Show Range" buttons again to check the
ranges selected for training and forecast.

ÑZoom:  The horizontal (X) and vertical (Y) scales can be zoomed in
and out.  Click the X (or Y) button, and then click the zoom in, zoom
out, zoom max, zoom min, zoom normal buttons to size the scales
accordingly.  Use the "Show Range" button to show the range if
necessary.

ÑScroll:  To view the portion of the chart outside the screen, use the
"Scroll" slider on the scroll bar at the bottom of the chart.  To
scroll the chart, move the cursor to the "Scroll" slider, press the
mouse button and move the mouse to the right or left, then release the
mouse button.

ÑOther Chart Functioins:  You may also refresh, reset or close the
chart by choosing the menu items from the "CHART" menu.

If you computer is linked to a printer, you could plot the chart by
clicking the "Print Screen" button.  To choose the printer, select
"Chooser" from the "CONTROL" menu;  to setup the page, select "Page
Setup" from the "FILE" menu.

Close the chart and return to the main screen now.

Step 4:  Learning
Click the "LEARN" button from the main screen.  You may proceed by
clicking "LEARN NOW" to begin the learning process using the pre-set
learning parameters.  You may also click the "Options.." button to
view and change the learning parameters.

The learning method and parameter settings can be viewed and
changed by clicking the "Learn Model" and "Options.." buttons
respectively.

For training tips and selection of neural network models, read the
User's Guide and reference book for more details.

ÑSTART:  Click the "START" button and the learning process will start.
The red (or gray on monochrome monitors) line represents the actual
target, and the blue (or the thick dark) line represents the forecast.

After the error has dropped to an acceptable level, the forecaster is
said to have "learned" the data, and the learning process can be
stopped by clicking the "STOP" button.  To resume, click the "Resume"
button. To exit learning, click the "Exit"  button.

There are 3 ways to stop the learning process by:

(1) visually comparing the actual (red) and the forecast (blue) and
the error values, and decide when to stop.  The absolute errors,
percentage errors and percentage errors are displayed on the screen.
You could click the "Testing" button to check how well the forecaster
has learned.  The Error Trace window also displays the percentage error
for the most recent 70 iterations.

(2) specifying the maximum iteration count from the "Options.."
button.  The training process will stop automatically when the
maximum iteration count is reached.

(3) specifying the stopping error value from the "Options.." button.
The training process will stop automatically when the stopping error
is reached.

Exit and return to the main screen.

Ñ Learning Frequency: Before you proceed to the next step, it is
advisable to make sure that the forecaster has learnt the entire
training samples in the learning range.  To do that, return to the main
screen, click the "CHART" button, select "SHOW ACTUAL" to display
the actual target, and select "Show Learn Frequency" from the "Chart"
menu.  If the learning is not evenly distributed across the entire
learning range, continue the learning.

Ñ Hidden Nodes' Learning Status:  Refer to the section on "Hidden
Layers and Hidden Nodes" on how to check the learning status of the
hidden nodes.

Step 5:  Testing/Forecasting/Classificaton
Now click the "FORECAST" button.  You may click the "FORECAST
NOW" button to proceed, or click the "Options.." to specify the file
name for storing the forecasted results.  Be sure to click the radio
button if you wish to store the forecasted results.  The forecasted
results will be stored in the "RESULT" folder with the filename
specified by you.

ÑSTART:  You may click the "START" button to perform forecast for
the entire forecast range, or click the "STEP" button repeatedly to do
single step forecast.

....more on forecasting....
The forecasting range overlapped with the learning range is for the
purpose of testing the performance of the forecaster for data it has
learned, and the forecasting range beyong the learning range is to
test the performance of the forecastor for data it has not learned.

If the learning range is carefully selected and the forecaster is well
trained in Step 4, it is able to forecast or classify for the
forecasting range, as illustrated by the examples provided.

At the end of the forecast, click the "Chart.." button to display the
forecast (or "EXIT" to the main screen and click the "CHART" button
from the main screen, then select the "Show Actual" and "Show
Forecast" from the "Charting" menu).

Click the "Pattern", "X", "Zoom In", "Zoom Out" and "Zoom Normal"
buttons, as well the "Scroll" slider to view the chart.

The chart can be printed by clicking the "Print Screen" button.

Now close the chart and return to the main screen.

If you have specified the optioin to store the forecasted results, you
could display them by selecting "Show Forecast File" from the
"FORECAST" menu.

You may also view the forecasted results by selecting the "Open.."
from the "FILE" menu, and open the file from the "RESULT" folder.
The "FILE" menu also allow you to print the forecasted results.

__________________________________________________________________
Accuracy:
__________________________________________________________________
If the result is not satisfactory, go back to Step 4 to continue the
learning process.  You may have to lower the "Learning Rate" using
the "LEARN" "OPTIOINS".

If the result is acceptable, train the forecaster up to the latest set of
data.  To do this, return to the main screen, click the "CHART" button,
select "SHOW ACTUAL" from the "CHART" menu.  Click and hold on to
the "Learn" slider and move it to the end of the data.   Do the same
for the "Forecast" slider.  Click the "Show Range" buttons to confirm
that the appropriate ranges have been selected.

Please see the Chapter on "Important Notes on Accuracy"  in the
reference book on factors affecting and techniques used to improve
the accuracy.

In general, classification problems or cross-sectional analysis can be
learned relatively easily and the performance is expected to be good
if the right set of indicators are used.  On the other hand, time-series
problems are more difficult because of the time-varying nature of the
problems, and you have to carry out  several experiment to find the
best forecast horizon, window size, number of input rows and the best
network paradigms.

To ensure that the forecaster is adequately trained, use the following
functions:

"Update & View Hidden Nodes' Learning..."
"Show Learn Frequency"
"Show All Min/Max"
"Show Distribution of Training Data"

__________________________________________________________________
Indicator Analysis:
__________________________________________________________________
Another important function performed by the forecaster is indicator
analysis.  It extracts two sets of indices during the learning process
(the Accumulated Error Index, or AEI) and after the learning process
(the Causal Index).

ÑINDICATORS:  To display the indices, select  "Update Indicator
Indices" and "Show Indicator Indices" from the "FORECAST" menu.
The indices are also stored in the "Result" folder, and can be viewed
and printed by selecting "OPEN.." from the "FILE" menu.

The indices show the significance of the various indicators.  A higher
index value means that the associated indicator is of higher
significance compared to others.

One could therefore start with a larger set of indicators, and obtain
the best combination of these indicators according to the significance
of their indices.

__________________________________________________________________
Additional Information:
__________________________________________________________________
Creating a new forecaster (Step 2)
__________________________________________________________________
To create a new forecaster, click the "FORECASTER" button and select
"New".  Use the default or enter appropriate values for the following

ÑNumber of input columns (number of indicators)
ÑNumber of hidden layers (for the network)
ÑNumber of input rows (window sze)
ÑNumber of steps ahead (forecast horizon)

The number of input indicators is automatically precomputed using
the information provided by the data file.  Therefore it is advisable
to first load the data so that this value is automatically determined.

Other fields may take on any reasonably large values.
NeuroForecaster¿ is capable of building large neural networks with
unlimited capacity (any number of input nodes, hidden layers,
hidden nodes, window size, forecast horizon).  The amount of free
RAM memory space is the only limit.

However, larger values would rapidly take up a lot of memory space
and slow down the computation tremendously, and sometimes may
cause the computer to hang if your computer does not have sufficient
RAM memory.

The number of input nodes is automatically calculated based on the
number of input indicators and the window size specified.  If you use
the default values, the number of hidden layers, hidden nodes,
window size, forecast horizon

ÑNumber of hidden layers
Usually 1 is sufficient; 2 or 3 for complex problems.  You probably
never have to use more than 3

ÑNumber of input rows (window size)
For classification or cross-sectional analysis such stock selection,
credit rating, property valuation, the window size is fixed at 1.  For
time series forecasting, a window size greater than 1 will always
yield better results, especially for markets that exhibit long-term
memory.  A larger window size is a good way to capture temporal
information contained in the time series.  If the indicators already
contain such temporal information, as in the case of some technical
indicators such as stochastics and moving averages, one could reduce
the window size to 2 or 1 to save memory space and speedup the
training.

ÑNumber of steps ahead (forecast horizon)
Sometimes it is necessary to do long-term forecasts, due to the delay
in obtaining the indicators or for planning purposes.  Several of the
examples show how to do 6-months ahead forecast.

Depending on the types of indicators used, the accuracy varies with
the forecast horizon.  Usually it deterioriates rapidly when the
forecast horizon increases.  To determine which is the best forecast
horizon, one has to train the forecaster for, say, 1-, 3- and 6-steps
ahead forecasts.

__________________________________________________________________
Saving the forecaster (Step 2)
__________________________________________________________________
A newly created forecaster needs not be saved upon creation.  You
only have to save a well-trained forecaster.  When a forecaster is
saved, its network parameters, learning parameters and the set of
weights (representing the knowledge acquired from training) are
stored in a file so that they can be reloaded the next time.

If the file name is already existing, NeuroForecaster¿ will prompt
you to use a new file name, or confirm to over-write the existing one.
Click the "Save & Exit" button to confirm, or "Cancel" button to
cancel. For a newly created forecaster, click "Exit" without saving
the forecaster.
__________________________________________________________________
Learn Options.... (Step 4)
__________________________________________________________________
When a forecaster is newly created,  the parameters have been set to
their optimal values.  Most of the time you only need to change the
Learning Rate (eta), say from 0.9 initially to 0.1 during the training,
and if the output is oscilatory, select Non-linear output instead of
Linear output.

Ñ Training Patterns:
Sequential:  when selected, the forecaster will learn the data points
in a sequential manner.  This training pattern is not recommended.

Random:  when selected, the forecaster will learn the data points in a
randomly selected manner.  This training pattern is recommended.

Ñ Save forecaster automatically on every XX iterations
When selected, the forecaster will be saved automatically after every
XX iterations.  This feature is useful if you want to leave the
training unattended for a long period of time, or in case of a power
failure.

Ñ  Learning Rate (eta):
This value controls the amount of current error for adjusting the
weights.

For stable systems, use a large value, eg. 0.9, initially, and reduce
it to 0.1 subsequently.  For oscillatory systems, use a smaller value,
eg. 0.6 or 0.3, and reduce it 0.1 subsequently.

Follow the examples given on the control of this value.

Ñ  Momentum rate (alpha):
This value controls the amount of previous error for adjusting the
weights.  It stablises and helps smooth out the training process and
prevents unusual cases (outlying data) from throwing the training off
track.

For problems with consistent, smoother data, the momentum rate can
be set to 0.

Ñ  Maximum Iteration, Stopping Error:
You could specify when to stop the training by setting the maximum
iteration count or the stop error.  The current iteration count and
stopping error are provided on the right for your reference.

Ñ  Input/Output Normalisation:
Two types of input normalisation can be carried out:

Column-wise: normalise each and every input indicator to XX % of its
maximum and minimum values;  Row-wise: normalise all input indicators
by the maximum and minimum in the current window.  The output from the
forecaster is normalised to YY % of the maximum and minimum values of
the target, and can be either linear or non-linear.

__________________________________________________________________
Hidden Nodes and Hidden Layers
__________________________________________________________________
Once a forecaster is created, its number of hidden layers is fixed, but
the number of hidden nodes for each hidden layer can be changed.

To change the number of hidden nodes, click the "FORECASTER"
button and then the "Hidden Nodes..." button.  You may now change
the numberof hidden nodes for the current hidden layer, or click the
"View Next Hidden Layer" button to switch to the next hidden layer,
and then change the number of hidden node for that layer.

ÑHidden Nodes' Learning Status: To view the learning status of each
hidden node, select "Update and View Hidden Nodes..." from the
"FOROECASTER" menu.  If a hidden node does not learn well during
the training, its AEI (Accumulative Error Index) will be very much
lower than that of other nodes, and will affect the accuracy of
forecasting or classification.

To remedy this problem, switch to another learning model, or rebuild
a new forecaster with fewer hidden nodes or hidden layers.

__________________________________________________________________
The MasterKey:
__________________________________________________________________
Ñ Precautions:
The MasterKey diskette should NEVER be write-protected.

Never attemp to read, write or duplicate the Masterkey provided.
Failure to do so will cause unrepairable damage to the MasterKey
diskette.

Ñ Full version: To run the full version NeuroForecaster¿, insert the
MasterKey and click the "VERIFY.." button.  The main screen with 5
buttons will appear.

Ñ Demo:  To proceed without the MasterKey, choose "DEMO" from the
screen or press the RETURN key.

If it is run as a DEMO only, the Verification button also appear on
the right bottom corner of the screen.  You can convert it into
the full version by clicking this button.

The full version allows you to create new forecasters and to save
them.  Subsequent use of the same forecaster does not require the
Masterkey diskette.  It is recommended that you create several
forecaster networks with different configurations and save them for
later use.

Insert the Masterkey ONLY when requested.
__________________________________________________________________
User's Guide and Reference Book:
__________________________________________________________________
Please refer to the User's Guide and reference book for matters not
covered in this QuickStart procedure, and the prerequisits and
techniques for achieving better accuracy.  Follow the application
examples provided for illustrations of how to:

(1) set up data files;
(2) select neural network models;
(3) do time-series forecasting for stocks and forex, etc.;
(4) do cross-sectional analysis including stock selection, property
valuation, credit rating etc.;
(5) analyse indicators.

__________________________________________________________________
TRADEMARK NOTICE:
Microsoft, MS/DOS, and Windows are trademarks of Microsoft
Corporation.  IBM, PC/XT, PC/AT, PS/2 and PC/DOS are are
trademarks of IBM Corporation.  Apple, MacIntosh SE, MacIntosh II,
MacIntosh IIci, MacIntosh Quardra are trademarks of Apple
Computer Inc.

DISCLAIMER NOTICE:
NeuroForecaster¿ is no substitute for real thinking or common sense.
The user should always review and check the results of
NeuroForecaster's processing and evaluate it against know references
and standards. The use of NeuroForecaster for investment, speculation,
gambling, or other similar or related purposes is at the user's risk.
Results generated by NeuroForecaster¿ are dependent on past
information and there is no guarantee that future results can be
forecast or predicted by NeuroForecaster¿.  Trading in stocks,
commodities and other securities or any form of speculation or gambling
is inherently risky and may result in loss.

LICENCE AGREEMENT:
This NeuroForecaster¿ software is licensed to you and you may only
use it under the terms of the software licence Agreement.  You are
granted a paid-up, non-transferrable personal licence to use it on one
computer only.  You do not have the right to copy or alter the
software, its data, user interface and any of the materials provided.
You are accountable for any violation of the Licence Agreement &
Copyright, Trademark or Trade Secret law.  The software developer
and distributors are not liable for any damages from the use of the
software & the accompanying information."

__________________________________________________________________
Contacts:
__________________________________________________________________
For upgrades, application examples, User's Guide, reference book
and MasterKey, please contact:

Francis Wong at Fax:[phone], Tel:[phone].
NeuroForecaster¿ Version 2.05.
Copyright ⌐ 1989-1992, ISS.  All rights reserved.