Home

General

What it does

Instructions

Tour

NEW:Try
it online

Registration as freeware

Test your copy of GGem

Downloads

NEW: Auto-train feature

Contact

loading requested page...

Fast download - about 30 seconds on a modem (2 MB)

Requires Windows 98/ME/XP/2000/Vista/7

GoldenGem^®

Neural network calculator ~ Stock market data source

Beginner's Intro

The GoldenGem console -- freeware neural network

predicted

General Information about Neural Networks

Background Information (from Wikipedia):

Since the early 90's when the first practically usable types emerged, artificial neural networks (ANNs) have rapidly grown in popularity. They are artificial intelligence adaptive software systems that have been inspired by how biological neural networks work. Their use comes in because they can learn to detect complex patterns in data. In mathematical terms, they are universal non-linear function approximators meaning that given the right data and configured correctly, they can capture and model any input-output relationships. This not only removes the need for human interpretation of charts or the series of rules for generating entry/exit signals but also provides a bridge to fundamental analysis as that type of data can be used as input. In addition, as ANNs are essentially non-linear statistical models, their accuracy and prediction capabilities can be both mathematically and empirically tested. In various studies neural networks used for generating trading signals have significantly outperformed buy-hold strategies as well as traditional linear technical analysis methods. While the advanced mathematical nature of such adaptive systems have kept neural networks for financial analysis mostly within academic research circles, in recent years more user friendly neural network software has made the technology more accessible to traders (read full article). [also follow the same link to the more important section on fundamental analysis]

Wikipedia, September 2006

Theory of Neural Networks (J. Moody)

Summary of operation:

The trader, wishing to quantify the relationship among a group of stock or share prices, and/or indices, enters the tickers in capital letters, separated by commas.
The needed historical and real time share price quotes and volumes are looked up and compared automatically.
The neural network searches for a nonlinear mathematical relationship (pattern) relating the prices and volumes to the ticker of interest, while the user participates by controlling a sensitivity (also called 'momentum') adjustment
When sensitivity is then set to zero, graphs show two years of correct and rigorous backtesting. through which the user may visually assess whether the relationship is valid throughout historical time.
The relationship is extended into the future to make a forecast, by the number of days the user has set on the slider during training.
There is no buy/sell indicator: the reliability of the forecast depends on the user's visual verification of the match between the two graphs obtained during backtesting, and the his estimation of the likelihood that the mathematical relationship which has been found will continue to hold in the future.

Now that you have read an outline summary of the operation of a neural network, try clicking on the 'Instructions' link on the left, and have a go memorizing the actual instructions. The technical operation of the program should be no more difficult than using a CD player, for example.

(If things go well, you'll have an opportunity to try them out on a mock-up practise tour, and if that goes well, you might even end up going online to try the real thing.)

User instructions

1. Think of a list of tickers that are likely to be mathematically related, over time. This is the hardest step -- mathematically related does not mean that they behave similarly. (see the Wikipedia article GoldenGem - Wikipedia).

2. Locate the label which says 'Related Group of Tickers.' Beneath this, delete the sample list of tickers which may be there, and enter the names of the tickers you are interested in analyzing, in capital letters, and separated by commas.

3. Choose a data source from the File: button at the top of the program. The simplest possibility is 'Load from Internet.' The (included) StockDownloader will appear. This program knows how to search the standard list of public domain websites including Google, Yahoo and MSN and place a plain text file called Internet.txt in the GoldenGem folder, which will have columns for Date, Ticker, Open, High, Low, Close, and Volume, with entries separated by commas. The default settings in StockDownloader are set to load the same list of stocks which you entered in GoldenGem, up to the most recent close price. If you are happy with this selection, press the 'Load' button on StockDownloader. (The StockDownloader is just a beginners' tool; it is expected that you will later graduate from technical analysis to fundamental analysis finding your own data from other websites;

click here to find out how to do this.)

4. Use the up and down arrows on your keyboard to choose which of the the various graphs you wish to see (or you can use the drop-down menu button, near the letters TK to the right of the screen).

Click to continue

5. You will see three coloured traces on each graph. The red trace is the graph of the selected share price throughout two years of historical time. The green graph shows the neural net's use of artificial intelligence to predict the red graph, throughout historical time, simultaneously combining all loaded volume and price information of all the shares. There is also a blue graph, which shows where the green graph would go if it were predicting perfectly.

6. Increase the sensitivity and wait until the green graph matches the blue graph, then decrease the sensitivity to zero (the bottom of the slider). This is best done in stages, over many iterations, as you are removing the training input. If the green and blue graphs match well when sensitivity is zero, this means the green curve has learned how to predict the advance copy of the red curve throughout historical time, by looking only at effects as many days in the past as you have selected on the Days slider. The green curve does not therefore need to finish at the 'Loaded Until' date, but can continue the calculation for an equal number of days yet to come. You see this in the first screenshot as the part of the green curve that extends three weeks past the verticical blue line.

7. The numerical prediction for the future increase or decrease will also be shown at the upper right of the display. A text display will say, for instance: Predicted change over the next 15 business days: +2.35.

Once you have memorized the instructions, then you are ready for the tour. Click the link on the left which says 'tour.'

For more detailed instructions go to the FAQ page (click here)
For technical specifications to click here


Other data sources: Besides using StockDownloader, there are other options for obtaining data. For three beginning examples, you can
	*choose 'Load from File,' to load data which you may have downloaded previously, such as the standard and poor 500 file sp500hst.txt which uses the same standard file conventions as GoldenGem, and can be obtained as a file called full_set.zip from biz.swcp.com/stocks/
	or
	*download a collection of .csv files (some of stock market data and some of Forex data) from for example www.forexrate.co.uk/forexhistoricaldata.php. The csv data export on that site works by scrolling with the up and down arrow keys. Select 500 data points, 1 day, and split by comma. Change the vol number in GoldenGem to column 4, enter the ticker names you wish into GoldenGem, and choose Import from a folder. Then browse to whatever folder contains the .csv files.
	or
	*paste or type in a text file of your own using Wordpad, as explained in the link below
	or
	*download a file of 2 years or so of daily data from the Bank of England Statistical database http://www.bankofengland.co.uk/statistics/index.htm Choose columnular (with titles if you wish). This creates a file called results.csv. Set the file position numbers tick, date close to 2,1,3 and leave the vol window blank. Type in the subset of ticker names you want to start with. Choose 'Browse for a new file' from GoldenGem's File button. Choose 'comma separated (csv) files' at the bottom of the file selecting box and select 'results.csv,' the file you've just downloaded from Bank of England.

(Click here for even further methods of importing and loading data from other applications)

Availability:

Try it online!

If you have a Windows XP operating system, with Internet Explorer or Mozilla you may find it is not necessary to install the program at all; you're invited to click here to try it online. For Internet Explorer choose 'run' instead of 'save'. For Mozilla choose 'save' then 'open.'

The link above is a small 504 K file called GoldenGem Viewer. Choose to 'load tickers from the internet' from the File: menu, then use the menu to the right of the screen to change between graphs. If it is the first time you've trained a neural network, think of it as a sort of video game, where the goal is to make both indicator lights remain green.

Full installation:

Local download Link

The installation is also at the very safe download site Download.com

Freeware version:

Auto-training version:

The program can also be found at PCMag.com (PC Magazine), by including the words PCMag and GoldenGem in a Google search.

The installation file setupv.exe (or setupva.exe for the automatic training version) is digitally signed by the original software author J Moody. The signature is recognized by a Microsoft Root Trusted Certificate, which verifies that the installation file has not been altered. If you use Internet Explorer, you have the capability to verify the certificate and the digital signature, and verify the following:

File Name: setupv.exe	File Size: 2.05 MB	Signed by: J Moody	Date Signed: April 24, 2009
Counter-signature: Comodo	Certificate: UTN	Time signed: 4 PM GMT

The Download.com site, which delivers the same file, verifies that software which they offer is free of adware, free of spyware, free of viruses, and that it correctly installs and completely uninstalls. It does not matter whether you download from the local link or from Download.com. The program works with Windows 98, 2000, NT, XP and Vista.

Registration as freeware:

The website and program are not for profit, and there is no charge at any time for the permanent use of the program. Google advertising revenue supports the website and delivery of the program. There is a non-free automatic training feature avaiable also which can be purchased elsewhere on the site.

Each time you start the program it will ask for a registration number. You may just close the registration window, or leave it blank and the fully-enabled program will appear. After 30 days you'll be asked to return to this website for a registration number. Write to the email address on your console or to goldengemnetwork@gmail.com and your registration key will be sent to you from England for free. If users have time, they will visit the Google ads below for competing products. Also when users write to request the free registration key, they should include advice for improving the program.

Here is an outside review of the program, really intelligently written
Download from an external mirror site with (sort of) independent review

Ads about competing Neural Networks, predicting the stock market, Math, Technical Analysis, Bonds, Exchange rates etc:

Recreational/educational from GoldenGem.co.uk

Portfolio analysis(free)
Visualize the output of a single neuron w/o bias (freeware)
Visualize a neural network with 2 inputs, 2 hidden neurons (freeware)
Terminal for writing in Tex (math) notation (freeware)
Terminal for kids, says letter and word sounds (shareware)
Which extra-solar planets could you reach -- special relativity (freeware)
Word unscrambler using fundamental thm of arithmetic (freeware)
Complex plane --educational, requires Mozilla (freeware)
Reuters news Ticker (freeware)

Test your copy of GoldenGem

We've made a file of tickers named x, y, z,w containing

10 sin(i/10), 10 cos(i/10), 10 cos²(i/10)sin(i/10)+10 cos(i/10), and 10 sin ²(i/10)-10 cos(i/10)-10.

for i=1 to 400. Right click proof2.txt and choose 'save target.' Set the ticker names x,y,z,w and file position numbers 1 (blank) 2 (blank) as in the screenshot below. Press the File: button, select 'Browse for new file' and locate the file proof2.txt which you have just downloaded. Then have fun!

The fact we've used repeating functions is just so the viewer can recognize whether the predictions are correct. The program does not use the fact that they are repeating functions in any way. Note that since the DAYS slider is set to 21, the predicted change is given 15 data points ahead, as the DAYS slider is in terms of actual days not including week-ends, so it will state that the prediction is for 15 business days in the future.

Graph of neural network's third prediction

neural network successfully predicts chart 4

Extra comment for the expert.

Set the sensitivity to zero immediately upon loading the data, and observe either x or y. Note that for some positions of the DAYS slider the indicator lights will both be green without even training the neural network, and for some positions the first light will be red. The reason is this: a sinewave shaped graph of period n days admits a perfect trading strategy already: if the average value is A, predict that in n/2 days it will return to the average value A. If the current value is A+c it will in fact take the value A-c so your prediction is half the actual change. The correlation between half the actual change, and the full actual change is 1. Our adjusted coefficient, which in most other cases punishes this strategy, in this case does not, because in this rare case when actual variance is built in, the variance of predicted changes is actually smaller variance of actual changes. Since predicted variance is twice actual, the adjusted coefficient is 2, and so the first indicator light measures the minimum min(1,2)=1 and detects a perfect strategy of buying or shorting in proportion to the negative of the current value. But until the net is trained, the two graphs do not intuitively look alike, and the prediction does not look believable; indeed the effectiveness of the trading strategy depends on, and measures, nothing but the cyclical nature of the graphs; whereas, once the neural net is trained, then again both lights are green, but now one sees a sound reason why; and what is detected now is the form of the relations between the functions expressed in the equations above.

The auto-train feature

A version of the program called AutoTrain is available, which rapidly trains automatically. In this version the speed button is replaced by an 'autotrain' button. Because training is facilitated, the condition for the leftmost light to be green is that the r value is 0.65 instead of 0.6 for the ordinary version. The preformance of the autotrain version is noticeably better and attains r values of .7 or higher. To download a trial version of AutoTrain go to http://www.goldengem.co.uk/setupva.exe

If you enjoy the auto-train feature and wish to register it after one month of use, this will cost 75 us dollars; at the end of one month of use a registration key for AutoTrain can be obtained through a paypal order by using the button below.

The theory of neural networks

Mathematically, the theory of neural networks is fairly trivial. That said, it is also true that the development of the theory would have proceeded more quickly if basic mathematical understanding had been applied at the beginning.

We assume one knows that a linear map

Rⁿ → R^m

is given by a matrix with n columns and m rows, whose entries are real numbers. A composite of such linear maps is again a linear map, and the matrix representing the composite is the product of the matrices representing the separate factors. Therefore, the set of functions which can be represented as a composite of linear maps is no larger than the set of functions which can be represented by a single matrix.

Just to make things concrete, if we attempt to find a function which determines the price of a share of IBM, in terms of the prices and volumes of five stocks, using a linear function, we are choosing one function out of a ten dimensional space of functions; or, we are choosing the ten entries of a matrix with a single row and ten columns. One notion of what those entries should be is the one which gives the least error, in the sense of least squares. Provided the share prices are normalized to have mean zero and standard deviation one, these ten numbers are what are called the linear regression coefficients.

The new ingredient in neural networks is that after applying a matrix to a vector, we then apply a transition function to each entry of the new vector. If repeat what we have just done, the result will not be much different. Some of the differences are that it is not guaranteed that there is a sequence of ten coefficients that gives the best fit, two sequences of ten might give an equally good fit. Also, it will be impossible to fit values which exceed the range of our transition function. These seem like disadvantages.

But there are tremendous compensating advantages. Namely, a composite of functions of this type (applying a matrix, then applying a transition function to each entry of the answer) is not the same as a single function of this type. For example, the composite of linear maps goes (if five shares are loaded) and there are ten neurons in the middle layer

R¹¹ → R¹⁰ → R¹⁰ → R⁵

To make a fair comparison, reagarding the calculation of one output, we are only looking at one of the factors in R⁵. So the number of matrix entries used for the same calculation is

11 × 10 + 10 × 10 + 10 × 1=220

So instead of a ten dimensional space of functions, we are looking at 220 dimensional space of functions.

Next, in order to simplify things, let us write all the various matrix entries as a sequence of variables

y₁, y₂, ..., y_m

where m is a number which may range up to, as we have seen, 220. And, let us write the input variables as x_1,...,x_n where in our example of five shares, n is just ten. In fact, n is eleven because we use a bias variable which is the constant input one.

Now, our neural network, or, the part of it which calculates a single output variable, is just a function

f(x₁,...,x_n; y₁,...,y_m).

We would like this function to match the actual share price of the share we are analyzing at a fixed number of days into the future, and we shall call the correct answer, when it is known, g(x_1,...,x_n). So our error, of course, is the difference

e= g(x₁,...,x_n)-f(x₁,...,x_n ; y₁,..,y_m).

The second term, if we were to write it down, would look complicated, it would involve composites of the transition function and matrix multiplications, and, if we include the normalization of variables it would look even more compli- cated. We shall now imagine a particular point in the training of the neural network. So x₁,...,x_n are fixed, they are par- ticular numbers. This means the first term g(x₁,...,x_n) is a number, and the second term is a rather complicated looking function involving all the thousands of variables y₁,...,y_m. But we may still ask, what is the best direction to change the y_i to improve the error the fastest. This is merely the value of e times the gradient of f viewed as a function of the y_i. The entries of the gradient of f are just the partial deri- vatives of f with respect to the y_i. Because f is a composite of separate functions Rⁱ to R^j for various values of i and j, the chain rule can be used, which says that the gradient of f is the product of the Jacobian matrices of these separate functions. Each function is in fact a matrix composed with a function that acts separately on each coordinate by the transition function. The Jacobian matrix of a matrix is the matrix itself, and the Jacobian matrix of the transition function acting on each coordiante separately is a diagonal matrix, with the derivative of the transition function in each diagonal entry. Some care must be taken to evaluate the derivative of the transition function at the appropriate value, but this is standared in multivariable calculus.

Let us now make this explicit. We have a composite

g₁ g₂ g₃
R¹¹ → R¹⁰ → R¹⁰ → R
and
f=g₃ o g₂ o g₁

a composite of three functions. For i=1,2,3 the function g_i is a composite M_i o h_i where M_i is a linear map, given by a matrix, and

h₁: R¹⁰ → R¹⁰
h_2: R¹⁰ → R¹⁰
h_3: R→ R

Note that h_i acts on the target of g_i.

The functions h merely apply the transition function to each variable.

Now, the derivatives of the h_i just apply the derivative of the transition function to each variable. Let us call these h_i'. We have

f=h₃ o M₃ o h₂ o M₂ o h₁ o M₁

so the chain rule says that the derivative of f evaluated at our fixed vector (x₁,...,x_n) is

h₃'( M₃ o h₂ o M₂ o h₁ o M₁(x₁,...,x_n)
o M₃(h₂ o M₂ o h₁ o M₁(x₁,...,x_n)
o h₂'(M₂ o h₁ o M₁ (x₁,...,x_n)
o M₂(h₁(M₁(x₁,...,x_n)
o h₁'(M₁(x₁,...,x_n)
o M₁(x₁,..,x_n)

Here we can view the h_i' once they are evaluated at the long expressions in parentheses, as square matrices with only the diagonal entries nonzero.

Now, I actually want to differentiate f(x₁,...,x_n) with respect to the variables y₁,...,y_m which are the entries of the matrices M₁,M₂, M₃ keeping the x_i as fixed numbers.

The calculation is less familiar but actually easier. To differentiat with respect to the (i,j) entry of M₁ replace the last term in the expression above with a column vector with x_i in position j. If I want to differentiate with respect to the (i,j) entry of M₂ then the last two terms will be removed, and the third from last term will be replaced by the i'th entry of h₁(M₁(x₁,...,x_n)) in position j and so-on.
This is just an application of the chain rule viewing the matrix entries as variables. The same calculation is also the same as the the 'backpropogation' calculation for artificial neural networks. To finish we must apply some multiple of the error -- actually GoldenGem uses also a sensitivity multiplier based on a logarithmic scale here. The particular transition function which we use is the inverse tangent function arctan. This is chosen because the other possible choice of a bipolar transition function is hyperbolic tangent. The hyperbolic tangent takes values between -1 and 1, and is almost always nearly equal to one of those values. It is often chosen if a neural network is meant to model digital data. Arctan varies more evenly, and is the better choice of bipolar transition function for analogue data. Why the neural network configuration --- this particular space of functions --- is chosen is easy to see. If the matrix entries in early stages are chosen small, it matches a linear function well, so it is in this sense at least as good as linear regression.

On the other pointer, it is possible to model a binary function very well, even with three levels. To see this, think of the output which ranges from -π /2 to π /2 as a digital output, thinking that an answer near pi/2 means 'yes' and one near the negative of that means 'no.' Now we can very easily choose matrix entries to arrange any truth table we wish. If I want an answer of 'yes' just if the first seven inputs are 'yes,' I can weight the matrix entries directly looking at the first seven inputs so that the sum of the contributions from these entries (plus a constant from a bias neuron) is positive if and only if all seven inputs say 'yes,' and in any case a very large positive or negative number. And so that the contributions from other inputs do not matter, I set other coefficients to zero. After applying the transition function to the sum, we see that one neuron in the middle level has an answer of π/2 if all seven are 'yes' and -π/2 otherwise. In this way, one sees, in the first stage I can model an 'and' of any subset of entries, or their negatives. Now, it is an easy fact of logic that any truth table can be expressed in two stages of negation and and. The more familiar fact is that any truth table can be expressed as an OR of a set of ANDS of the input variables and their negations. But then one can use the tautology that

OR = NOT AND NOT.

This shows that a three level neural network can model both linear and logical phenomena. A more precise fact is any continuous function whose domain is a bounded subset of Rⁿ can be approximated uniformly by a sequence of functions f₁,f₂,... where f_i is realized by a three level perceptron with i neurons in the hidden (middle) layer. This was first discovered by Funahashi (in the Journal Neural Networks, vol 2) and independently by Hornik et al in the same journal. Hornik et al applied the phrase `universal nonlinear function approximator' to this three level configuration, and this is the default configuration in GoldenGem.

In conclusion, the chain rule is an element of under- graduate mathematics, it was not applied in neural network theory for many years. When it was applied, it was called `backpropogation,' because the chain rule takes account of all the values and derivatives of the separate functions we have composed to create f.

Another concluding remark: in data mining, some people run several neural networks in parallel. This is very much the same as running a single neural net with a very large middle layer, but it may be better, because there is a concern about overtraining and poor generalization if a single three layer perceptron is used with a large middle layer. For that reason the number of neurons in the middle layer of GoldenGem by default is no more than the square root of the product of the number in levels 1 and 3, and this means it may be useful to run many copies of the program simultaneously, or run the program many times. With this in view we have added a 'speed' button which allows rapid training, and quick indicator lights which can be used to assess statistical significance of the training set, and an adjustment in the stock downloader which allows one to choose a validation data set.

It is my view that the challenge now is not better software, but better use of the existing algorithms. I hope I have successfully brought the GoldenGem console into the 21'st century, and it will remain a reliable slide rule of neural calculation in the field.

Advantages:

Benchmark predictions of abstract mathematical functions, such as shown on this site, have been extensively verified.
The technical specifications are those agreed to be most effective in stock market prediction.
The algorithm has been widely used for many years in finance, trading and investment, portfolio management. It is established as a reliable and valuable calculation.
The algorithm is the only reliable way in which it is possible to simultaneously consider the combined effects of a number prices and volumes.

Contact Us

Postal Address: GoldenGem Neural Networks, 12 Armorial Road, Coventry CV3 6GJ, England
Telephone (if dialled from USA): 01144-2476-523490
Software author (if dialled from USA): 01144-2476-417946
Support: GoldenGemNetwork@gmail.com

GoldenGem neural network

Beginner's Introduction

Currently, the site is visited by programmers, serious investors, students (mostly Masters and PhD), some emeritus prof.'s and hobbyists, most of whom are already involved in neural networks.

We would like to change that, and we are working on making the site and the program more accessible.

Starting on this page, the instructions in blue will guide you through, and will tell you exactly where to go at each stage. The questions and answers below should give you a starting point.

Questions and Answers

What are the main types of financial analysis?

There are two main types. Technical Analysis attempts to extrapolate a price using only earlier values of that same price. Only a few papers show a statistically significant advantage over random trading.

The second type is Fundamental Analysis, where data from financial statements, interest rates, volumes, competitors' prices, prices of raw materials or other variables known to affect the target prices are used.

As explained by Wikipedia (see the General Information link on the left), neural networks act as a 'bridge' between Technical Analysis and the more highly regarded Fundamental Analysis. Future values of prices are approximated, not by nonlinear extrapolation of earlier values, but by hypothesizing and testing actual causal connections.

It is possible to load a set of prices and volumes of the most well-known shares, bonds, and indices using StockDownloader, by pressing one button. But unless one has the skill to find a meaningful relationship among these variables, StockDownloader is really only for beginners. A competent user will graduate to using other types of data, available on the internet and also accessible by GoldenGem with a bit more work; the instructions link on the left leads to a series of links explaining further types of data that can be imported, and we're willing to modify the program to make it compatible with new formats as they come up.

Developing this implementation over time we have needed to confront subtle questions before the algorithm began to work well, as it now does. The Verification link on the left shows GoldenGem predicting abstract mathematical functions. The program does not use the fact that the functions are repeating; only today's values of the input variables are used in making the last day of prediction. In cases when earlier values than the last known value may have an effect you should also include `stochastics' as input variables. The program does not take into account 'oscillations' such as described in the Elliot wave 'theory,' a theory which we do not subscribe to; however, note that the program is able to predict functions that oscillate. Finally, the program is able to predict any one of the functions knowing the others, but there is no preferred process of extrapolation that should would work for a single function treated by itself.

How reliable is a prediction that this program gives me?

We have added a pair of indicator lights to help answer this in each case. The first indicator light refers to a number r during backtesting, which is defined to be the minimum of the correlation coefficient of predicted versus actual daily changes, on the one hand, and the correlation coefficient times the ratio (actual variance)/(predicted variance), on the other hand. To give some idea of the meaning of this: if a person were to buy or short, every day during backtesting, in proportion to its predicted percentage change, assuming things are normally distributed, the percentage profit over n days of trading during that time in the past would have been

(percentage gain over n days of backtesting) =

r (

)1/2

(

365

)1/2

where V is the maximum of actual and predicted annual volatility expressed in percentage points.

This taking of V to be the maximum of actual and predicted volatility seems contrived, but it is exactly what one wants. If r and V were defined using only the actual volatility, then a strategy of relying upon a posteriori information would exist to attain deceptively good returns during backtesting which do not really result from any prediction: a sluggish response, in which the green curve stays near the 2 year average, would correspond to a strategy of predicting always a sudden return to the 2 year average, which during backtesting includes knowledge of future days and would unfairly reward the correlation coefficient alone. Whereas if r and V were defined using only the predicted volatility, there would be no intrinsic relation between r and the actual percentage gain: a large r value could arise from a prediction with very low variance. The value of r as we have defined it rules out both these problems, and appears to correspond with what looks like intuitively good backtesting.

The first light is yellow when the r value is larger than 0.39 and green when it is larger than 0.6. The second light goes from red to yellow to green as the training input is removed. You will need to try different combinations of input variables before you will be able to make both lights remain green at the same time. If the lights cannot be made to remain green, the answer to your question is, the prediction is meaningless. If the lights do remain green, then that means a relationship has been found which has been able to make succesful predictions during the backtesting interval.

When both lights have remained green, does this imply the prediction can be trusted? Not yet. Even taking account the variance ratio as we have, the formulation in terms of profit shows that this number could be high enough to set the green light, just because some of the hypothetical trades were extremely profitable, others not at all. You also need to to actually look at the behaviour of the prediction line, the part of the green line extending into the future, past the red line, throughout backtesting, and see qualitatively how consistenly it is correct. When sensitivity is set to zero there is no training input, and the green graph is calculated only using data values of all variables from the time of the earlier red graph, and any prediction you see therefore shows a real mathematical relationship during backtesting.

Finally, you still are not quite done. Even when you have assessed both statistically and visually that the predictions throughout backtesting are good, to be really sure the variables you are looking at are related, you should set 'today's date' in StockDownloader to a time in the past, or otherwise load data only up various times in the past, and train the net to predict a range of values which you actually already know. This is a 'validation data set,' and the next version of GoldenGem will make this last stage of validation easier.

There do exist relationships between variables which are known to affect prices, the ones which are well-known can't be exploited unless you have knowledge of the input variables advance of the trading public. It is not true that all existing relationships are well-known. Insider trading is perfectly legal if you exploit public domain information through your own intelligence.

What will happen the first time I try it?

A good training strategy is to start with a high sensitivity, and to bring it down in stages. Assuming the variables actually were expected to be related, and your training strategy was correct, you are likely to end up with the first light turning red, signifying inadequate correlation, by the time the second light turns green. This is usually for one of three reasons:

1. If you see vertical green spikes, and a message 'Press the Reset button,' then you have traumatized the net. Like a human or animal, it will take a very long time to recover. Similar to a good nights' sleep might do for an animal, the Reset button gives a fresh start, and all is forgiven, but it will need to be trained again from the beginning.

2. If the green line is flat, this is because it was trained inadequately. Raise the sensitivity slider again and wait a while before bringing it down (ideally in stages).

3. If the green line looks just like the red line, but shifted to the right by the amount on the Days slider, you are seeing a situation where the expected future value is always nothing but the last known value. If all graphs are like that, congratulations, you have found a 'Markovian' set of shares: interesting but with no opportunity for arbitrage, assuming the neural network has found the best possible solution.

An automatic training version is also available which helps avoid some of these pitfalls.

Conclusion

Financial analysis is something you do, not something you buy.

A neural network requires involvement by the user. You have to choose what data you think is relevant, you have to learn how to train the net, and it is up to you to evaluate the backtesting. The most important thing to remember is that although our display shows only two graphs at a time (actual and predicted), the predicted graph is generated by taking account of the mathematical relations among the prices and volumes of all loaded variables simultaneously and so the choice of ungraphed tickers affects the quality of the match between the two graphs you are observing.

If you think you are ready to walk through the process of training up a neural net, welcome aboard.

The next thing you need to do then, is click on the "What it does" link, on the left of your screen, and study the main steps. When you are done we'll tell you what to do next.

Right Click, Select all, Copy, and Paste the text below into your HTML code if you would like to embed the free Reuters newsTicker on your site.

(Never made a website before? Don't worry, just paste the code into Notepad, add whatever text you like, and save as a file called favourite.html on your desktop. You have made your first website. You can email it to colleagues or put upload it to a server.)

GoldenGem®

General Information about Neural Networks

Background Information (from Wikipedia):

Theory of Neural Networks (J. Moody)

Summary of operation:

User instructions

Availability:

Try it online!

Full installation:

Registration as freeware:

Recreational/educational from GoldenGem.co.uk

Test your copy of GoldenGem

Extra comment for the expert.

The auto-train feature

Advantages:

Contact Us

Beginner's Introduction

Questions and Answers

Conclusion

GoldenGem^®