[Ward Systems Group Support]

One of the examples we used to show at our training seminars was a neural net with perhaps 15 to 20 columns of random numbers predicting another column of random numbers.  The example worked very well.  The purpose of the example was to teach methods to prevent overfitting.

1.  Keep the number of inputs low (from 5 to 10 inputs)
2. Make sure training sets are large enough to cover up, down, and sideways markets, but not so large as to include obsolete patterns
3. Don’t use too many hidden neurons
4. Make sure your inputs are detrended
5. Optimizing over all chart pages can help.

The best method to judge the success of a neural net is to look at performance in out-of-sample data.  Are the trading signals hitting peaks and valleys.  Is the profit based on one or two “lucky” trades or does it cover multiple trades.

 

[Ward Systems Group Support]

AT2 only recalculates itself and not the other indicators.

[Ward Systems Group Support]

In the Trading Strategy Advanced tab, there is a parameter you can set called “Shortest desired trade span” that you could set to 5 days.

There is a similar option on the Optimization tab of the prediction wizard.

[Ward Systems Group Support]

If you are talking about custom indicators from the Trader Add-ons, they can’t be modified and re-saved as custom indicators.

If you want to combine indicators that are included in NeuroShell Trader and save them with a different name and and parameter settings, you can do that.

 

 

[Ward Systems Group Support]

> I am trying to determine how the TP2 when it is set to 1 bar forward
> and retrain on each bar is actually working.

As far as the Tprop indicators are concerned, they are simply an indicator that returns a stream of values.  TProp indicators are absolutely no different than a moving average calculation, except that it does a more complicated computation.

So for instance a Moving Average stream of data is computed as follows:
1) It takes a window size (say 200 bars) and produces a value for the last bar (i.e. the average of the last 200 bars).
2) It then moves it’s window forward one bar and produces a value for the new bar at the end (i.e. the average of the last 200 bars).
3) Each time it shifts forward in time, it uses the prior 200 bars to create the last bar’s value using a completely independent calculation from the prior windows.

In TProp’s case, it does exactly the same thing as the Moving Average.  The Tprop indicator set to retrain on every new bar is computed as follows:
1) The Tprop indicator takes a window size (say 200 training set size) and produces a value for the last bar (i.e. it trains a neural network on the specified input values over the earlier 200 bars not including the most recent bar and then applies that neural network to the input values of the most recent bar … i.e. input values it wasn’t trained on … aka “out of sample” value).
2) It then moves it’s window forward one bar and produces a value for the new bar at the end (i.e. an entirely new neural network trained over the prior 200 bars and then applied to the last bar’s input values.
3) Each time it shifts forward in time it uses the prior 200 bars to create the last bar’s value using a completely independent calculation from the prior windows.

Notice how the only difference between 1 thru 3 for the Simple Moving Average and the Tprop is that the complexity of the calculation.  The Moving Average does a summation and divides by the length of the window size.  Tprop on the other hand builds a model on the prior data and then uses that model to “predict” the last bar’s value.   They both then “walkforward” the calculation one bar and do completely new calculations to compute the next value.

> So if it finds a better solution on 1 bar, does it then applies that solution to the other bars ?

No, as explained above, every new walkforward (i.e. shifting of the calculation window) uses a completely independent set of calculations … whether you are talking about a Moving Average that walks-forward it’s calculation window by 1 bar or a Tprop indicator that walks-forward it’s training set by 1 bar.

> Then when it is in testing mode , i see the number of winning trades move around
> by a large number , if it is walking forward why does not it change by One
> trade at a time ?

By testing mode, do you mean optimization?  If so, let’s make up a Moving Average trading system that is roughly equivalent to your TProp trading system for explanation purposes (and because we’ve established that the simple moving average and Tprop are almost the same except for the calculations “behind the curtain”)

Moving Average Trading System:
Long Entry when Simple Moving Average ( Stochastic %D ( 10 ), 200 ) is greater than 50
Long Exit when Simple Moving Average ( Stochastic %D ( 10 ), 200 ) is less than 50

When you optimize the Simple Moving Average system above, the genetic optimizer is going to randomly try out different combinations of Moving Average window sizes and Stochastic Window Sizes.  So for instance it might try {10,200), (4, 111), (16, 57), (12, 237), etc.  Each time it tries a new set of potentially “optimal” parameters, it is going to create an entirely different stream of data which varies around 50 entirely differently than the prior set of values and thus creates a different number of trades.  Because it is trying the values randomly at first, you would expect the number of trades to jump around all over the place with each new set of parameters tried.  Even as it narrows in on a solution, it still tries some random mutations, so the number of trades may still jump around.

A Tprop optimization is no different than the Simple Moving Average optimization above.  Depending upon what you are optimizing, it tries different input parameters (i.e. stochastic %D parameters, etc.), different window sizes (aka training set size), different number of bars before retrain, etc. etc. etc.  With each random set of potentially “optimal” values, you are getting a completely different stream of data values which you would expect to give wildly varying number of trades.  Once again because it is trying the values randomly at first, you would expect the number of trades to jump around all over the place with each new set of parameters tried.  Even as it narrows in on a solution, it still tries some random mutations, so the number of trades may still jump around.

[Ward Systems Group Support]

One other note, you can let the genetic algorithm optimize the lookback period.

[Ward Systems Group Support]

The following from the help file may be of interest to this discussion:

Most neural nets, including Turboprop2 in the Prediction Wizard, need far more training bars than the 20 to 200 we have recommended. We believe that smaller training set sizes are appropriate with AT2 because the training window is so dynamic, and an out-of-sample prediction only lasts for one bar. You may disagree with this (and in fact some neural network experts may disagree, too), in which case you are free to use up to 10,000 bars.

 

[Ward Systems Group Support]

We will pass your thoughts to the developers.

[Ward Systems Group Support]

Attached is an example chart that fills the Long Entry on a limit order.

Attachments:

You must be logged in to access attached files.

[Ward Systems Group Support]

If you use an AT2 net inside of the prediction wizard, the AT2 value will be updated every bar if that is how you set it up.  However, the overall prediction will not be retrained unless you hit the retrain button.  This would be similar to new values for indicator inputs that show up every bar.  The net from the prediction is applied to the new values, but the model itself doesn’t change.

[Ward Systems Group Support]

You can’t perform walk forward optimization on a prediction.  However, the Adaptive TurboProp 2 add-on allows you to retrain the model in as many bars as you set in the indicator.  For example, you can retrain the model every daily bar or every week.

[Ward Systems Group Support]

Can you send your chart with saved data to support@wardsystems.com so we can look at all of the settings?

[Ward Systems Group Support]

We suggest you give them a call.  They should be happy to speak with a potential new customer.

Contact Information:

• Justin Choi or Joseph Bangkot
• Phone: 1-770-999-4511, option #3

 

[Ward Systems Group Support]

The Trader does not include a specific method for adding a value for Forex market rollovers.  You can simulate it by either increasing commissions or margin requirements based on an average commission.

[Ward Systems Group Support]

Glad the chart gave you a new area to explore.  I’m also working on volume indicators, as you can tell from the recent newsletters.

Marge