Noxa indicators for Neuroshell

[Rengoku]

I'm thinking that this has something to do with 64 bit vs. 32 bit! im getting ready to install 32 bit on my laptop... I'll let you know!

 

[Rengoku]

Sure enough! If anyone is having issues running this addon... it just might be you are on a 64 bit system! As soon as I installed Windows 7 32 bit... I had zero issues! Now the fun begins!

 

[nevis]

Re: My most successful neural nets

Actually my systems are adaptive to some extent even though CSSA is not. It’s all built in the distributions at the back end. To make a story short, I look at these 3 cases: CSSA is right and generates equity, CSSA is right but does not generate equity, CSSA is wrong. I trade only the first case.

In fact I have little faith in adaptive indicators because they need a lot of bars to reliably reflect a change. Also, models in NeuroShell are globals so they don’t adapt. To me adaptive inputs are quite irrelevant. One other possibility that I trust more is to make the inputs more stationary by taking the first derivatives. This way NeuroShell learns first order changes. But the problem of market changing the changes remains. I found that adapting to market changes at the back end is more robust.

Nevis

Hi Nevis,

Thanks for your explanation, your statement still making me to confirm that CSSA should be set up manually. Some time it will not follow the market change and the parameter to be readjusted. So far there is no auto adjustment/adaptive features, and it is difficult to use NN or GA to find the proper CSSA parameters.

Please advice if you can state it is a robust system if its not an adaptive to a market change.

Thanks,
Arryex

 

[amit_rwd]

Arry master!

Any idea how one could train the neural net for double tops? I have attached the pic.

i normalised data like High previous day - high today , high previous day - low previous day .. scraching my head .. i used input selection to predict 4 bar percentage in open after a top has been established

 

[amit_rwd]

I noticed on while using neural nets results can vary drastically. I wanted some suggestions to avoid this. I have a scenario in which i trained eur/jpy from 01/01/2008-01/01/2009 and used the training set to forecast 02/01/2009 - present The results were approx 10% returns annualised. I saved this template .

Now when I opened the same template on other pc (using the same datafeed) the results changed and were now 5% annualised returns . I know now neural nets are data sensitive but is there any way i can save the data it has learned and use the same learning process on different pc ? All this has happened in neuroshell trader 5

 

[Highfreq]

Arry master!

Any idea how one could train the neural net for double tops? I have attached the pic.

i normalised data like High previous day - high today , high previous day - low previous day .. scraching my head .. i used input selection to predict 4 bar percentage in open after a top has been established

You can turn a neural net into a pattern detector by training it only on bars that represent best this pattern. For that you can use the null entry feature in NST. For example in pseudo code you could do:

If (top is detected at open 4 bars ago) and (current % in open = x%) then open else *

That substitute N/A (the null entry) for training bars that do not conform to your pattern. This only trains the prediction on inputs that resemble the pattern you want to predict. You can create several of these nets to predict different values for x%.

One way to detect for the tops is to use CSSA QPhase:
If QPhase(at Open 4 bars ago) crosses above 0

 

[amit_rwd]

You can turn a neural net into a pattern detector by training it only on bars that represent best this pattern. For that you can use the null entry feature in NST. For example in pseudo code you could do:

If (top is detected at open 4 bars ago) and (current % in open = x%) then open else *

That substitute N/A (the null entry) for training bars that do not conform to your pattern. This only trains the prediction on inputs that resemble the pattern you want to predict. You can create several of these nets to predict different values for x%.

One way to detect for the tops is to use CSSA QPhase:
If QPhase(at Open 4 bars ago) crosses above 0

high freq thx for reply..
are you suggesting that I should insert this as a condition in neuroshell prediction wizard?

 

[Highfreq]

high freq thx for reply..
are you suggesting that I should insert this as a condition in neuroshell prediction wizard?

I like to create my inputs separately first (Insert>New Indicator…). By doing so I can have a look at them; it is important to have enough learning instances to avoid overfitting. Then I add them as fixed inputs in the prediction wizard.:smart:

 

[windsurfing_stew]

One way to detect for the tops is to use CSSA QPhase:
If QPhase(at Open 4 bars ago) crosses above 0

I find that for shorter wavelenths (up to say 10 bars) this works well. However for longer cycles (10 to 40 bars) it's sweeter to get in a bit late and exit early. That is, only trade when the cycle is at its steepest.

I tend to also use the conditional null * to selectively train nets. Seems to work well. Glad to see someone else has picked up on this.

Highfreq - If you'd like to post your chart it would be interesting to see examine the example you were discussing.

Interestingly CSSA Cycle and MESA8's StaticBandPass are highly similar. I tend to use the latter as you have more control over the frequency. Anyone else found this?

 

[windsurfing_stew]

Arry master!

Any idea how one could train the neural net for double tops? I have attached the pic.

i normalised data like High previous day - high today , high previous day - low previous day .. scraching my head .. i used input selection to predict 4 bar percentage in open after a top has been established

Amit - One thing that is very curious about the chart you have posted is that it once the price starts moving in a given direction (up or down) it tends to continue to move that way. To me that is an even more favourable strategy for that chart than trying to pick double tops etc.

 

[amit_rwd]

Amit - One thing that is very curious about the chart you have posted is that it once the price starts moving in a given direction (up or down) it tends to continue to move that way. To me that is an even more favourable strategy for that chart than trying to pick double tops etc.

WS is difficult to predict the direction as you mentioned because one would not know which way its moving. However after the first top is formed the net could be trained on the characteristics of first top ( first arrow structure) and the net could buy/sell appropriately. Please notice the very close similarity on the two top. It seems that some hedge funds are running such adaptive nets. Note the aggresive buying which is long bullish green candles on left leg, whereas the slow selling and many small red candles on right leg of each top.

The null feature you guys talking about can you post a NS chart with example please

 

[amit_rwd]

Which is a better neural net model from your perspective
model1>
training set 1 year and out of sample set 1 year
training(backtest) return 35% , no of trades 700
out of sample (backtest)return 40%, no of trades 1000
sharpe of 5 on out of sample

model 2>
unoptimised (backtest) net prediction for 1 year , 1000 trades and return of 55% with a sharpe of 6.

I am giving preferance to systems which have good performance on the basis of return/no of trades. I think the no trades greater than 1000 for a particular model is a good indication of reliable model

 

[Highfreq]

Which is a better neural net model from your perspective
model1>
training set 1 year and out of sample set 1 year
training(backtest) return 35% , no of trades 700
out of sample (backtest)return 40%, no of trades 1000
sharpe of 5 on out of sample

model 2>
unoptimised (backtest) net prediction for 1 year , 1000 trades and return of 55% with a sharpe of 6.

I am giving preferance to systems which have good performance on the basis of return/no of trades. I think the no trades greater than 1000 for a particular model is a good indication of reliable model

Why do you think the number of trades matters? To me, the equity line is what matters the most.
I found this interview interesting:
http://championship.mql5.com/2010/en/news/16

 

[windsurfing_stew]

Amit,

There isn't necessarily a right or wrong answer as to how many trades is optimal other than -

* If you haven't already factored in transaction costs then you might give preference to fewer trades
* If you are not fully automated then you may once again give preference to fewer trades
* Statistically you will have more confidence with more trades
* As a rough rule of thumb, you should have a minimum of 30-100 trades in order to have any statistical confidence in your results.

Is your test on one stock or is this a summary across a number of symbols?

Take a look at the list of profit per trade. Check that your averages are not being skewed by one or two large outliers too.

Both systems sound ok. A Sharpe of 5-6 is excellent.

 

[amit_rwd]

* If you haven't already factored in transaction costs then you might give preference to fewer trades
Yes I have taken into consideration, All my systems are good for ECN spreads and even better for usage at HF funds.
* If you are not fully automated then you may once again give preference to fewer trades
* Statistically you will have more confidence with more trades
True this is what I am looking for , 5000 bars or 1000 trades per year
* As a rough rule of thumb, you should have a minimum of 30-100 trades in order to have any statistical confidence in your results.
no sure about 30-100 trades

Is your test on one stock or is this a summary across a number of symbols?
I have tested on just 1 fx pair based trading strategy having 3 nets which has statistically generated monthly 150 pips minimum . This TS is still on test phase.

Take a look at the list of profit per trade. Check that your averages are not being skewed by one or two large outliers too.
no large outliers , TS has performed in both up and down moves equally

Both systems sound ok. A Sharpe of 5-6 is excellent.[/QUOTE]

Thx for your response as usual

 

[windsurfing_stew]

Perhaps add more FX pairs to your strategy (say another 9). Keep the best few pairs from optimisation for your out-of-sample testing. Usually if a model works on one pair then it won't work on all pairs but will work on at least a couple of others.

 

[amit_rwd]

Thiis pattern has repeated again today on 15min eurusd

 

[newt24]

stupid question this is, but all the people that show images with the buy/sell arrows, are they using real live data? or is it all back testing / optimising, which don't really help in the real world.

So people really trade of the charts?

thanks

 

[supremegizmo]

has anyone here has problems in transferring neuroshell template from one pc to another?
I have saved a model in my 1st pc which has approx returns fo 30% when i reopened it on another pc with same historical feed the model returns changed to 22% . Is there any way to make sure the neuroshell model saves its learning and can use the same learning in another pc?

 

[Highfreq]

has anyone here has problems in transferring neuroshell template from one pc to another?
I have saved a model in my 1st pc which has approx returns fo 30% when i reopened it on another pc with same historical feed the model returns changed to 22% . Is there any way to make sure the neuroshell model saves its learning and can use the same learning in another pc?

I don't have that problem. I run NeuroShell on several systems (Intel, AMD, 32 and 64bits); my charts are identical on all of them. Have you tried saving your charts with the data attached? One reason for a chart to change is that the input series have changed...