Noxa indicators for Neuroshell

[alby1714]

Hi Krzysiaczek99,
I found your tests very interesting. Clearly noise affects the performance as Noxa themselves seem to admit. Nevetheless CSSA remain profitable as the noise change and this is not a mean achievement unless it were due to pure chance !
Your consideration that PF should "not fall at least in case when we move to lower noise conditions. " is thought provoking and I would invite Noxa to comment on that specific point.
The bottomline is perhaps that one cannot expect perfect performance and therefore it would be of value to all of us to be able to assess the limits of CSSA in comparison with other available indicators. Have you any views about this ?
Thanks.

alby1714

First I agree MACD PF rise is coincidental and not important here, it's simple EMA strategy and it does not distinguish between noisy and not nosy markets.

Regarding CSSA. In my opinion it simple learns cycle and noise, this is the reason that i was calling this 'curve fitting' to noise. So when noise level changes it starts to under
perform because is not adapted to level of a new noise. So it doesnt extract cycle correctly, if it would extract correctly (clean signal which was the same in GOLD5 and -9db) PF would not fall at least in case when we move to lower noise conditions. The settings which were extracting cycle correctly in one case should extract cycle in another because it is the same cycle.

Than it is very risky business to use it if it is not able to filter out noise well. Does it has filtering mechanism build in ?? Which filters if any ?? On the other hand GroupStart and GroupDepth settings should filter noise i guess. So why PF falls ??

Krzysztof

 

[Krzysiaczek99]

remain profitable

Hi Krzysiaczek99,
I found your tests very interesting. Clearly noise affects the performance as Noxa themselves seem to admit. Nevetheless CSSA remain profitable as the noise change and this is not a mean achievement unless it were due to pure chance !
Your consideration that PF should "not fall at least in case when we move to lower noise conditions. " is thought provoking and I would invite Noxa to comment on that specific point.
The bottomline is perhaps that one cannot expect perfect performance and therefore it would be of value to all of us to be able to assess the limits of CSSA in comparison with other available indicators. Have you any views about this ?
Thanks.

alby1714

Hi alby,

I love your mail, it shows how easy is to buy an advertisement. The tests are done on artefficial signals which CSSA must recognize !!! Not non linear and non stationary FOREX/STOCK MARKET signals with who knows which type of noise.
If i would make the tests on those signals i would never find what i found. See post 89 for more details.

So absolute profitability doesn't mean anything here, only the direction of changes means.

So last mail from NOXA except that he confirmed noise impact i would classify as 'bussiness talk'. The fact that NOXA don't want to give settings for real out of sample test means something.....

There is no similar indicator on the market except DT_SSA from Russia which I didn't try.

Krzysztof

 

[Krzysiaczek99]

exact noise impact on CSSA

Hello everybody,

As last response from NOXA was very enigmatic in my opinion i decided to make an additional test to find out exact impact of changing noise conditions for CSSA E0 performance. This is simulation of change in market conditions
which happens all the time in the market. For those examples noise is only changing, underlying cycle is all time the same and is plottet together with noised signal to compare. I also plot CSSA cycles.

In order to do it i created 3 triple charts with 3 csv file each

-9db
GOLD5
GOLD15

for readers info

-9db file S/N === -18db S/N based on 20log(s/N)
GOLD5 file S=2.5 N=6.5 S/N= -8db
GOLD15 clean signal = underlying cycle

Fourier spectrum of first two in post 138, clearly seen that -9db is more noisy, in GOLD5 two peaks of underlying
cycle 0.1 and 0.5 HZ

than i collect 'perfect settings' settings which are giving the best performance (money wise) for each signal
they are:

-9db NOXA
L 109,3,2,1,0046,1,435
S 102,3,1,0,0.054,1,435

GOLD15 perfect
L 12,2,5,0,0,1,435
S 50,2,1,0,0,1,435

GOLD5 perfect
L 112,2,1,1,0.021,1,435
S 58,3,2,1,0.943,1,435

and i rerun those settings against files with different S/N level. Here are the results
Charts are attached.

First chart NOISE_test_5_9db_15_settings with perfect settings for clean underlying cycle (GOLD15) and printing trades for underlying cycle

Number of trades 55, hit=100%, 1Ytrades=76352% so perfection !!!

 

[Krzysiaczek99]

against GOLD5 -8db

than I switch a chart and run against gold 5 file so noise increased to -8db. Than it looks very different. Please compare underlying cycle and trades. Buying on tops and selling on bottoms !!!! Wow, it looks that underlying cycle is not in control of trade decision at all !!

Look what happen to CSSA cycle does not look like a sin at all for long entries because in perfect setting m-history is set to 12 which was giving the best performance. Anyway I thought that because we set GroupStart and GRoupdepth to proper eigenvectors we filter the noise

Statistic

Number of trades 75, 1Yrt=12145% PF=1.75, hit=57%

COMMENT

Number of trades grown so next confirmation than underlying cycle is not in control

 

[Krzysiaczek99]

against -9db

than i switched to -9bd. Similar situation like in GOLD5 case,

Statistic:

Number of trades 103 ---- grown again, 1yrt 76353% so higher than in GOLD5 even noise is much stronger !!

PF=1.52, hit=59.2%

I made an analysis only for one chart i.e. with perfect settings for GOLD15 - underlying cycle, i attach three charts with other 'perfect setting' so you can make analysis yourself

CONCLUSION is very simple. Even if we are adjusted by GrupuStart and Depth to underlying cycle due to noise change those settings don't work at all in other noise conditions simply because there is no noise filtering so noise takes control for making trading decisions, perfect setting are adjusted to 'old noise' level.

Anyway don't expect much more posts from me as after two weeks I feel i know enough
to move to another area - professor Codebreaker and GOERTZEL. If anybody already
made an indicator for Goertzel for NS (dll) or want to do it, let me know.

Krzysztof

 

[Noxa]

Noise issue figured out?

As i promised I made a key test for me. In order to check my suspicion that curve fitting occurs in CSSA i made following test

First I fetched GOLD5 file (noise -15db) with the settings given by NOXA, the results were superb, PF for CSSA 605/80% for MACD 0.46/30%

Than I fetched optimized settings from GOLD5 for CSSA and MACD and used them as an input setting for file with noise -9db. as a signal behind it the same just masked by noise with different level i expected CSSA to perform even better because of higher S/N ratio !!! This situation is equal to changing market conditions (noise level) to actually better conditions - lower noise.

Than look what happened. The performance of MACD has grown but performance of CSSA has fallen immediatelly. MACD outperform CSSA by 50% during the same period.

So my conclusion is that even if settings are for cycle extraction CSSA makes curve fitting as well and this is not acceptable.

FIles are attached and charts

-9db -combined signal (GOLD15+ 9db_noise)
-9db_noise - just noise from NS

Other files you have

GOLD1 - gauss noise
GOLD15 - clean signal
GOLD5 = GOLD1+GOLD15

Krzysztof

Further to this post and posts #144, #149:
http://www.trade2win.com/boards/666974-post144.html
http://www.trade2win.com/boards/667254-post149.html

You said that when the noise is reduced, the Ratio Gross Profit/Loss from CSSA drops therefore there is curve fitting. In the following experiment I show that a drop in the ratio can simply be an artifact from noise. I use neither CSSA nor learning; I simply use the perfect entries from GOLD15. I then add some noise to it. The entries remain fixed whatever the noise; so we have a perfect system that is not impacted by noise. Bottom curve (grey line) shows GOLD15 (sin + cos) wave; troughs trigger long entries (blue signals), peaks trigger short entries (red signals). The top graph shows GOLD15 + noise.

I use two different noises: KS#3 is a uniform noise; KS#110 is a Gaussian noise. The peak to peak amplitude of the Gaussian noise KS#110 (s/n = -13dB) is twice as large as the peak to peak amplitude of the uniform noise KS#3 (s/n = -7dB).

See what happens:
Gaussian noise -13dB: Ratio Gross Profit/Loss = 12.37
Uniform Deviates -7dB: Ratio Gross Profit/Loss = 3.92

When noise is reduced by a ratio of two, the Ratio Gross Profit/Loss dropped from 12.37 to 3.92. In this case this is not curve fitting but an artifact from noise of different characteristics. Have you ever considered this in your results?

Noxa

 

[Noxa]

than I switch a chart and run against gold 5 file so noise increased to -8db. Than it looks very different. Please compare underlying cycle and trades. Buying on tops and selling on bottoms !!!! Wow, it looks that underlying cycle is not in control of trade decision at all !!

Look what happen to CSSA cycle does not look like a sin at all for long entries because in perfect setting m-history is set to 12 which was giving the best performance. Anyway I thought that because we set GroupStart and GRoupdepth to proper eigenvectors we filter the noise

Statistic

Number of trades 75, 1Yrt=12145% PF=1.75, hit=57%

COMMENT

Number of trades grown so next confirmation than underlying cycle is not in control

I am really impressed with all the efforts you spent trying to make CSSA look bad at any price but it can’t be that bad!!!!

>> Wow, it looks that underlying cycle is not in control of trade decision at all!!
The reason for such bad result is an excessively poor setting!
L 12,2,5,0,0,1,435
m-histories = 12
GroupStart = 2
GroupDepth = 5
I wonder who is out of control! Look at the group you chose (see screen shot of ShowEigenvector). It is just noise!!! Your signals are clearly out of phase indicating that they are not optimal at all!!! Was that intentional?

As you can see in my own results (see graphs, chart attached), the long and short cycles came out almost identical which is expected (see bottom graph) and the entries remain unchanged after adding the noise.

>> Anyway I thought that because we set GroupStart and GRoupdepth to proper eigenvectors we filter the noise
That is correct given that you set the parameters properly.

Comments?

Noxa

 

[Krzysiaczek99]

clarification

I am really impressed with all the efforts you spent trying to make CSSA look bad at any price but it can’t be that bad!!!!

>> Wow, it looks that underlying cycle is not in control of trade decision at all!!
The reason for such bad result is an excessively poor setting!
L 12,2,5,0,0,1,435
m-histories = 12
GroupStart = 2
GroupDepth = 5
I wonder who is out of control! Look at the group you chose (see screen shot of ShowEigenvector). It is just noise!!! Your signals are clearly out of phase indicating that they are not optimal at all!!! Was that intentional?

As you can see in my own results (see graphs, chart attached), the long and short cycles came out almost identical which is expected (see bottom graph) and the entries remain unchanged after adding the noise.

>> Anyway I thought that because we set GroupStart and GRoupdepth to proper eigenvectors we filter the noise
That is correct given that you set the parameters properly.

Comments?

Noxa

Hi,

Than please rerun exact settings which you gave now against -9db. I also found out that settings from 'noisy' envirnoment can work for clean signal so what you did
you found perfect settings for GOLD5 and rerun in GOLD15 --> noise decrease to 0. If you will try to rerun it in another noisy file it will always collapse.

My example was opposite I found settings for GOLD15 and rerun in GOLD5 and -9db
so noise increase. Other charts are doing different combinations. See also next post

key point is that one setting can be completelly wrong after noise change and we are never sure how it can perform in the future, change of trades numbers and my post confirms that underlying cycle is loosing of control in trading decision.


>>The reason for such bad result is an excessively poor setting!
L 12,2,5,0,0,1,435
m-histories = 12
GroupStart = 2
GroupDepth = 5
I wonder who is out of control! Look at the group you chose (see screen shot of ShowEigenvector). It is just noise!!! Your signals are clearly out of phase indicating that they are not optimal at all!!! Was that intentional?

it was the settings which gave the best performance money wise for GOLD15. GA has chosen then (100% hit). So 100% hit setting collapsed imediatelly for another noise level.

I just checked cross chart NOISE_test_15_9db_5_settings
so perfect setting given by you for GOLD5. See chart after moving to -9db

PF down from 9.78 to 0.79, number of trades from 49 to 443 !! a lot of buying tops...so even with your perfect settings it collapses

GOLD5 perfect
112,2,1,1,0.021,1,435
58,3,2,1,0.943,1,435

For perfect -9db setting when we move to GOLD5 (lower noise) PF goes down from
11.6 to 6.05, number of trades the same so maybe underlying cycle is in control
As i see from mail above PF can fall because by noise type change. Easy to verify, maybe I will try it.

Anyway as I told you i don't have more time for this, simple after two weeks i realized the complexity and unstability of settings with unclear optimization procedure (post 127 and 128 not answered) so I prefer spend my time on more 'controllable' and clear things.

The fact that nobody interfere those discussions i think says something, anyway if somebody would show up
claiming big succesess with CSSA i would ask him immediatelly for exact settings for out of sample test with real market.

Krzysztof

 

[Krzysiaczek99]

poor settings ??

>>Wow, it looks that underlying cycle is not in control of trade decision at all!![/COLOR]
The reason for such bad result is an excessively poor setting!
L 12,2,5,0,0,1,435
m-histories = 12
GroupStart = 2
GroupDepth = 5
I wonder who is out of control! Look at the group you chose (see screen shot of ShowEigenvector). It is just noise!!! Your signals are clearly out of phase indicating that they are not optimal at all!!! Was that intentional?

See screenshot below for GOLD15. 5 eigenvectors

so

GroupStart = 2
GroupDepth = 5

or maybe

GroupDepth = 4

or maybe

GroupDepth = 3

anyway i changed by hand to 4 and 3 and again 100% in GOLD15 and collapse in GOLD5 so it didn't change anything here.

I hope is clear for everybody that the noise (market conditions) change happens during trading so we trade with old settings but we suppouse to be adjusted to underlying cycle so this change should not hurt us. That is normal market conditions as noise level changes during day and resiliance to it is a key.

Krzysztof

 

[Noxa]

Hi,

Than please rerun exact settings which you gave now against -9db. I also found out that settings from 'noisy' envirnoment can work for clean signal so what you did
you found perfect settings for GOLD5 and rerun in GOLD15 --> noise decrease to 0. If you will try to rerun it in another noisy file it will always collapse.

My example was opposite I found settings for GOLD15 and rerun in GOLD5 and -9db
so noise increase. Other charts are doing different combinations. See also next post

key point is that one setting can be completelly wrong after noise change and we are never sure how it can perform in the future, change of trades numbers and my post confirms that underlying cycle is loosing of control in trading decision.


>>The reason for such bad result is an excessively poor setting!
L 12,2,5,0,0,1,435
m-histories = 12
GroupStart = 2
GroupDepth = 5
I wonder who is out of control! Look at the group you chose (see screen shot of ShowEigenvector). It is just noise!!! Your signals are clearly out of phase indicating that they are not optimal at all!!! Was that intentional?

it was the settings which gave the best performance money wise for GOLD15. GA has chosen then (100% hit). So 100% hit setting collapsed imediatelly for another noise level.

I just checked cross chart NOISE_test_15_9db_5_settings
so perfect setting given by you for GOLD5. See chart after moving to -9db

PF down from 9.78 to 0.79, number of trades from 49 to 443 !! a lot of buying tops...so even with your perfect settings it collapses

GOLD5 perfect
112,2,1,1,0.021,1,435
58,3,2,1,0.943,1,435

For perfect -9db setting when we move to GOLD5 (lower noise) PF goes down from
11.6 to 6.05, number of trades the same so maybe underlying cycle is in control
As i see from mail above PF can fall because by noise type change. Easy to verify, maybe I will try it.

Anyway as I told you i don't have more time for this, simple after two weeks i realized the complexity and unstability of settings with unclear optimization procedure (post 127 and 128 not answered) so I prefer spend my time on more 'controllable' and clear things.

The fact that nobody interfere those discussions i think says something, anyway if somebody would show up
claiming big succesess with CSSA i would ask him immediatelly for exact settings for out of sample test with real market.

Krzysztof

Krzysztof

I find your posts extremely perceptive and interesting but also misleading and provocative to say the least. So, please, do not distort or misinterpret my words. I have neither the time nor the willingness to go down that route with you.

The experiment I did was an exact repeat of what you did. Learning was done on the clear signal GOLD15 and test was done on noisy signal GOLD5.

>> If you will try to rerun it in another noisy file it will always collapse.
I did rerun on different noises as you suggested. See screen shots; chart attached.

KS#102 PF = 8.77 %profitable trades = 75%
KS#104 PF = 2 %profitable trades = 50%
KS#106 PF = 0.51 %profitable trades = 41.7%
KS#108 PF = 20.42 %profitable trades = 75%
KS3#109 PF = 111.50 %profitable trades = 91.7%

In all cases the long and short cycles, remained almost identical (see bottom graphs) in all noise configurations (s/n=-8dB) meaning that any disparity on the results you can see is only due to the artifact from noise that I described in my post#166 and has nothing to do with CSSA.

By the way could you comment on my post#166? http://www.trade2win.com/boards/670462-post166.html


>> key point is that one setting can be completelly wrong after noise change and we are never sure how it can perform in the future, change of trades numbers and my post confirms that underlying cycle is loosing of control in trading decision.
As you saw in my experiment, the settings were almost perfect after applying noise but disparity on the PF was large (from 0.51 to 111.5). Some charts made money (KS#102, KS#108, KS#109) some other charts lost money (KS#104, KS#106). The reason for that is the artifact from noise which has very little to do with CSSA since the cycles remained unchanged under noise conditions.
Results collapsing immediately as you said can be a byproduct of noise which has nothing to do with CSSA. Don’t you agree?

I’ll touch base in a later post about another noise artifact that might also play a role.


>> Anyway as I told you i don't have more time for this, simple after two weeks i realized the complexity and unstability of settings with unclear optimization procedure (post 127 and 128 not answered) so I prefer spend my time on more 'controllable' and clear things.
I believe that all the trouble comes from your inexperience in running an auto-optimization in NeuroShell Trader. You told me you have 4 months of experience with it. I thought it was a typo as you really got 4 weeks and you know what I am talking about.

Also don’t tell me there is a simple solution to markets. This is misleading information. CSSA is very flexible therefore very powerful to handle all markets conditions. Don’t expect it to be as easy as clicking on an optimize button.

Noxa

 

[Krzysiaczek99]

new GOLD15 perfect settings ??? why ??

Hi,

my setting were

GOLD15 perfect
12,2,5,0,0,1,435
50,2,1,0,0,1,435

which were correct for GOLD15, giving 100%, perhaps you should clarify to the readers that you made mistake in post 167 because you didnt (I hope) understand the test and I clarified this in next post.

Than now you changed them to 250,4,2,0...

This is backward 'curve fitting' to 'cook' the results and you know it so point stays as it was

one setting can be completelly wrong after noise change and we are never sure how it can perform in the future, change of trades numbers and my post confirms that underlying cycle is loosing of control in trading decision.

My setting was found in completelly correct way

The same problem occurs for GOLD5 perfect settings found by you when we move
to other noise level like i shown in previous mail

Regarding my expirience with NS and in general. I think this is the biggest unluck which
NOXA could face here......I know and you know very well that 99% of people involved
in FOREX/STOCK MARKET would not be able to do those tests and made such discussion but i was able somehow....

Yes, optimization is big problem here specjally if you are using NS only. I made investigation myself meanwhile and perhaps post something but good you mentioned it.

Regarding post you mention to comment. I even didn't look because if underlying cycle can lose control thats end for me......but i will have a look if you wish.

Krzysztof

P.S. I have NS 5.5 now so can open your charts. Somebody offered me your NEI indicators to public test....

 

[Noxa]

Very bad settings indeed

>>Wow, it looks that underlying cycle is not in control of trade decision at all!![/COLOR]
The reason for such bad result is an excessively poor setting!
L 12,2,5,0,0,1,435
m-histories = 12
GroupStart = 2
GroupDepth = 5
I wonder who is out of control! Look at the group you chose (see screen shot of ShowEigenvector). It is just noise!!! Your signals are clearly out of phase indicating that they are not optimal at all!!! Was that intentional?

See screenshot below for GOLD15. 5 eigenvectors

so

GroupStart = 2
GroupDepth = 5

or maybe

GroupDepth = 4

or maybe

GroupDepth = 3

anyway i changed by hand to 4 and 3 and again 100% in GOLD15 and collapse in GOLD5 so it didn't change anything here.

I hope is clear for everybody that the noise (market conditions) change happens during trading so we trade with old settings but we suppouse to be adjusted to underlying cycle so this change should not hurt us. That is normal market conditions as noise level changes during day and resiliance to it is a key.

Krzysztof

Yes your setting is very poor. Very hard to believe that the GA produced it!
m-histories = 12
GroupStart = 2
GroupDepth = 5

The problem is m-histories = 12 meaning that you have decomposed price in 12 components. My optimal setting is about 200. Moreover you have shown ShowEigenvector as if m-histories was 50; it is not 50, it is 12. May be a mistake may be not, but very misleading! See the screen shot of the ShowEigenvector window for GOLD15 with m-histories = 12 the way it should be. Your setting chose a group made of components 2 to 5. As you can see component 5 is noise so the group has noise in it.

Noxa

 

[Noxa]

Nothing to do with optimization; it's manual setting to show the slow cycle

Hi,

my setting were

GOLD15 perfect
12,2,5,0,0,1,435
50,2,1,0,0,1,435

which were correct for GOLD15, giving 100%, perhaps you should clarify to the readers that you made mistake in post 167 because you didnt (I hope) understand the test and I clarified this in next post.

Than now you changed them to 250,4,2,0...

This is backward 'curve fitting' to 'cook' the results and you know it so point stays as it was

one setting can be completelly wrong after noise change and we are never sure how it can perform in the future, change of trades numbers and my post confirms that underlying cycle is loosing of control in trading decision.

My setting was found in completelly correct way

The same problem occurs for GOLD5 perfect settings found by you when we move
to other noise level like i shown in previous mail

Regarding my expirience with NS and in general. I think this is the biggest unluck which
NOXA could face here......I know and you know very well that 99% of people involved
in FOREX/STOCK MARKET would not be able to do those tests and made such discussion but i was able somehow....

Yes, optimization is big problem here specjally if you are using NS only. I made investigation myself meanwhile and perhaps post something but good you mentioned it.

Regarding post you mention to comment. I even didn't look because if underlying cycle can lose control thats end for me......but i will have a look if you wish.

Krzysztof

P.S. I have NS 5.5 now so can open your charts. Somebody offered me your NEI indicators to public test....

>> GOLD15 perfect
12,2,5,0,0,1,435
50,2,1,0,0,1,435
which were correct for GOLD15, giving 100%, perhaps you should clarify to the readers that you made mistake in post 167 because you didnt (I hope) understand the test and I clarified this in next post.
Read my previous post for a clarification as to why your setting was poor to start with. What mistake are you talking about? May be you can tell. I won’t repeat it again; the test I did is the exact repeat of your test.

>> Than now you changed them to 250,4,2,0...
This is backward 'curve fitting' to 'cook' the results and you know it so point stays as it was
250, 4, 2, 0 is the manual setting for the slow cycle. It has nothing to do with our concern. Just to show you how easy it is to find the slow cycle.

>> one setting can be completelly wrong after noise change and we are never sure how it can perform in the future, change of trades numbers and my post confirms that underlying cycle is loosing of control in trading decision.

My setting was found in completelly correct way
This is not true; your setting incorporated component 5 which is noise as shown in my previous post.

>> The same problem occurs for GOLD5 perfect settings found by you when we move to other noise level like i shown in previous mail
The perfect settings were from GOLD15 the same way you did. Check by yourself, you have the chart.

Noxa

 

[Krzysiaczek99]

group depth 4 or 5

Yes your setting is very poor. Very hard to believe that the GA produced it!
m-histories = 12
GroupStart = 2
GroupDepth = 5

The problem is m-histories = 12 meaning that you have decomposed price in 12 components. My optimal setting is about 200. Moreover you have shown ShowEigenvector as if m-histories was 50; it is not 50, it is 12. May be a mistake may be not, but very misleading! See the screen shot of the ShowEigenvector window for GOLD15 with m-histories = 12 the way it should be. Your setting chose a group made of components 2 to 5. As you can see component 5 is noise so the group has noise in it.

Noxa

it was a mistake. But i told you I changed by hand gropudepth and i doesn't change anything. Here is changed to 4. Hit 100% for GOLD15 and collapse in GOLD5 and -9db

Krzysztof

 

[Krzysiaczek99]

Gold5

>> The same problem occurs for GOLD5 perfect settings found by you when we move to other noise level like i shown in previous mail
The perfect settings were from GOLD15 the same way you did. Check by yourself, you have the chart.


This is misunderstanding. I talk about perfect setting for GOLD5 from cross chart
NOISE_test_15_9db_5_settings

GOLD5 perfect
112,2,1,1,0.021,1,435
58,3,2,1,0.943,1,435

see screenshots 443 trades in -9db !!!

you didn't do this test. Settings were given by you in previous mails.

Krzysztof

 

[Krzysiaczek99]

GOLD5 settings post 115

here are the screenshots of your settings for GOLD5 from post 115, They were 'perfect' for you at this time and they cause this 443 trades and buying tops in -9db. And than what ??
They should be adjusted again manually to slow cycle or something ??? You dont see slow cycle here at all because of noise !! Than it will be a "backward curve fitting" again !!

Obviously it does not make any logic to readjust setting which give high hit with suspicion
that they will not perform in the future simple because we don't know the future and this you did for GOLD15, change of GroupDepth to 4 didn't change anything, they are still perfect and still falling in the future (post174). If the noise conditions would not change
in the future those setting still would be 'perfect' but if the conditions will change or not we don't know.

Krzysztof

 

[Krzysiaczek99]

artifact of noise

Further to this post and posts #144, #149:
http://www.trade2win.com/boards/666974-post144.html
http://www.trade2win.com/boards/667254-post149.html

You said that when the noise is reduced, the Ratio Gross Profit/Loss from CSSA drops therefore there is curve fitting. In the following experiment I show that a drop in the ratio can simply be an artifact from noise. I use neither CSSA nor learning; I simply use the perfect entries from GOLD15. I then add some noise to it. The entries remain fixed whatever the noise; so we have a perfect system that is not impacted by noise. Bottom curve (grey line) shows GOLD15 (sin + cos) wave; troughs trigger long entries (blue signals), peaks trigger short entries (red signals). The top graph shows GOLD15 + noise.

I use two different noises: KS#3 is a uniform noise; KS#110 is a Gaussian noise. The peak to peak amplitude of the Gaussian noise KS#110 (s/n = -13dB) is twice as large as the peak to peak amplitude of the uniform noise KS#3 (s/n = -7dB).

See what happens:
Gaussian noise -13dB: Ratio Gross Profit/Loss = 12.37
Uniform Deviates -7dB: Ratio Gross Profit/Loss = 3.92

When noise is reduced by a ratio of two, the Ratio Gross Profit/Loss dropped from 12.37 to 3.92. In this case this is not curve fitting but an artifact from noise of different characteristics. Have you ever considered this in your results?

Noxa

Hi,

English is not my mother language but i understand that 'artifact' means product.
Yes I agree, when the noise goes up, PF was going up i also noticed this.

My explanation was that CSSA 'curve fits' or learn noise and PF goes up because of higher amplitude and mean reversion. The best test would be test on noise with Hurst
exponent=0.8 i think because this mean reverting is misleading.

Trading Strategies Based On Digital Filters - Page 41 - Forex Trading

post 403

You can generate those things with MATLAB WAVELET toolbox.

Krzysztof

 

[Krzysiaczek99]

optimization of CSSA

Hello everybody,

In order to conclude my research about CSSA I would like to point one very important thing which is no clear for me. Optimization of CSSA. Here I compiled answers from NOXA about the subject from different posts

- I periodically halted the optimization (at each new epoch) to avoid over-fitting.

Over-fitting occurs when an optimizer learns the noise in the data. One simple way to avoid it is to periodically halt the optimization. The optimization is then continued from the point at which it was halted. Nothing fancy here! Just a way to give you an access to interim results so that you can make an educated guess on when to stop the optimization!


>> So your procedure of optimization for market conditions is simple adjust parameters, optimize, break optimization from time to time to see that model is making money out of sample, if equity curve is rising than model is OK ?? Is that correct ??

It is not fully correct. The interim results can be many things that can help you assess the validity of your model and find market states. There are many things you can do; for example we look for multiple up and down trending segments in the equity lines and some willingness to reproduce that behavior out-of-sample...

post 127

> How you know when epoch is finished in NS ??

This is one of them; simply watch the fitness clusters for the training and out-of-sample data sets. They will move in unison until they diverge.

Can you make a screenschots of it ??

no answer here

post 128 - the same results which proves how easy is to overfit CSSA

post 133

It seems that you are pretty selective in your answers, you didn't comment post 127 and 128 and answer the question if you connect special tool during optimization.

no answer as well


My opinion is that using just Neuroshell is very easy to overfit CSSA optimization, it happened a lot of times for me see post 128. Optimization envirnoment of Neuroshlell is very simple: off, on, pause and paper trading. So no control of other parameters like "look for multiple up and down trending segments in the equity lines and some willingness to reproduce that behavior out-of-sample... or

"simply watch the fitness clusters for the training and out-of-sample data sets. They will move in unison until they diverge."

So my conclusion is that NOXA is using external optimizer to optimize properly CSSA and avoid overfittig. However I couldn't find this tool as an addon for NS but I found as an addon for Tradestation See link and screenshoots. I think NOXA will never confirm using such tool perhaps due to bussiness relations with Ward or not to show to the people that GA from NS is not advanced enough to optimize CSSA properly.

Power Walk Forward Optimizer

If anybody knows about possibility of using external optimizers with NS let me know.

Krzysztof

 

[at120]

hi guys!

A question to Noxa, is it possible to give a sample setting for forex?
Just to understand...

Thanks a lot for all this infos!

Bye, AT

 

[Krzysiaczek99]

sample settings EURUSD30

Hi,

Here you have one. Seems to be perfect for time being. Last 2 days out of sample - 80% hit. Letą see how they will perform in the future. I hope NOXA can give own settings also
and comment for this ones because for me there is no need to change them.

Krzysztof