My current systems - walking forward

[CanOz]

Sinner,

Knowing only a very little about Neural Networks and assuming without a Google that ANN is Adaptive Neural Network i can tell you that i have no experience with these.

I am currently using a genetic program called Adaptrade Builder to construct an algorithm in Easy Language code that i then test on more OOS data, then optimize and walk forward.

Were you thinking i was using some kind of neural network software?

Cheers,


CanOz

 

[tech/a]

I agree with Index futs.

What I'm seeing here is a system WITHOUT an edge.

Low win ratio over a shorter time frame with no opportunity to catch any decent trends.
with such a low win ratio you need much higher winning trades.
Or a lot lower losses.

Its only churning numbers.

The second system the HSI is similar if you take out the largest winning trade.

 

[CanOz]

Yes, i agree on the poor metrics....thats why i like that formula %Wins*PF*RW/L

Its a volitility system, not trend following. Non are trend following systems.

My SPI system seems the best so far in short OOS data, but i have a strange price anomoly in the data that i'm trying to work out...once i get that sorted i'll test it on two years of data and post some results.

In the meantime i think its back to the drawing board until i get something that meets my metrics. Let see what we can find...hopefully something like this one or better.

Thanks for the feedback...

Cheers,


CanOz

 

[tech/a]

With slippage and commissions your probably looksas good.

 

[CanOz]

Yeah it really Irks me that most vendors don't put in S&C. At the very least you can include the IB commissions, and a point of slippage each way.

CanOz

 

[sinner]

Sinner,

Knowing only a very little about Neural Networks and assuming without a Google that ANN is Adaptive Neural Network i can tell you that i have no experience with these.

I am currently using a genetic program called Adaptrade Builder to construct an algorithm in Easy Language code that i then test on more OOS data, then optimize and walk forward.

Were you thinking i was using some kind of neural network software?

Cheers,


CanOz

Hi canoz, was just pointing out that although my adaptive model uses much more 'advanced' principles of machine learning than genetic algorithms or ANN (without a very solid understanding of the ins and outs, these are really like the moving averages of machine learning), my own research confirmed what is coming out in the financial science (and that blog) recently about requiring an 'aggregate' rather than individual decision for machine learning trading signals.

For example, my machine learning model which has pretty good prediction rates for weekly moves in an equity index (incl magnitude) is actually a highly abstracted hierarchy of smaller models operating under a larger one.

 

[CanOz]

my machine learning model which has pretty good prediction rates for weekly moves in an equity index (incl magnitude) is actually a highly abstracted hierarchy of smaller models operating under a larger one.

Thats the 'aggregated decision' then...?

You know it wouldn't be difficult for me to aggregate decisions of several similar systems. The Builder has a habit of constructing very similar algorithms (makes sense on the same data) if you run it over the same data with only slight variations in the indicators or settings...I wonder if it would make sense to use several of these systems and wait until they all generated the same signal before taking a position...then test this on OOS data and see what the results are like....just thinking out load...



CanOz

 

[tech/a]

Sinner
What software are you using?

 

[sinner]

Sinner
What software are you using?

tech, I write all the vectorisers, training wrappers and modellers in Python (numPy and sciPy usually can do the maths but if not I plug the script into R) myself and use SVM-light (a free, very robust Support Vector Machine implementation for most Operating Systems) as my learning and classifying mechanism.

http://svmlight.joachims.org/

This is all an 'adaptation' of stuff I learned in CompSci at uni (now years ago) to do stuff like ranking http links for usefulness (think Google PageRank, our prof was a goog PhD) and spam filtering.

Here is a 2004 example from the scientific literature using SVM in a very very (I can't stress how very) simple model to predict (not incl magnitude) weekly direction in the Nikkei futs. It's a really good, inspiring paper though, recommended reading.

http://www.sciencedirect.com/science/article/pii/S0305054804000681

I notice this guys most successful techniques use SVM too

http://themarketpredictor.com/blog/

Hope you don't mind the hijack there canoz :/

Thats the 'aggregated decision' then...?

You know it wouldn't be difficult for me to aggregate decisions of several similar systems. The Builder has a habit of constructing very similar algorithms (makes sense on the same data) if you run it over the same data with only slight variations in the indicators or settings...I wonder if it would make sense to use several of these systems and wait until they all generated the same signal before taking a position...then test this on OOS data and see what the results are like....just thinking out load...



CanOz

Check out that adaptivetrading blog link I posted earlier, go back through the posts, it gives very useful hints as to

* Adaptive techniques
* How to value individual bots in a swarm
* Different techniques to trade the aggregate decision (hint: waiting for everyone to agree is not it, think more 'democratic' or 'representative', or net long/net short)

 

[CanOz]

No worries Sinner, i welcome any systematic discussion.....machine generated or otherwise...

I can't access Blogs/You-tube/FB etc...

CanOz

 

[sinner]

No worries Sinner, i welcome any systematic discussion.....machine generated or otherwise...

I can't access Blogs/You-tube/FB etc...

CanOz

Is that the China Firewall thing?

I will try and PM you some excerpts if I get a chance then.

 

[CanOz]

Is that the China Firewall thing?

I will try and PM you some excerpts if I get a chance then.

We call it the Great Fire Wall.....

Appreciate that Sinner.

I've scrapped all of my trial work from Builder and have started all over again. The first month was a learning period anyway. Now that i have my objective function sorted and a standard protocal for develop, testing, Optimization and Walk Forward I'll give all my markets another go.

I may change to my old thread "Developing a Trading system from Scratch" and show the progress, gives me a bit of a journal as well.

Cheers,


CanOz

 

[McLovin]

We call it the Great Fire Wall.....

A VPN will get around it. I have used VyprVPN in China and have had no issues with it. Whenver I travel I use Vypr, it's great if you're constantly using unsecure WiFi.

(sorry the post sounds a bit like an advert)

 

[CanOz]

A VPN will get around it. I have used VyprVPN in China and have had no issues with it. Whenver I travel I use Vypr, it's great if you're constantly using unsecure WiFi.

(sorry the post sounds a bit like an advert)

Yes and i used one for a while, they shut that down or somehow made it so slow it was useless.

I have great internet speed now and i can access what i need to conduct my business, so I'm happy for now. My wife wants a VPN for FB though, so we may get one that she can use on the iPad. Thanks McLovin.

Cheers,


CanOz