System Development with acrary

[ig0r]

Quote from acrary:

No, they aren't curve fit. I've done some work on model 1 last year because it has an edge and I wanted to use that part of it for another model. The other two have been around for ages but I'm not doing anything with them because they have no edge. The results are based on 1 unit size (not 1 contract). A unit is determined by largest market volatility divided by the current market volatility so the number of contracts varies to keep the results consistent with market volatility. No money management was applied to any of this (if I did, it'd be obvious).

ok, cool. One more quick question, are these intraday or overnight/swing trades? Also, I agree with you that combining non-correlated systems improves results but won't they occasionally contradict themselves? For example, a trend following system and a range trading system, one says it's a breakout and to buy while the other is signalling a sell

 

[acrary]

Quote from axeman:

I may have missed it...but what is the starting capital for these results?

peace

axeman

I didn't post starting capital because these aren't complete systems. They're just volatility adjusted so I can easily compare one year with another. Here's the annual summary for model2 so you can see that say 2003 was doing about as well as 2001. It's pretty obvious 2002 was way outside the normal bounds of the model.

 

[rickty]

acrary,

I've been following this thread with considerable interest, thanks for starting it.

I have a question (maybe not too related to this thread) that I've been wanting to ask you for a while now. A few years back you wrote at Chuck LeBeau's site that you had used a neural net to predict trend up days with accuracy of better than 80%. I've been wondering what sort of neural net you used? Was it a regular feedforward network with a backpropagation learning algorithm?
Or do you need to use a recurrent network for this sort of problem. (You mentioned you used Brainmaker, which I believe also has a recurrent network). Any information would be much appreciated. (Please PM me if you think this is more appropriate).

Thanks,

Richard

 

[acrary]

Quote from opm8:

acrary,

Are your systems entering orders themselves or do they require someone to do it?

I'm working on my own system and obviously, automated order entry is a big step in terms of "letting go."

Do you see a benefit to one way or the other?

--opm8

When I'm around I trade against the models. I use them as competitors. I'm a competitive freak. If I know where the models want to buy or sell I watch the market and try to front run them and stomp on their results. When I'm not around someone enters the orders and trades them mechanically. I'm no technology guru and really have no interest in automating my trading.

I don't know much about "letting go" as I know I can beat any model I develop. If I couldn't then I'd pack it in and let them all be traded mechanically. My only involvement would be to monitor the edge within the models to ensure it hasn't started deteriorating.

 

[acrary]

Quote from ig0r:

ok, cool. One more quick question, are these intraday or overnight/swing trades? Also, I agree with you that combining non-correlated systems improves results but won't they occasionally contradict themselves? For example, a trend following system and a range trading system, one says it's a breakout and to buy while the other is signalling a sell

All intraday breakout models based on different criteria.

At this point I'm just posting the material as background info.
If the methods are randomly correlated or non-correlated, sometimes the results will cancel each other out and sometimes they'll pile on with multiple orders going the same way. The idea is to improve consistency to the point where you feel very confident. Once you're there you can apply aggressive mmgt strategies to maximize the profitability within the consistency.

If I started out with entries and exits we'd never get to consistency and most everyone would bail on a good method once it has a small drawdown. Since the account can only be either making equity highs or in a drawdown I think this topic should be addressed first.

 

[ig0r]

Quote from acrary:

All intraday breakout models based on different criteria.

As this point I'm just posting the material as background info.
If the methods are randomly correlated or non-correlated, sometimes the results will cancel each other out and sometimes they'll pile on with multiple orders going the same way. The idea is to improve consistency to the point where you feel very confident. Once you're there you can apply aggressive mmgt strategies to maximize the profitability within the consistency.

If I started out with entries and exits we'd never get to consistency and most everyone would bail on a good method once it has a small drawdown. Since the account can only be either making equity highs or in a drawdown I think this topic should be addressed first.

excellent response, thank you

 

[acrary]

Quote from rickty:

acrary,

I've been following this thread with considerable interest, thanks for starting it.

I have a question (maybe not too related to this thread) that I've been wanting to ask you for a while now. A few years back you wrote at Chuck LeBeau's site that you had used a neural net to predict trend up days with accuracy of better than 80%. I've been wondering what sort of neural net you used? Was it a regular feedforward network with a backpropagation learning algorithm?
Or do you need to use a recurrent network for this sort of problem. (You mentioned you used Brainmaker, which I believe also has a recurrent network). Any information would be much appreciated. (Please PM me if you think this is more appropriate).

Thanks,

Richard

I'll make this quick because it's way off the current topic. I have both Brainmaker Pro with the GTO component and Neuralworks Professional II/Plus. I no longer use neural networks for anything (replaced with something better). At the time I used Brainmaker for the GTO component to learn what inputs were significant to the output. I then used Neuralworks to find the weights between the inputs to determine what level of importance each input had. It was a backprop network.

 

[rickty]

Quote from acrary:

I'll make this quick because it's way off the current topic. I have both Brainmaker Pro with the GTO component and Neuralworks Professional II/Plus. I no longer use neural networks for anything (replaced with something better). At the time I used Brainmaker for the GTO component to learn what inputs were significant to the output. I then used Neuralworks to find the weights between the inputs to determine what level of importance each input had. It was a backprop network.

I appreciate your reply. Thanks.

Richard

 

[acrary]

One other thing about correlation I wanted to post was that even if we found a negative correlation, do a couple checks.
If we had say -.4 between two models you might think you found a great negative method to complement your primary model. Another way I found to check it is to create a table of correlations with rolling periods (in the case of monthly 12 periods). If the correlation isn't consistent and smooth then you're probably looking at fools gold (expect the correlation to revert to random over time in the future).
Using our 3 models I've put in columns for the correlations between each of the models. You can see by inspection that the 12 month rolling correlations are moving around wildly. This is a pretty clear example of random correlations between the models.

 

[acrary]

Correlation cont'd.

Here's a screenshot of crude versus unleaded gas using 30 day average correlations between the two. Over the past 5 years the correlation was greater than 95% so you'd expect them to be highly correlated. This is what it looks like when they are.