Has anybody tried Neural Networks? Does it work?

[dtrader98]

Quote from EliteThink:

we pulled the plug on nn research when the need for data filtering became aparent. If this is no longer the case 'would love to know...

One of the things covered in that blog was precisely the need for data filtering. You are already ahead of 99% of the crowd in finding a useful requirement. Congrats on that. Keep in mind there are also many ways to filter, some better than others.

I've found other applications where filtering is not a requirement, however, opportunities are much more sparse (think 3+ sigma information). Trade-offs will always be present.

 

[promagma]

Great blog, thanks.

Would it be correct to say -

If you plug in the AAPL stock price and MACD values and expect the neural network to calculate a winning system, you will be disappointed.

But say you are trading oil, and your main strategy is noticing that the gold price tends to lead oil. So you use the gold price to determine long and short opportunities. For the sake of discussion lets say that it is a real edge, a trained monkey could profit from it. I think this would be a perfect scenario to apply a neural network to discover additional filtering critera, money management strategies, etc. to squeeze every last penny.

 

[dtrader98]

Quote from promagma:

Great blog, thanks.

Would it be correct to say -

If you plug in the AAPL stock price and MACD values and expect the neural network to calculate a winning system, you will be disappointed.

But say you are trading oil, and your main strategy is noticing that the gold price tends to lead oil. So you use the gold price to determine long and short opportunities. For the sake of discussion lets say that it is a real edge, a trained monkey could profit from it. I think this would be a perfect scenario to apply a neural network to discover additional filtering critera, money management strategies, etc. to squeeze every last penny.

Suppose you did a scatterplot of many different instruments (sometimes known as draftsman's plots) showing correlations. Ok, and in a perfect world they were all straight lines. Well, believe it or not, this type of research is prominent in financial quantitative circles (see BARRA/factor analysis). BUT, what if the instruments did not show a straight line fit but some other odd shape? You might throw away the results if the correlation (linear) was very low. A NN, however, thrives on this type of situation.

 

[mizhael]

Quote from hoodooman:

I followed a guy for a while who claimed to be an expert on NN. His picks weren't good.

Okay, your observation is: Outright picking using NN doesn't work.

 

[mizhael]

Quote from stevegee58:

Exactly what I said before. You have to put some bounds on what you're measuring.

For instance there have been statistical studies of candlestick patterns. It would be straightforward to train a NN to find which 1-, 2-, 3- etc candlestick patterns are predictive. For instance certain candlestick patterns have a 60% or 70% reliability for daily bars for stocks. A NN could find if there are candlestick patterns with similar reliability on other instruments and time frames.

Any pointers to more details? Interested!

 

[mizhael]

Quote from promagma:

Great blog, thanks.

Would it be correct to say -

If you plug in the AAPL stock price and MACD values and expect the neural network to calculate a winning system, you will be disappointed.

But say you are trading oil, and your main strategy is noticing that the gold price tends to lead oil. So you use the gold price to determine long and short opportunities. For the sake of discussion lets say that it is a real edge, a trained monkey could profit from it. I think this would be a perfect scenario to apply a neural network to discover additional filtering critera, money management strategies, etc. to squeeze every last penny.

How is NN related to money management strategies. Curious...

And how do you conclude gold leads oil?

 

[mizhael]

Quote from dtrader98:

Suppose you did a scatterplot of many different instruments (sometimes known as draftsman's plots) showing correlations. Ok, and in a perfect world they were all straight lines. Well, believe it or not, this type of research is prominent in financial quantitative circles (see BARRA/factor analysis). BUT, what if the instruments did not show a straight line fit but some other odd shape? You might throw away the results if the correlation (linear) was very low. A NN, however, thrives on this type of situation.

Correlation is not causality or leading/lagging relation.

You need causality or leading/lagging relation in order to trade...

 

[EliteThink]

Quote from dtrader98:

One of the things covered in that blog was precisely the need for data filtering. You are already ahead of 99% of the crowd in finding a useful requirement. Congrats on that.

Wow, thanks for making me feel smart.

Seriously, it is difficult enough working with real clean data. The idea of a NN is enticing but I don't know one trader who uses it successfully. I'm sure someone is out there though...

 

[dtrader98]

Quote from mizhael:

Correlation is not causality or leading/lagging relation.

You need causality or leading/lagging relation in order to trade...

I think you misunderstand. Correlation is one type of relationship; a very simple, linear, and limited one at that. There is no need to limit yourself to linear relationships, nor non-lead/lagged relationships for that matter.

 

[promagma]

Quote from mizhael:

How is NN related to money management strategies. Curious...

Sorry I was confusing NN and genetic optimization .... for money management you could do a genetic optimization.