'Watching The Market' vs. Machine Learning.

[Craig66]

Here is a question I have been considering and I can't really think of a good answer...Many experienced traders on this site have recommended 'Watching The Market' in order to pick up repeating patterns which could form the basis of trading systems, the human mind is designed to pick up patterns, so no doubt, patterns will be found. On the other hand, another school of thought advocates the use of ML to discover patterns, this usually falls victim to the usual problems, given enough systems, some are bound to look good by accident if one permutes enough. The problem I can't answer is, how is a pattern noticed by a trader anymore valid than a pattern found by a machine? What is the difference? You still have to go and back-test the idea, and it still might be just a random pattern that you've noticed.

The only answer I can think of that is the experienced trader looks at the market in a completely different way to a novice, and that somehow this domain knowledge makes the difference. 'Watching The Market' seems like a pointless exercise unless you have the correct framework in which to view it, and maybe experienced traders tend to forget this when making the recommendation.

What do you guys think?

 

[Eyez]

It's all random patterns... only thing that matters is if you can make money from it or not. ML is used to parameter optimization.. but again you're fitting from the past data as nothing, even "solid non random patterns" is indicative of the future.

The difference is exp traders will look for 'patterns' or trade ideas that fit into a particular R/R profile. and also consider patterns that lose horribly.

Say you have a strategy that shows a loss 9 out of 10 times, but the winner is 10x the average loss. Is this a good strategy? Can a novice (mentally) handle 9/10 losses? regardless of patterns, at the EOD it requires trust in your system... and trust in your system builds over the time you monitor/use your system. if your a novice you don't have this experience (especially if you don't have a system, let alone a system you trust) and even if a novice has a 'good' system, its ruined by the lack of follow-thru (ie. mentally unable to handle losing streak, even tho the winner is 10x the average loss)

there is always risk-- not necessarily only market risk, but the mentality of the trader, surge, hurricane, terrorist attacks, sharknado, etc. even if the pattern is 'not random'

Quote from Craig66:

Here is a question I have been considering and I can't really think of a good answer...Many experienced traders on this site have recommended 'Watching The Market' in order to pick up repeating patterns which could form the basis of trading systems, the human mind is designed to pick up patterns, so no doubt, patterns will be found. On the other hand, another school of thought advocates the use of ML to discover patterns, this usually falls victim to the usual problems, given enough systems, some are bound to look good by accident if one permutes enough. The problem I can't answer is, how is a pattern noticed by a trader anymore valid than a pattern found by a machine? What is the difference? You still have to go and back-test the idea, and it still might be just a random pattern that you've noticed.

The only answer I can think of that is the experienced trader looks at the market in a completely different way to a novice, and that somehow this domain knowledge makes the difference. 'Watching The Market' seems like a pointless exercise unless you have the correct framework in which to view it, and maybe experienced traders tend to forget this when making the recommendation.

What do you guys think?

 

[rwk]

Quote from Craig66:
The only answer I can think of that is the experienced trader looks at the market in a completely different way to a novice, and that somehow this domain knowledge makes the difference. 'Watching The Market' seems like a pointless exercise unless you have the correct framework in which to view it, and maybe experienced traders tend to forget this when making the recommendation.

I think this is an astute observation. Some people seem to better than others at spotting patterns, but often the problem becomes one of too many, rather than too few. Trading on patterns that are not predictive get very expensive.

I don't spend much time watching the market. I find it mind-numbingly boring. And I don't think I have ever has an entirely original idea in my life. Yet I can still trade profitably. There is creativity in rearranging pieces of the puzzle, not just in finding new ones.

 

[HurricaneUS]

Quote from Craig66:

Here is a question I have been considering and I can't really think of a good answer...Many experienced traders on this site have recommended 'Watching The Market' in order to pick up repeating patterns which could form the basis of trading systems, the human mind is designed to pick up patterns, so no doubt, patterns will be found. On the other hand, another school of thought advocates the use of ML to discover patterns, this usually falls victim to the usual problems, given enough systems, some are bound to look good by accident if one permutes enough. The problem I can't answer is, how is a pattern noticed by a trader anymore valid than a pattern found by a machine? What is the difference? You still have to go and back-test the idea, and it still might be just a random pattern that you've noticed.

The only answer I can think of that is the experienced trader looks at the market in a completely different way to a novice, and that somehow this domain knowledge makes the difference. 'Watching The Market' seems like a pointless exercise unless you have the correct framework in which to view it, and maybe experienced traders tend to forget this when making the recommendation.

What do you guys think?

you can not program experience, intuition and hunches into a machine....the day that happens we will all be batteries :eek:

 

[vicirek]

Brain operates on different principle than computers.

For all practical purposes computer processes information sequentially and each complex task has to be divided into simple rudimentary operations. Once the complexity of the problem increases it is difficult to create viable working algorithm using computers.

Machine learning has many types of algorithms but it is still limited what it can do. How we recognize patterns is largely unknown and neural networks or fuzzy logic is probably poor and even incorrect approximation of what is involved in learning.

Market provides us with much simpler information. It is not 3D, it is black and white, does not have shades, can be simplified and represented as chain of numbers or simple 2D chart.

For this type of problem I would argue that if human generated algorithm based on market observation is executed as computer program and does not result in consistent performance then this observation must have be based on chain of random events mistakenly recognized as market pattern.

 

[vinc]

Quote from HurricaneUS:

you can not program experience, intuition and hunches into a machine....the day that happens we will all be batteries :eek:

actually that's a big advantage of programming ..
by the way, how do you have sex ? regular or ' out of the box' way?? / forgot to ask you in the other thread/
just curious

 

[Emil]

Of course there's no difference if the patterns is invented by a machine or a human, the only thing that matters is the performance metrics and the statistical significance of the results. Letting a computer do the testing will just speed up the process of discovery, in the end it is up to the human operator to determine if it's just a fluke or something viable.
The trick is of course how to set up the search process. I don't really think there are that many ways to skin the cat, apart from the usually stat arb strategies, liquidity provision, fundamental analysis, seasonality and momentum, what do you really expect to find? I would say it's better to start with a general theory about how markets move and let the computer do the rest of the number crunching than to throw a 100-node neural network together with emerging markets' butter production numbers and see how well it fits the S&P500.

 

[Craig66]

I suspect that this is correct.
The ability to 'frame' the problem in the correct way is the key.
Feeding the last 5 daily bars of SPY into a neural network is probably just as useless as a beginner trader taking watching the last 5 daily bars of SPY.

 

[ammo]

the market moves in herds,the patterns are just hints at continuation or reversal,sometimes they work,sometimes not,if you watch several instruments as the herd, one or two usually start to lose momentum or turn first, probably able to automate that, actually the way things work, i am sure someone has ,or its one big play and you look for the author to leave hints in the script, either way, you have to be a detective and not take anything at face value, only a piece of the puzzle

 

[goodgoing]

Quote from Eyez:

It's all random patterns... only thing that matters is if you can make money from it or not. ML is used to parameter optimization.. but again you're fitting from the past data as nothing, even "solid non random patterns" is indicative of the future.

Although QIM hasn't been doing all that great lately probably because of sheer size, maybe it pays to know what its founder has said about systems and patterns

"In the book Hedge Fund Market Wizards (May 2012), there is an interview of Jaffray Woodriff, the founder of QIM. When he was asked during the interview about data-mining he said amongst other things that “…the human tendency is to take the models that continue to do well in the out-of-sample data and choose those models for trading.” He also went on to explain how ”this process simply turns the out-of-sample data into part of the training data because it cherry-picks the models that did best in the out-of-sample period.” He then argued that instead one must look for patterns “where, on average, all the models out-of-sample continue to do well” and that “you are really getting somewhere if the out-of-sample results are more than 50 percent of the in-sample.”"

Source


If you understand in depth what Jaffray Woodriff is talking about, then you may realize why the world is full of losers. It's not ML or data-mining, it's how people use them constrained by limited knowledge and experience. I have two programs that work along those lines but are different and both have generated very robust models.