Machine Learning in trading
In my experience, systems that perform well outside the sample period have only a few optimizable parameters.
The problem with machine learning is you will introduce many more (hyper)parameters. This in turn makes it prone to curve fitting as can be seen on most if not all machine learning articles on this website resulting in poor performance out of sample.
I have done my fair share of research and practice with full machine learning algorithms as well assisted/hybrid algorithms and my conclusion is that a problem need solving often is best solved by the simplest solution (baseline) and that using machine learning to solve the same problem only introduces more problems and in most cases does not outperform the baseline. In the cases it did outperform it was difficult to understand exactly why and thus make it stable over time ie when to retrain/train continues/ how many epochs etc etc etc.
In my experience, systems that perform well outside the sample period have only a few optimizable parameters.
The problem with machine learning is you will introduce many more (hyper)parameters. This in turn makes it prone to curve fitting as can be seen on most if not all machine learning articles on this website resulting in poor performance out of sample.
I have done my fair share of research and practice with full machine learning algorithms as well assisted/hybrid algorithms and my conclusion is that a problem need solving often is best solved by the simplest solution (baseline) and that using machine learning to solve the same problem only introduces more problems and in most cases does not outperform the baseline. In the cases it did outperform it was difficult to understand exactly why and thus make it stable over time ie when to retrain/train continues/ how many epochs etc etc etc.
You got really great point. Machine learning is considered as a black box system, which mean, for most of us, it is hard to understand what is going on inside the system.
But still, if one prefers the blackbox system and trained sufficiently to deal with them, you can go for it.
Anyway, soon or later, we will be seeing it is robot broadcasing all forex and stock market news and analysis in the near future. :)
It is only a black box for those who do not understand or for self proclaimed data scientists adding to the hype that something mystical is going on. In reality there is nothing black box about it.
Granted, it is difficult to *debug* in lack of a better word, but there are methods to gain more insight. This however, adds to the complexity.
Most machine learning EAs are basically a person saying "give me a currency chart from last year, and I teach you how to trade this chart and earn millions".
All technical analysis is based on what happened in the past. If you think that there is not a certain type of reproducibility in the behaviour of the markets, there would be no profitable trading...
I think machine learning is good in finding patterns and trade those patterns.
Just don't overoptimize the strategy because it will fit to the data point.
You can read book " Advances in Financial Machine Learning " and you will know why machine learning algorithm failed in financial market and how you can overcome this issue.
I think machine learning is good in finding patterns and trade those patterns.
Just don't overoptimize the strategy because it will fit to the data point.
You can read book " Advances in Financial Machine Learning " and you will know why machine learning algorithm failed in financial market and how you can overcome this issue.
Thanks a lot. I already read that book, it is really interesting. I believe that sets of parameters which were profitable in the past 9 years cannot stop all together giving profit... What do you think?
It is just some backtesting results. It is very hard to say that past performance will continue in the future.
It is just some backtesting results. It is very hard to say that past performance will continue in the future. Defintely not.
Even if the optimization is renewed every week?
That is probably tough question to answer.
But still, there is saying that Past performance is not indicative of future results .
Optimization is based on the past data, isn't it?
I think it is should be the case, although you might have some control over this by exploring different neural network models and different optimization techniques.
If your robot really understand the human language and they can undersatnd the impact of what US president says to the financial market, it might be different story though. :)
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
You agree to website policy and terms of use
Hello,
I would like to know from all of you what do you think about the possibility to automatically generate, every week, new strategies based on continuous optimization of trading parameters through backtesting.