There is no requirement that correlation be either linear or simple. Correlation is any departure of two or more random variables from statistical independence.
You are using an out of date browser. It may not display this or other websites correctly.
You should upgrade or use an alternative browser.
You should upgrade or use an alternative browser.
Has anybody tried Neural Networks? Does it work?
- Thread starter mizhael
- Start date
Quote from jimbojim:
Data mining for rule-based statistically significant patterns of price action is the way to go. There is one program that I know of which combines data mining with entry/exit to find patterns from historical data that exceed certain limits in performance parameters values you set in advance. It says now on the website that is not sold any longer (website) There is another one I have found, TSL, but its price tag is prohibitive (website). This one is based on Genetic Programming.
I just checked the APS website and I do not see the note you mentioned. Regardless, APS is one of the better investments in a trading tool I have ever made. It can help you understand price action and if you know how to use it properly you can get great results. This is not a tool for beginners who are looking to be spoon-fed. The same applies to TSL. This kind of advanced tools require an understanding of trading system evaluation and application. I think only 1 out of 1000 traders can use them properly. If probability and statistics was not one of your favorite courses in college, then I do not recommend getting any of these tools, they will not help you at all.
Most of the times these question seems to be asked in order to find out if there exist a tool that gives you the holy grail trading system. So far I havenât seen such a tool.
Even so, I find neural networks useful. But, as with all tools, you need to understand what it does and how it does it before it becomes useful. As dtrader98 among others often points out, it is important to know what information to use if youâre going to be succesful.
Moreover I believe another major problem is how people classify their training examples. You must remember that the outcome (profit/loss) from a financial time series is very different from other types of problems where NN is used (e.g. character recognition). This means that the outcome in the training examples are unstable with respect to the input values, so that input values for one example that are very close to another example in feature space could have a very different outcome (one having a 10% profit and the other a 10% loss). This often confuses the standard learning algorithms relying on error measurements for training. To get âbetterâ training results users often adds a lot of extra nodes in hidden layers in order to separate these two examples, even if the difference in results was the due to some corporate news and could never been deducted from the input values. In the end this reduces the walk-forward performance because of the over-fitting to unstable examples.
I have addressed these problems by using a different learning algorithms where we do not use individual examples (sometimes called unsupervised learning) and different goal functions that is not based on error measures. This has lead me to using a genetic algorithm to optimize the network. This gives the freedom to use almost any goal function.
Even so, I find neural networks useful. But, as with all tools, you need to understand what it does and how it does it before it becomes useful. As dtrader98 among others often points out, it is important to know what information to use if youâre going to be succesful.
Moreover I believe another major problem is how people classify their training examples. You must remember that the outcome (profit/loss) from a financial time series is very different from other types of problems where NN is used (e.g. character recognition). This means that the outcome in the training examples are unstable with respect to the input values, so that input values for one example that are very close to another example in feature space could have a very different outcome (one having a 10% profit and the other a 10% loss). This often confuses the standard learning algorithms relying on error measurements for training. To get âbetterâ training results users often adds a lot of extra nodes in hidden layers in order to separate these two examples, even if the difference in results was the due to some corporate news and could never been deducted from the input values. In the end this reduces the walk-forward performance because of the over-fitting to unstable examples.
I have addressed these problems by using a different learning algorithms where we do not use individual examples (sometimes called unsupervised learning) and different goal functions that is not based on error measures. This has lead me to using a genetic algorithm to optimize the network. This gives the freedom to use almost any goal function.