Many/Single indicators with standard settings

Hi,

A single indicator with standard settings is considered a good and not curve fitting indicator.

However, When I put many indicators with all standard settings on the chart, It is considered a curve fitting system although it is profitable in back test?

Is curve fitting is not only about parameters?

Thanks in advance.

Yes, the more indicators you're using for your signals, the higher is the danger of overfitting, even with standard settings.

The reason is that overfitting does not depend merely on parameters, it depends on degrees of freedom. Any parameter is a degree of freedom, but any additional indicator also.

Thanks.

Is MTF(each TF is used only one indicator) also in the danger of overfitting?

Sure thing. Just think in terms of degrees of freedom, as jcl nicely put. The first degree of freedom is what is your first TF? The second one is your second TF, and so forth... Then when you come to indicators, say you're using MACD. It has three degrees of freedom by itself, because of three parameters it takes to be exactly defined (12, 26, 9). Apply accordingly to others indicators also. Soon you'll see that you have so many parameters to choose from, that your search for a holy grail has quite some chance to succeed. Unfortunately, it will be so perfect only for the time period of your backtest, aka past time. But will fail miserably going forward, out of sample. Obviously, you overfitted. Or said in other words, you've been fooled by randomness.

There's one technique that can help you to avoid overfit, and that is to carefully check your parameter space after optimization. For example, once I had a "great" trading strategy with these parameters: 17, 9, 1. Backtested over 5 years, it gave excellent results. But, as soon as I changed 17 to 18, it would fail. Obviously, not a good trading system. Your final parameters should allow for quite some tolerance, if you expect the trading system to be robust out of sample.

One other way to attack this issue is to use discretized values, so to skip large parts of parameters space and thus lower the chance of overfit. For example, if you're tuning the period of moving average between 50 and 200, don't use all values, but go in steps of say 5, like 75, 80, 85, etc... But also apply the principle from my last paragraph, if you're doing great with SMA80, but results are bad with SMA75, then you probably overfitted.

Back to your example of MTF. Say you're going to use 5 timeframes. That's 5 parameters that can produce a very large number of possible combinations. You're practically guaranteed you will find at least one combination which will produce excellent results. Do you really think it would survive the test of time?

I hope my english is good enough to attempt to explain these very important concepts in automated trading.

And after such a long reply, maybe these additional short answers would make more sense.



Obviously, you're reserving the right to choose which indicator exactly to use? So, that's you're degree of freedom. And yes, you've just increased the possibility to overfit, because you granted yourself that degree of freedom. It doesn't matter that you will use only standard parameters, because there's so many indicators to choose from.



Now you also have very large number of indicator combinations to choose from. And you can optimize the hell out of it, until they perfectly fit. Not bad for only standard settings.



Obviously not.

Thank you for the great answers.

So it looks like that a forward testing is the king to distinguish between a profitable system or not.

Maybe a curve fitting system would make money in the short period of that test, but in the end it would be a looser.

Is there really no good systems that use many indicators?
Do you know an example?

Yeah, forward test is definitely a king. Also it doesn't necessarily mean that a system with many indicators can't be a winner. It all depends how you stitched everything together. If you make sure that your system is not overly sensitive to any of the indicators involved, you're on the good path, I guess.

I can see that there is a system with many indicators in forum like forexfactory.
But most of them are a manual system.
Does degree of freedom apply for it?

Generally speaking, yes.

Does MACD without a parameter input option mean that there are no parameters with it?
I don't think so.
Also, I think that there is a parameter that is separate with degree of freedom, like an alert input with indicators.

MACD without parameter just means that the defaults are used.

I learned that in an uncertain world like trading the important is highly-diverse.
But at same time I think it would lead to curve fitting in trading?

What do you think about?
Is it not a good thing to spread a net as far as possible?