Methods of carrying out a rolling forward - page 12
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
We can debate on this topic. If you have 10,000 transactions (points in a multidimensional space that we limit ourselves to selected dimensions - variables), it's enough to enter 10,000 variables to fit all your points on a hyper-plane. This is already a barefaced fit. So, what I'm saying is.
it's wrong in principle, but there are times when it's not critical.
The validity can be detected. If changing some variable with the others fixed does not change the outcome much, significance is low. However, the interaction with other variables is lost. In fact, a system is an interaction of several variables.
It is simpler than that.
The topic of this thread is volking-forward. The fitting in its terms is optimizing on the history and substitutingthe result of optimization for the actual performance of the Expert Advisor. The lack of fitting is the analysis of results on the non-optimized, forward part of the history. If you look only at the non-optimized parts, that is the evaluation without adjustment.
That is, the only criterion for the validity of having or not having a variable in this sense is whether or not it helps to earn in unpredictable future conditions.
In other words - the best kind of fitting I know is forward fitting. But, unfortunately, you can't achieve it just by optimization, you have to change the code
It's simpler than that.
The topic of this topic is a wolfing forward. Fitting in its terms is optimization on the history and substitutingthe result of optimization for the real performance of the Expert Advisor. The lack of fitting is the analysis of results on the non-optimized, forward part of the history. If you look only at the non-optimized parts, that is the evaluation without adjustment.
That is, the only criterion for the validity of having or not having a variable in this sense is whether or not it helps to earn in unpredictable future conditions.
In other words - the best kind of fitting I know is forward fitting. But, unfortunately, it can't be achieved by simple optimization, you have to change the code.
I agree with this.
Market is full of Expert Advisors worth thousands of dollars that have at best 500 trades in optimization. It is a loss, yes.
It's not that simple, but there is another question: how much time should we spend on pre-tuning systems (learning the noise) and then seeing the bad forwards, if there is a way to pre-design a system which will be less prone to learning the noise.
I agree with that.
The market is full of EAs for thousands of quid with at best 500 trades in optimisation. This is a loss, yes.
It's not easy, it's a question of how much time to spend on pre-tuned systems (learning noise) and then seeing bad forwards if there is a way to pre-design a system that will be less prone to learning noise.
No one has yet suggested anything better than volking-forward in this sense.
Then I'll suggest it.
Dancing Forward is a new word in testing.
On the green teach, on the red test. Repeat many times.
)
Dancing Forward is a new word in testing.
Then I'll suggest it.
Dancing Forward is a new word in testing.
On the green teach, on the red test. Repeat many times.
)
It's quite an old idea. I've tried it with volcano, with my hands, but it didn't work. And both training on successful sections and on the subsequent unsuccessful ones.
The point is that we lose inertia and get a lottery in return. If only there was a way to account for the specific samplicity of the story.... -I have a couple of ideas about that and it still needs to be tested.
The trick is that there are no reliable ways to identify specific market characteristics in history and confirm their cyclicality. And without that, the process of selecting sections of the past is completely random.
What do we get as a result?
And what do we get as a result?
It's a joke of course )
But this method is similar to crossvalidation with repetitions. This approach is used when estimating model parameters on a test set. That is, validation of the results must be performed on a separate set.
Described my Walk Forward scheme using the staff tester here.
I do not use it at all, my GA has its own tester, everything is on MQL5.
PS In 2011 I got tired of asking MQ to implement this or that. I have written all my own. I use the built-in tester only for debugging before starting the demo mode in realtime.