Advisor for an article. Testing for all comers.
It is quite clear that a successful forward is not necessarily in the best optimization results, i.e. it still needs to be found by gradually running the tests.
It is quite clear that a successful forward does not necessarily have to be in the best optimization results, i.e. it still needs to be found by gradually running tests.
And why look for a successful forward? It turns out to be a fit, but only manually. Have you tried, by the way, some other criteria for over-optimisation. Here are, for example, two options:
1. optimise parameters after the maximum drawdown from the previous optimisation has been reached.
2. optimizing parameters if there is no profit within a certain period of time, like in the previous optimization.
These variants can even be tried together. And the search of a successful forward is, in my opinion, some unnecessary activity.
If you really really need it, you can get this one (layers and neurons in any number):
If you really need it badly, you can get this one (layers and neurons in any number):
https://www.mql5.com/ru/articles/252
Why do you need to look for a successful forward?
If you have an Expert Advisor that never adjusts itself and always produces successful forwards for all successful optimization results, then of course there is no need to look for anything.
Not everything is lying on the surface in gold bars the size of your fist.
It turns out fitting, but only manually.
Maybe it does, maybe it doesn't? No one will give a 100% guarantee, as not all that glitters is gold. But confidence in an EA which has given positive results outside of optimization sampling, i.e. on data which the EA has not even guessed, is always higher. Checking for lousiness with an exam is not my invention.
Have you tried any other criteria for over-optimisation, by the way. For example, here are two options:
1. Optimising the parameters after the maximum drawdown from the previous optimisation has been reached.
2. optimizing parameters if there is no profit within a certain period of time, like in the previous optimization.
You can even try to use these variants together.
No, of course not. After all, I have carefully read Grandfather Krylov's fable called "Quartet".
If you are so eager, you made it up, try it yourself. Maybe it will work?
If no one is forcing or coercing you, you do not have to look for it. Testing is voluntary, not forced.
Reshetov:
...
Maybe it does, maybe it doesn't?
...
You're reacting painfully. You can keep your bayonets. You make it up, you test it. :)
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
You agree to website policy and terms of use
I am now completing the article titled "Why is the neural network retrained?
Created an EA for it, which uses an algorithm that corrects and makes it difficult to retrain the network.
It needs to be tested on different instruments, timeframes, quotes from different brokerage companies.
How to do testing?
1. Loading Expert Advisor in experts folder of the terminal
2. Load the quotes, the more the more the better.
3. In the strategy tester set the appropriate symbol, timeframe, simulation by opening prices, download settings from attached archive.
4. Running a single test of the Expert Advisor
5. Look at the number of deals in the results. Divide the number of deals by two.
6. Using the mouse on the test chart, look for a deal with a half number from step 5. In the pop-up tip, find out its date.
7. We look from which date the quotes are loaded, set the date in the Strategy Tester from the beginning of quotes and to the date found in step 6.
8. Launch optimization (it should be quick, even on weak machines, only a few minutes).
9. Disable the date in the tester and look for successful forward testing in the optimization results, i.e. Profit with not too significant drawdowns must be from the beginning of the chart to the end.
And in this topic it is desirable to share results and opinions.
It is quite clear that successful forward does not necessarily have to be in the best optimization results, i.e. we should still find it when gradually running the tests.
For example, the top test results may contain the following result (optimization from 1 to 356 deals, then forward):
And below this one (optimization from 1 to 471 trades, then forward):
Expert Advisor's settings (in attached ZIP archive):
x0, x2 ... x7 - weighting coefficients. It is better not to touch them, but leave them as they are (they will be optimized).
sl - Stop Loss and Take Profit in pips (optimized).
mn - magic number (not optimized)
d - number of significant digits in position volumes, i.e. for lots (not optimized). If you leave the default setting as 1 lot, you can leave it unchanged.
I have already checked all parameters that have to be optimized.
The Expert Advisor's code is compiled. But there are no limitations in the code itself, so if someone wants to test it not only in the Strategy Tester, I won't mind. The source code of EAs on mql4 and mql5 will be published in the article (I don't want to open it before the article is published, but anyone who needs it will know how to dig it out).