Optimise an EA and get the best of the optimised ones. - page 3

 
Serqey Nikitin:

Reading statements like that, I always want to ask: "You must be God...?"

Even the most distant person from trading, by simple reasoning, will tell you that knowing everything is impossible in principle, and there will always be CULIBINES who will be able to create a profitable strategy! This is the law of development of complex systems...

You don't have to be in the soup to get a rough idea of how a chicken feels. It's enough to stick your finger in the boiling water.

Certainly, given the infinity of the universe, one might think that there are TCs that are always profitable.

But I haven't met them.

And in my experience - the balance of profitable and unprofitable time - depends very little on the complexity of the system. That's why I came to the conclusion that efforts should be directed not at creating a complex super-system, but at creating many simple systems.

Which is what I am doing, and what I invite those who wish to participate in.

 
Aleksey Panfilov:

The suggestion seems understated, from experience how much would it cost to optimise with your algorithm in "Cloud".

Or, alternatively, how long would it take to optimize if you allocate 4 average processors to it?

Btw, first question on the merits.

For me, it takes four hours on four i5 cores. I don't see the point of engaging the Cloud, although if someone wants to, no problem, it greatly speeds up testing. But - it's "on their own". Those who help - I give access to the best TC from the general "pool".

Let's do this - I lay out the most "resource-intensive" TC, you run it, and assess whether it is worth messing with it.

In half an hour, I will get to my computer, and I will post the expert.
 
Roman Kutemov:
I don't have private messages.
On this topic https://www.mql5.com/ru/forum/231536/

Skype me please

(hilarious... Skype says my wi-fi connection is wrong... Correct for browser, correct for terminal, but wrong for Skype... (Hilarious...)

 
George Merts:

One of the indispensable characteristics of a working vehicle is stability...

Stability is the 'blue dream'...

I found it in "constant variability"... (that's where I live)


 
prikolnyjkent:

(woof woof woof ... Skype says my wi-fi connection isn't right ... It's right for my browser, it's right for my terminal, but not for Skype... (Geez...)


There was something similar, I don't remember how I solved it. Try it from the computer.
 

All of my EAs use the simplest patterns and rules.

The main trick is to "cover" as many variants of market behaviour as possible, and let in those experts who are currently showing good results. Experts showing critical results are stopped (automatically) and need to be re-optimised.

I am attaching one of the most resource-intensive (for MT5, but working versions - I can offer it for MT4 as well).

-----------------------------------------------------------------

Optimization interval - one year. 13.03.17 - 13.03.18

The timeframe - any from M5 to H4, the Expert Advisor selects the best. I usually put it on H1

Forward - Custom, 13.08.17 (five months back, seven months forward).

Mode: No lag, OHLC on M1.

Initial deposit 10000, leverage 1:100

Optimisation is fast, Custom max (the fitness function shows the "gravity percentage" - an integral score from several components)

I attach the set-file, it contains desired ranges, total number of variants - 7993824300.

Optimisation - fast.

--------------------------------------------------

You can estimate the duration of optimisation. It is better to use cross-pair, say NZDJPY (it is slower)

Files:
EMATrendDTS.zip  762 kb
 
Petros Shatakhtsyan:

It's obvious that you have no practical experience.

Firstly, it raises the question: where did you get so much "TC", from the junk drawer?

You have no idea how much time and money you need to spend for optimization in MQL5 clod to develop only one profitable Expert Advisor.

I have all these TCs in operation.

I don't need optimization in the cloud, I optimize everything on my home computer, and I'm satisfied.

No practical experience? Probably. So, my suggestion - you just won't be interested. It's perfectly clear. I show TC I have (I repeat, all of them with very simple principles, anyone can write them without much trouble). Anyone who wants to spend a year optimizing my TC, and gets for 3 months any TC of the "favorites".


 
Maxim Romanov:

It took me seven months to develop the last robot, just to develop the algorithm, not even to program it yet. So 200 is a big number, of course.

My first robot took over a year. It used 20 complicated patterns, trendlines, smart money management... And after three months of work it started to fail, and I lost everything I had earned. At the same time, as I see it, there are simple TS that give exactly the same result.

So I don't want to write one complicated trading robot anymore. There will be a huge pile of simple ones. And I will be engaged in selecting from them working ones.

 
Mihail Marchukajtes:
All the same is implemented with the help of neural networks, only much easier. You always have the same TS and principle of approach to the market. You train repeatedly, always getting different models, then choose the one that will work. Instead of a bunch of EAs, each with its own algorithm, 200 pieces. HORROR!!!

All algorithms are the same, there are 16 of them per symbol. Two options for trend detection, two entry directions, forward/reverse trailing, reversals, fixed TP-SL. Everything is very simple.

A neural network is a dead end. A neuronet extracts patterns but it doesn't possess common sense. As a result - it will find "statistical artefacts" much more likely to get in the way.

I take 16 variants of TS, in my opinion, covering the maximum number of different market behavior. Some of them will definitely work.

 
Petros Shatakhtsyan:

Not just time. A good TC, once the optimum mode has been selected, should not often require over-optimisation.

I have a simpler approach to this question. There are "limit parameters", as soon as the TS exceeds them - that's it, it stops automatically, and I direct it to re-optimization.

It is clear that TS which doesn't correspond to the market (say, trend-following on flat symbols) - will often exceed the limits, and often requires re-optimization. But it is necessary to see that the market has not changed, and that the trend TS still does not work on this symbol.

I repeat - if anyone is not interested, I'm not imposing anything.