The perfect mechanical trading system.

New comment
 
Hello programmers and coders, philosophers and pragmatists :) I propose to develop an idea of creating a subgame.
For this we will share our thoughts and try to jointly write this incredible wonder :)
In general, in order. About me: I'm not a programmer, but I have a big trading experience (more than 5 years) I'm completely satisfied with my trading and earning. But even after all these years, I always come to the conclusion that the experience and knowledge of finanosmic markets is impossible to comprehend completely and thoroughly. There is always a learning process. What changes over time? The percentage of intuitive, analytically correct conclusions about the direction of the market. That is probably all. Like all of you I have probably been searching for grails for a long time and have become convinced of their absence, just like you. Actually the end of demogoguery - let's get back to the topic :).
And so. Let's take a priori the rule that there are a lot of profitable mechanical trading systems, and in a certain period of time and for a certain currency, mts can give good results. This is why whatever mechanical system we use as the basis, the main thing that should be in it - not a single user's tuning parameters. All parameters must be calculated independently, i.e. it will be an automatic mechanical trading system (AMTS) that is optimized according to the volatile market.
In this branch I propose to share my thoughts on this subject and write little by little some parts of the code that would implement the stages of optimization.
I repeat - I'm not a programmer, so I will start with the simple things.
We will divide self-adjustable parameters of AMTS into main and additional.
Basic:
1. Number of lots.
2. Take profit
3. Stop Loss
4. Trailing Stop
Additional
5. All those parameters on the basis of which AMTS will be written.

As the basis of MM we shall take: entrance 5% of equity, profit factor not lower than 2, take profit and stop loss ratio: 60/40.
1. Well first, the number of lots in many systems is implemented: this is the LotOptimizator function.
Let's assume that our system will be intraday, respectively.
2. Take Profit will be calculated as 70% of the average daily price movement of the last month.
3. Stop Loss, respectively, as 40% of Take Profit.
I suggest to calculate Stop Loss and Take Profit levels at 01:00 a.m., after going to the next day.

In general, if the topic is interesting, let us express our ideas.
 
You can retrain the neural network at each new bar. Or retrain it. The task comes down to building the model.
 

You can retrain the neural network at each new bar. Or retrain it. The task comes down to building the model.

You must have worked with neural networks. And you know that multilayer neural networks do not learn quickly. And if the timeframes are short, the network will not have enough time to work out a solution.
I haven't worked with neural networks, I'm just reading. How much progress have you made?

My idea in this direction is self-adaptive genetic algorithm LGAP. I'm working on it. Almost ready dll - template.
 
My imho - it would be a fit. I tried neural networks. It wasn't much use, to be honest. And I used a pretty good one!

The idea always lies with the trader, because one cannot do without a fundamental reasoning here. And a network will be engaged in data fitting - it has too many degrees of freedom - I've checked it in practice.
 
kniff писал (а):
My imho - it would be a fit. I tried neural networks. It wasn't much use, to be honest. And I used a pretty good one!

The idea always lies with the trader, because one cannot do without a fundamental reasoning here. And a network will be engaged in data fitting - it has too many degrees of freedom - I checked it in practice.

That's the way it should be. The market is volatile: volatility, volumes, price ranges. Parameters have to change quickly.
 
quality писал (а):
Hello programmers and coders, philosophers and pragmatists :) I propose to develop an idea of creating a subgame.

Personally, I am very interested in this subject! Ready to take part in every way!

About selftuning parameters: with basic everything is clear, but what to take as the basis for additional parameters, ie, what indicators, levels, channels, or what?

I had such an idea:
- put several indicators on the chart (e.g. RSI, Stoch, CCI, MACD, etc.) pick up "approximately" the values of these indicators;
- then, look at the history for approximate price reversals (i.e., where it is clearly visible "here we have to buy, here we have to sell");
- then write down the values of all indicators in these points, for buy and sell, in an array or in a file;
- further, in the Expert Advisor - check it (taking into account the deviation of indicator values in percentage from the ideal values), ie, for example, in the array, the RSI value for buy turned out to be 20, then, if the percentage of triggering is 10, then the buy will trigger from 18 to 22, as well as with all other indicators;
- Dale, you can (or should:) also add to the check crossing indicators of different levels or their signal lines;

I haven't checked it myself (although I started to write an experimental EA, I have no results yet), so I can't say anything about it working or not.
 
quality писал (а):

That's the way the fit should be. The market is volatile: volatility, volumes, price ranges. Parameters have to change quickly.

By the way, parameters of all indicators (in this case, we will have to make them self-adjusting as well, but how, I don't know yet). In my example (above) you can try to change the percentage of deviation of indicator values from the ideal, i.e. try to follow the market changes, you can and probably should add new indicator value combinations to the array (file).
 
I've been thinking about it for a long time myself. I just don't know from which end to approach it. If there are good, well-founded ideas, I'm ready to implement them. At the moment I am very interested in the pattern ideas outlined in 'Self Learning EXPERT' in combination with a neural network.
 
kniff, it seems to me that history matching is a very, very useful thing if it is TIME-appropriate. That is, if the system adapts quickly enough. So there's no need to be afraid of it. It's better to think about how to adapt to the story fast enough. For example, the system possesses some vector of settings X. Each moment of time t, for this vector there exists some optimal value X' that doesn't change too quickly. I.e. X'(t0) is close to X'(t1), if t0 and t1 are close. If in time dt=t1-t0, the system manages to correctly determine the value of X'(t0), then by time t1 it will be the value of the setting. I.e. not optimal (the optimal value will be X'(t1)), but close enough to the optimal one. In this connection here is a question for my colleagues who were quite heavily engaged in testing (alas, I can not boast with it). How abruptly do naive (not adaptive, rigidly adjustable) systems start to fail when market conditions change? Is there some kind of transition point from profitability to plummeting ? Or does it happen abruptly and catastrophically?
 
favoritex, could you elaborate on what LGAP is? I couldn't find anything about it on Yandex, except a mention that it's such a thing. And the context of the mention seemed interesting to me.
 
eugenk1:
It seems to me that history adjustment is a very, very useful thing, if it is TIME-appropriate. That is, if the system adapts quickly enough. So you don't have to be afraid of it. It's better to think about how to adapt to the story fast enough. For example, like this: The system possesses some vector of settings X. Each moment of time, this vector has some optimal value X', which doesn't change too quickly. I.e. X'(t0) is close to X'(t1), if t0 and t1 are close. If in time dt=t1-t0, the system manages to correctly determine the value X'(t0), then by time t1 it will be the value of the settings. I.e. not optimal (optimal will be X'(t1)), but close enough to the optimal. In this connection here is a question for my colleagues who were quite heavily engaged in testing (alas, I can not boast with it). How abruptly do naive (not adaptive, rigidly adjustable) systems start to fail when market conditions change? Is there some kind of transition point from profitability to plummeting ? Or does it happen abruptly and catastrophically?
This is exactly what I would like, but no :( Apparently, it all depends on the system. So far I have the impression that everything happens exactly "abruptly and catastrophically". In general, try Phoenix with standard settings - before February 13 this year 100% loss, and after - abruptly begins a high level of profitability with high - stability. Imagine the reverse transition.
123456789101112
New comment