Machine learning for robots - page 10
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
Interesting to see GBPJPY, EURJPY, AUDJPY. Somehow it seems to me that one of these pairs would be better.
Is the EA trending or flat?
In this experiment, I have not divided by trend and flat, but in other experiments I've observed that it works in the same way as with volatility, time, news filters, etc.
But all this should be methodically tested and refined, for this reason I suggested a topic with templates that would clearly apply, edit, etc., although so far there are few willing to practical participation:)
You probably didn't run the tests on MetaQuotes-Demo and also used other timeframes.
I have only trained for OHLC USDCHF H4 - it's an experiment for MetaQuotes-Demo since it has a big base and the quotes from other providers may be very different.
The problem of unification of learning, so the Expert Advisor would not be sensitive to differences in quotes and that it summarizes information from different timeframes, is another problem.
For this I experiment with formulas of defining training patterns and integral price bar characteristics.
Recently an interesting solution was suggested by one of the programmers in the English part of the forum, if you have any ideas in this direction, suggest them.
https://www.mql5.com/en/forum/281402/page4
I have one direction - trend-following technologies and systems and automatic operation within the working day and working week on the most profitable currency pairs and timeframes.
I have one focus - trend following technologies and systems and automatic work within the working day and working week on the most profitable currency pairs and timeframes.
Well, I read your offer recently with a willingness to participate in EA creation and didn't have time to prepare a reply before the message disappeared... :)
The thing is that manually generated EAs are very difficult to correct, firstly, they may contain megabytes of code and sometimes I have to use command line compiler, because the built-in optimization in the editor is slow, and secondly, there are arrays of constants and weighting factors that are difficult to logically comprehend.
That's why I had to generate a new, minimized EA with short learning period on GBPUSD M15, 3 bars pattern and model with decision trees, for example, to show you some logic.
Here are some tests of this EA for different instruments, timeframes and brokers.
But to solve the main task of MO - forecasting, we need more experiments with different initial data, models, training parameters and forward testing. Eventually we need to understand and learn to use the market memory or make sure there is no memory at all :)
Well, I recently read your offer to participate in EA creation and didn't have time to prepare a reply before the message disappeared... :)
The thing is that the machine-generated EAs are very difficult to correct by hand. First of all, it may be megabytes of code, sometimes I have to use command line compiler, because the built-in editor with optimization is slow, and secondly, it is an array of constants, weighting factors, which is difficult to make sense out of logically.
Therefore, for the sake of example, I had to generate a new, minimized Expert Advisor with short learning period on GBPUSD M15, a 3-bar pattern and decision tree model, so that you could at least look through the logic.
Here are some tests of this EA on different instruments, timeframes, brokers.
GBPUSD M30 RoboForex
EURUSD M15 InstaForex
GBPUSD M15 Alpari
AUDUSD H1 MetaQuotes
These results give indirect evidence that the EA model has some generalizing ability, but to solve the main task of MO - prediction, we need more experiments with different input data, models, training parameters and forward testing, we finally need to understand and learn to use the market memory or finally make sure there is no memory at all :)
I apologize for my deleted post - I thought nobody needed it, so I have removed it somewhere else.
What a mess you have made of your EA. And all this despite the fact that everything is much easier and more accessible.
You know the nature of price movement in Forex, and the closest patterns of this movement - rising and falling Trends of different lengths and Zigzags of larger or smaller volumes.
And you can easily coordinate your buys and sells with the beginning and the end of these objects, and make almost all of the resulting profit (minus losses from the spread and insufficient quality of the working program).
I apologise for my deleted post - I didn't think anyone needed it, so I put it away.
You've made a lot of noise for your advisor. And this despite the fact that everything is much easier and more accessible.
You know the nature of price movement in Forex, and the closest patterns of this movement - rising and falling Trends of different lengths and Zigzags of larger or smaller volumes.
And you can easily match your buys and sells to the start and end of these objects and make almost all of the resulting profit (minus the losses from the spread and lack of quality of the working program).
You have explained everything simply, but I will try to simplify it, without going into the nature of currency movements, models, trends and program development, because all that, IMHO, has been covered over and over again, and it can be thought about infinitely.
It is quite another matter to use machine learning to follow the market memory: just teach the robot to trade on the price history peaks and troughs.
Of course, the learning has to be quick and of high quality, and I may have to do it often, but all this can be solved by a simple automation, especially because I already have it.
The only thing left to do is to check in practice how much the trained robot can trade by inertia and how often it needs to be changed or retrained, and which parts of the history to take.
It's like going downhill and jumping off a ski jump, acceleration, jump and fly as long as you can, then back uphill again, which is even easier:)
You have explained everything simply, but I will try to simplify it, without going into the nature of currency movements, patterns, trends and program development, because all this, IMHO, has been covered over and over again and one can think about it endlessly.
Another thing is to sit on the tail of market memory on machine learning, there is nothing to think about, just teach the bot to trade on the peaks and troughs of price history.
Of course, you have to teach it quickly and qualitatively, and you may have to do it often, but all this can be solved by primitive automation, moreover, I already have it.
The only thing left to do is to check on practice how much the trained robot can trade by inertia and how often it needs to be changed or retrained, and which parts of the history it needs to study.
It's like going downhill and jumping off a ski jump, acceleration, jump and fly as long as you can, then back uphill again, which is even easier:)
Perhaps this is also a certain variant of achieving the desired effect. Try it, maybe something will work out.
In general, the most desirable for everyone here is to get from
Trading as much as possible, or even better - ALL of the profits of the day and each transaction,
and with minimum effort of your own time and money.
Ivan Negreshniy:
Another thing to do is to sit on the tail of the market memory with machine learning, there is nothing to think about, just teach the bot to trade on the peaks and troughs of the price history.
Not by price history, but by increments - they form the price (the integral of all the increments is actually the price from the starting point).
Fortunately, for neural network experts, the 1st condition for Kolmogorov's forecasting (expectation =0) for such BP is held.
The 2nd condition - stationarity - is not satisfied.
I propose to input to the NS, besides the increments themselves, their moments: variance, skewness, kurtosis... and the autocorrelation coefficient. The NS is simply obliged to find regularities in this junk.
Well, I recently read your offer to participate in EA creation and didn't have time to prepare a reply before the message disappeared... :)
The thing is that the machine-generated EAs are very difficult to correct by hand. First of all, it may be megabytes of code, sometimes I have to use command line compiler, because the built-in editor with optimization is slow, and secondly, it is an array of constants, weighting factors, which is difficult to make sense out of logically.
Therefore, for the sake of example, I had to generate a new, minimized Expert Advisor with short learning period on GBPUSD M15, a 3-bar pattern and decision tree model, so that you could at least look through the logic.
Here are some tests of this EA for different instruments, timeframes, brokers.
GBPUSD M30 RoboForex
EURUSD M15 InstaForex
GBPUSD M15 Alpari
AUDUSD H1 MetaQuotes
But to solve the main task of MO - forecasting, we need more experiments with different input data, models, training parameters and forward testing, we need to finally understand and learn how to use the market memory or finally make sure there is no memory at all :)
Forget about forecasting - follow the price
The picture shows gold and entry points, i.e. any system always follows the price.
You have explained everything simply, but I will try to simplify it, without going into the nature of currency movements, patterns, trends and program development, because all this, IMHO, has already been done over and over again and one can think about it endlessly.
The market memory on machine learning is another thing, there is nothing to think about, just teach the bot to trade on the peaks and troughs of the price history.
Of course, the learning has to be quick and of high quality, and I may have to do it often, but all this can be solved by a simple automation, moreover, I already have it.
It only remains to find out in practice how often a trained robot can trade by inertia and how often it needs to be changed or retrained, and which parts of the history it needs to study.
It's like going downhill and jumping off a ski jump, acceleration, jump and fly as long as you can, then back uphill again, which is even easier:)
The market is constantly changing and the bot based on one algorithm will fail and let everything go down the drain.
I have not seen a better one yet.
The market is constantly changing and the bot on one algorithm will fail and blow everything down the drain...
The market is constantly changing and the bot will fail with one algorithm and it will go down the drain.