Discussion of article "Machine learning in Grid and Martingale trading systems. Would you bet on it?"

 

New article Machine learning in Grid and Martingale trading systems. Would you bet on it? has been published:

This article describes the machine learning technique applied to grid and martingale trading. Surprisingly, this approach has little to no coverage in the global network. After reading the article, you will be able to create your own trading bots.

Testing should be performed on the timeframe on which the bot was trained. In this case it is H1. It can be tested using open prices, since the bot has an explicit control of bar opening. However, since a grid is used, M1 OHLC can be selected for greater accuracy.

This particular bot was trained in the following period:

START_DATE = datetime(2020, 5, 1)
TSTART_DATE = datetime(2019, 1, 1)
FULL_DATE = datetime(2018, 1, 1)
END_DATE = datetime(2022, 1, 1)

  • The interval from the fifth month of 2020 to the present day is a training period, which is divided 50/50 into training and validation subsamples. 
  • From the 1st month of 2019, the model was evaluated according to R^2 and the best one was chosen.
  • From the 1st month of 2018, the model was tested in a custom tester.
  • Synthetic data was used for training (generated by the Gaussian mixture model)
  • The CatBoost model has a strong regularization which helps to avoid overfitting on the training sample.

All these factors indicate (which is also confirmed by the custom tester) that we have found a certain pattern in the interval from 2018 to the present day.

Lest us view how it looks like in the MetaTrader 5 Strategy Tester.


With the exception that we can now see equity drawdowns, the balance chart looks the same as in my custom tester. It is good news. Let us make sure that the bot is trading exactly the grid and nothing else.


Author: Maxim Dmitrievsky

 
found a definite pattern over the interval from 2018 to now

Overshooting.

 
  1. Huge respect to the Author! In the Russian-speaking segment, he is a real pioneer in MO publications in trading.
  2. If the author's enthusiasm does not run out, please continue to share the material he has accumulated.
 
fxsaber:
  1. Huge respect to the Author! In the Russian-speaking segment, he is a real pioneer in MO publications in trading.
  2. If the author's enthusiasm does not run out, please continue to share the material he has accumulated.

Thanks. List of topics I would like to check out:

  • synthetic time series and training on them (in detail)
  • portfolios/pair trading
  • reverse engineering of signals/bots
With martingale everything is clear. The same +- same patterns are found, no new dependencies are found
 
secret:

Over-sitting.

A transplant of what? It's a martingale.

without it, it's the exact same correlation from 2017. It's a head-to-head without filtering

I thought you'd find other patterns through a grid. Most likely any grid is just a cover for what would work without it

 
Unfortunately, the data for training and testing are once again slipped backwards , which significantly devalues the result. The author's persistence in unwillingness, if not to adhere completely, at least to compare his approach with canonical schemes is disappointing.
 
Stanislav Korotky:
Unfortunately, the data for training and testing are once again slipped backwards , which significantly devalues the result. The author's persistence in unwillingness, if not to adhere completely, at least to compare his approach with canonical schemes is disappointing.

Unfortunately, you persistently do not want to read what I wrote last time and gave examples

When the computer will be free I will do the opposite. But it has already been done by me and others in the 2nd article.

1, 2
Обсуждение статьи "Продвинутый ресемплинг и выбор CatBoost моделей брутфорс методом"
Обсуждение статьи "Продвинутый ресемплинг и выбор CatBoost моделей брутфорс методом"
  • 2020.12.01
  • www.mql5.com
Опубликована статья Продвинутый ресемплинг и выбор CatBoost моделей брутфорс методом : Автор: Maxim Dmitrievsky...
 

Some brokers (I will not point fingers) have a grail history of quotes in the past, on which it is pointless to train, unless you cut out rollovers and other artefacts, because it will be caught not a pattern, but nonsense

But it will work in the opposite direction, because the training was not done on grail splices

If you don't check the history with your eyes, you can torture the MO for a long time without getting anywhere.

 
Maxim Dmitrievsky:

Thank you. List of topics I'd like to check out:

  • synthetic time series and learning from them (in detail)
  • portfolios/pair trading
  • reverse engineering of signals/bots
everything is clear with martingale. The same +- same patterns are found, no new dependencies are found

Yes, pair trading is an extremely interesting topic.

 
Maxim Dmitrievsky:

Unfortunately, you stubbornly don't want to read what I wrote last time and gave examples.

I will do the opposite when my computer is free. But already done by me and other people in the 2nd article.

1, 2

What is the point of writing important things in the discussion sheets rather than including them in the articles themselves?

 
Stanislav Korotky:

What's the point of finishing important things in discussion sheets rather than including them in the articles themselves?

I have written several times why the article approach is preferable, at least for me.

In a global sense there is no difference. Fighting superstition is not really something I want to do.