Evgeniy Scherbina

Folks, I think I have made the breakthrough! I have been fighting for quite a while with the "actor-critic" technique in Reinforcement Learning. It did not work well until I noticed another erroneous code example on GitHub. I always used a single trade as a game while the erroneous code used a number of days with different trades as a game. That's when I realized what was wrong with my code! The erroneous code did not bother about the validation and only performed one pass, which was ok because it was erroneous. But the idea was good. Suddenly, I started seeing positive dynamics in my tests.

Below are 2 pictures - one is training and the other is validation with a small part of test data (the profile does not allow 2 pictures so it is only the training, check the Comments section of "Fast Neurons" for the second picture). It is good, I did it! I am going to put it on the signal starting next week. If it shows no bugs I will publish it as an update to the strategy of "Fast Neurons". It is going to be the first strategy on the mql5 market, which uses the "actor-critic" technique in Reinforcement Learning (specifically "Advantage Actor Critic" or A2C).

What is the best option among the many popular options in Reinforcement Learning? It is clearly "Actor Critic". They post a lot of code to demonstrate they understand what it is about on GitHub. It is usually a very complicated and ineffecient code or it is coded to play the famous Atari games. Most games are won by playing the longer the better. It is not the case for Forex trading. You do not win by holding an open position too long. I have found that the technique "Advantage Actor Critic" or A2C is very sensitive to rewards.

And then, a complicated technique should at least outperform simpler techniques, otherwise it is useless. Alright, I will try to understand why they insist on using the log function for probabilities just to make sure it does not overcomplicate things. If I get through, this will be my most advanced and sophisticated system for Forex trading yet.

Hi, everyone!! Another year and I am still fighting for the best neural strategy ever. I have added an "ante" one-year test (a test which precedes the training period) and 2 validations, one to fit the training curve and the other to check the final result. Should I call it a triple validation?! I have also started using 3 charts - monthly, weekly, and daily.

All of this is being added to my strategies right now. "New York" is currently the most advanced strategy which balances wins and loses. "Multiq" combines data from multiple symbols to train a single neural network. And the ever-lasting "Excelsior" uses 2 competing neural networks. Check them out!

I am also working over a strategy for intraday trading, which will include 4 charts - monthly, weekly, daily, and 6-hourly! I am thinking to call it "Fast Neurons", check it out soon, too!

Hi, everyone. I published recently the newest update for Excelsior. The important difference is that it uses more information (relative price changes) in addition to MACD. It may also close negative trades.

But we are not there yet.

The biggest jump will be using 4 threshold values instead of 2 as it is now.

Right now 2 networks in this strategy use 2 threshold values to open and close trades. For example, any signal over 0.8 is to open a trade, while any signal below the oppositve value of 0.2 is to close that opened trade. There is a very long run between 0.8 and 0.2, and the strategy loses time and does not close when it should. So I thought any signal over 0.8 to open and below 0.6 to close would be better.

It takes time to test this new approach. For the previous approach, it was 4 x 4 = 16 total options. Now it is 4 x 4 x 4 x 4 = 256 possible options to be tested to see if it performs better. I am doing it and when I am done with it we shall triumph!

This month of September I guess.

Hi, everyone. I have made this nice table to summarize similarities and difference of the 3 new strategies I created. While these strategies are already available or will be available soon, I would like to explain how I developed them.

It was "Excelsior" that pushed me into the realm of possibilities offered by recurrent neural networks. As none of the standard approaches produced a robust result, I decided to experiment and combine two competing neural networks to make one trading decision. It did work in some periods, and it did not work so well in other periods.

Then I read "Reinforcement Learning: An Introduction" by Sutton and Barto and, after months of experiments and tests, I came up with a nice strategy called "Pipsovar". The one difference of reinforcement learning from recurrent networks is that it does not know its decisions and outcomes from the start. Instead, it explores the future by making all possible decisions on the go.

And finally, I created "Arsene Lupin", a strategy that brushes off the complexity forced by indicators and uses only relative prices changes. "Arsene Lupin" leverages my newest techniques to train and verify recurrent neural networks.

Hi, everyone.

Excelsior is going to be updated this week. I have found a way to combine data from weekly and daily charts into one multivariate training.

Live trading has shown the drawbacks of separate training on weekly and daily charts. Week trading seems to be lagging too much behind. While day trading does not have enough of past perspective to rely on and to make correct trading decisions. It will now be 30 weekly bars + 30 daily bars combined into one multivariate thread of 30 states to forecast 1 (or 2) week ahead.

The tests I have made so far show a more reliable trading in the future in the past. We shall see if it can outperform itself in the real future!

I have recently started working on a new strategy using a reinforcement learning... For now, it cannot even reproduce its historical trading, but I am determined to find a good solution one way or another...

Hi, everyone.

I have recently came accross "The Motley Fool Investing Philosophy":

1) Buy 25+ Companies Over Time
2) Hold Stocks for 5+ Years
3) Add New Savings Regularly
4) Hold Through Market Volatility
5) Let Winners Run
6) Target Long-Term Returns

I guess all of it is applicable to Forex trading and, in particular, to my newest advisor "Excelsior". It uses 8 symbols all traded from one chart. It trades in both weekly and daily charts. It can compensate losing trades if market volatility rises. And it uses a dynamic trail to get the bigget profit possible from trending winning trades. I am sure the advisor "Excelsior" can get the job done!

And may this clear-cut and well-designed approach from the Motley Fool help us get the profit in the coming year 2022!!

Добавил тему Ресурс типа double перестал импортироваться

Hi, everyone.

The next update of Excelsior is for the next week, with ever more options in the training to make it an ever more stable and profitable system. Nothing guarantees a profit but I am sure I have made yet another step towards defeating the market noise.

I will publish a signal at that point, which has been running for about 2 months and has made over 50% of profit.

And the price for Excelsior will double up.

Why? Why are you not selling yet? Is it because you are not using Excelsior?

Awaiting the Fed decision with all those sells...

Hi, everyone.

If you thought I was done with ANT after a few negative reviews, you were wrong! The truth is TensorFlow and Python have tremendous potential and provide quite a bunch of possibilities to train Forex strategies. As I said earlier, I found a way to speed up the training process 4000 times. This was not talking. There are many articles out there that explore how to optimize iterations by using NumPy arrays with floats and indices instead of clumsy Pandas arrays. And then I finally understood how to appease the retracing error in TensorFlow. So this is real!

With a much faster training, I can clinch more options into the training to see if it can outperform its previous results. And finally, here it comes - my newest technical test with the future in the past! It shows what works and what doesn't and why. I am currently testing a linear activation against a sigmoid activation for ANT. I see that the "cc" trades are unstable while the "c" and "r-in" trades can be way more profitable. The new version of ANT has been delayed somewhat but it is coming this month and it is sure to have gorgeous improvements! So stay tuned!

Hi, everyone.

I have done a few more tests with the newest advisor Excelsior. And I think I have come to an almost perfect configuration of options.

Below is a picture of 2 trainings: one with the current version 2.1 and the other with the next version 2.2. Both charts contain the top 20 rows with profits for training, validation, and test periods. Over a 150 thousand rows for each training were sorted by the validation profit.

What we can clearly see, is that the current version 2.1 produces a few configurations that secured a modest profit in the test period (passes 6, 12, 16, 17, 18, and 19). But the next version 2.2 has a much better performance in the unknown test period. While it is still a history test and nothing guarantees a profit on the volatile Forex market, it is a very clear indication that this system of 2 competing networks, that I am tuning right now, is absolutely wonderful.

Hi, everyone. So far, I have created these 3 strategies which utilize the power of neural networks.


1. Ain't No Trend (or "ANT" for a short name) was my first attempt to understand how neural networks may help evaluate the market noise. It was initially developed completely in MQL, and then I transported the training part to Python. All the various configurations of fully-connected neural networks have enabled this strategy to produce a ridiculously low number of trades. It should be considered as the most conservative and least profitable.

Number of symbols: 6
Timeframes: H4
Trading strategy: MACD indicator
Training strategy: fully-connected neural network
Number of trades in real-time: less than 10 per month

A major upgrade is due in August.


2. Neural Bar Predictor (or "NBP" for a short name) was made with a more sophisticated approach and super-tricky-complicated recurrent neural networks. I spent 2 hours figuring out a way to transport the weights from a TensorFlow fully-connected neural network of ANT, and it took me 2 weeks to finally do it the right way for the recurrent neural network of NBP. It was various configurations of training, too, with very many options still to be explored. The basic idea for NBP was to make it generate by far much more trades than ANT did. I got to it by packing 5 timeframes and different types of signals together. The next big step for this strategy will be to introduce more fine-tuning and exclude tons of unrealistic trading results in order to achieve a more stable version and keep a decent number of trades.

Number of symbols: 6
Timeframes: D1, H12, H8, H6, H4
Trading strategy: ADX indicator
Training strategy: recurrent neural network
Number of trades in real-time: over 40 per month

A major upgrade is due in August.


3. Excelsior (no short name this time) is currently my best and most advanced effort at harnessing a popular variant of the recurrent neural networks and making it forecast the most reliable daily charts of 8 symbols. With this approach, I understood that recurrent networks could forecast any time series. To make it a stable system, it should be 2 independent competing networks. The trading strategy relies simply on daily bars so all trading decisions are made by a combination of the recurrent neural networks. Also, this enables the advisor to reconsider its trading decisions on every new day and even during a day! To overcome all the new obstacles of this approach, I had to find a way to speed up training and validation calculations about 4000 times.

Number of symbols: 8
Timeframes: D1
Trading strategy: daily bars
Training strategy: 2 competing recurrent neural networks
Number of trades in real-time: stat expected soon

The latest version 2.0 has recently been published.