Neural Network self-optimizer (best Expert advisor)

The solution for forex: a self-optimizing neural network, which has been trained tick by tick with variable spread for at least 10 years, to that network add a news filter, so that after being broadcast, the network is self-optimizing , and then the regular self-optimization coefficient is recomposed, with a stop loss equal to or less than the take profit, and a system with a threshold that reinforces the stop loss and take profit

More or less, something like this, tick by tick with variable spread from 2003 to date, starts with 250 dollars, and ends with 25,000,000, very good.

Have you ever forward tested this EA on real market?

Ah, you're about 10 years out of date. When neural networks were first becoming more mainstream, they were applied to forex/stocks very quickly and found to be lacking. The issue is that you don't have enough of a data "story" to train the models well.

Sure, you can go as granular as tick based, but the movements are so miniscule and volatile that it doesn't get a cohesive strategy. If you think about how the neural network will work in terms of minimising loss (assuming you're using either a recurrent neural net), it will essentially be averaging these moves anyway as a lot will counter each other. So then going that low makes the move less worthy. For trying to predict actual moves, the window you'll need to monitor will be huge, which then means you may as well have been doing higher granularity candles.

To get genuine recognisable moves that would make reliable training, you probably want to be on 1H data or higher and there just aren't enough data points to establish patterns, especially considering that the market moves different in different years (if you trained a model during COVID, or the recession, or the gulf war, or brexit, then how will it perform outside of these?).

Now, I'm not saying neural networks won't have a potential of yielding positive results (the amount of quants being recruited to banks say otherwise), but if you think it's as easy as feeding raw price data and a simple news filter in, then unfortunately you're sadly mistaken.

The network was trained between 2010 and 2020, then a backward test was carried out, and it was found that the 2008 crisis passed without problems, which was one of the most terrible, and the future tests from the beginning of 2020 to 2021, continued being profitable, without falling into the curve, self-optimization, notably improves this, and works well in both 30 minutes and 1 hour. Thank you very much for your concepts, which of course are very valid.
Now tell me, which expert advisor works well for your critic, from 2003 to date, starting from a small account. It would be interesting to see an expert advisor work well, tick by tick, without destroying the account, and of course in 20 years obtain an important profitability.

I'm an algo programmer, having coded them for myself, discord groups, and on the market place, and can manipulate the majority of my library to show such results with the right parameter sets; I also have a particular interest in AI, however I concentrate on collecting more abnormal datasets to prevent myself recreating what the masses have already done. That above graph is easy to achieve given lot scaling, and picking a backtest which produces a good forward test. Due to its exponential nature, it doesn't even need to start performing well until around 2018 on that graph.

It's also got all the trademarks of an overfitted algo... 60% DD to start with a relative DD of 97%, but also your biggest loss was $500k when the peak was $25mil.. Even assuming that the largest losing trade happened at the absolute peak of account value, that's a 2% loss on a $25mil account which is a bit risky to say the least. But this then doesn't make sense considering the large DD happened at around $10mil and lost like $6mil; you'd have had to hit such a poor stroke of luck to lose 60% of your account assuming an absolute peak loss of 2%... Looking further, and utilising past experience, I can see that your lot size peaked in around the last 10% of the graph where the large step-up happened, (see how we start exponential and then in the last 2 grid horizontals, it stops and goes linear?) meaning you were able to hit such losses at a much smaller equity. This is very reminiscent of an overfitted algo on super aggressive settings. What is your risk % per trade on this bot?

As an example of overfitting on a more respectable scale, I've attached a screenshot of a random BB algo I have loaded up as we speak showing $1000 - $35,000 in a year, using static lot sizing of 0.1 lot.. If I bothered coding scaling in, that would outperform your equity curve in the same timeframe. Worth noting, it's a static trend as long as I run it, and doesn't hit anywhere near 60% DD... unfortunately I'm on my macbook which doesn't have decent histories downloaded due to its HDD size.



Not going to get in to an e-peen conversation, however my parting wisdom is this: What you are showing above can be recreated using widely available python libraries in less than 100 lines of code (tensorflow/sk-learn for the NN, NLTK for sentiment analysis of news, GoogleNews to get the news and OandaPyV20 to trade). In MQL there's a basic NN tutorial here: https://www.mql5.com/en/articles/497 from 8 years ago. There are also thousands of medium articles and youtube videos on creating forex/stock NN models... If it's that easy, why have none of these spotted the 100% APR potential and cashed in? I get that in manual trading, the simple methods often can be profitable due to people's intuition/luck/discretion playing a major part. In algo trading it just isn't that simple, otherwise that MA crossover bot that every youtube video tells you to code would be gifting everybody a Lambo!

Edit: Sorry, I'm also curious. If you trained on data from 2010-2020, which is a statically defined range, how can it then be self optimising? How would you reoptimise a model thats already been trained on over 60% of the dataset that you're then running it on...That's two conflicting model types? Either it's dynamically modelling itself from a continuous stream of data and adjusting its weights and biases on the fly, or it's trained from a dataset and then has fixed parameters... The only alternative I can assume is that you've essentially trained it twice on pretty much the same dataset, proving overfitting, as you're then "optimising" it on a dataset it's already seen and attempted to get perfect a perfect model on already...

Edit 2: Apologies again, just seen your post history and can clearly see that you're not a programmer. Unfortunately you've stumbled in to the naive trap with the above algo. I wish you the best of luck in your quest of finding the holy grail being handed out in forums!

First of all sorry, because of my English, it's not good, I only speak Spanish, and I use the translator: The neural network was trained for 10 years, without auto-optimizing, and logically, without time filters, when the test is done backwards. and forward, without auto-optimizing as the graph shows, it is absolutely profitable, but then when the auto-optimization option is applied, the graph improves, and then the time filter option is applied, with modification of the auto-optimization time of the network, and get even better. I'm not saying this is the holy grail. But I believe that using recurrent and self-optimizing neural networks, they generate a high probability of success, in the operations, regarding the lots to be put into play in each operation, if it is profitable, and the capital justifies it in the following option the lot to put into play, is the same or is increased, but if it is not profitable, in the next option the lot is reduced. Generate only one single operation at a time. Look for it on the internet as the title indicates, and try it, it is the best way to see if it works or is a fraud.