Machine learning in trading: theory, models, practice and algo-trading - page 559
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
And here I make a forecasting thingy, then on some history I immediately evaluate its quality... if the quality is normal then I teach the second one to trade according to forecasts, the other one just defines where it is better to buy/sell
(at first it just had fuzzy logic, but then I decided to replace it with the Boolean one)
The same problem - the optimization of the target function.
Perhaps the purpose of both optimization in MT and learning NS is to find some kind of target function. But the resulting functions are very different, including in the physical sense.
In addition to the previous post.
For my NS, I don't need any clouds. Everything is done on my home computer. Yes, it takes a long time, really within 24 hours. By the way, I don't need MQL for those VS, neither its features, nor limitations).
Probably the purpose of optimization in MT and NS training is a search for some target function. But resulting functions are very different, including physical sense.
I need a compromise - I don't want to wait 24 hours for training, and I don't want to retrain by hand... I need fast, reproducible in the tester over a long period, with auto retraining when needed
Considering how non-trivial the task is, there are only fast algorithms and some optimization
I don't want to write all sorts of testers / shapers by myself, use different program combinations, let the developers do it... and I'm just a trader
If there will be an excellent integration with R\Python, which they kind of want to add - then we'll be able to move on...
My timing - if I will not have a robust system on NS by the beginning of spring, then the hell with it, I'm not proud :)
And third, when teaching, don't impose your solution on the MLP, as most local articles do.
Do you mean self-training? How can this be done?
I need a compromise - I don't want to wait 24 hours for training, and I don't want to retrain by hand... I need fast, reproducible in the tester over a long period, with auto retraining when needed
Given how non-trivial the task is, there are only fast algorithms and some optimization
Estimate of time - if until the beginning of spring I won't get a robust system on NS, the hell with it, I'm not proud :)
I've been working with NS for about half a year, just working on the projectile approach. And only then, gradually, moved on to market tasks. And it's better to train 24 hours with more or less known result, than to try different fast algorithms and variants.
By the way, I got an acceptable result at the 4th or 5th attempt. It took me only about 2 months. (Of which only 5 days directly to the training).
Self-learning you mean? How can you do it?
No, not self-learning. The usual BP with intermediate settings between epochs.
Just imagine learning the multiplication table with a teacher who doesn't even really know it himself. That's exactly where I started with this thought).
Came to what Heikin described - the ratio of right/wrong answers in the learning sequence should correspond to reality.
I spent about half a year working with NS just practicing my approach to the projectile. And only then, gradually, I switched to market problems. And it is better to train for 24 hours with more or less known result, than to try different fast algorithms and variants.
By the way, I got an acceptable result at the 4th or 5th attempt. It took me only about 2 months. Of these, only 5 days directly to the training).
also about half a year ... it kind of earns, but should be more, this is the NS! :) 5% a month on average is not enough to be satisfied ... after arbitrage, when hundreds of percent were made
It's also about six months old... it kind of earns, but it should earn more, it's the NS! :) 5% per month on average is not enough to feel satisfied... after arbitrage, when hundreds of percent were made
I gave a chart of the test results somewhere above. There seems to be nothing wrong with it.
There's no problem retraining once every 3 months. Haven't needed it yet.
Previously warned - no specific data from the real world is not published. I don't confirm or deny speculations about it).
Yuriy Asaulenko:
And third, when teaching, do not impose your solution on the MLP, as most local articles do.
...
No, not self-training. The usual BP with intermediate settings between epochs.
Just imagine learning the multiplication table with a teacher who doesn't even really know it himself. That's exactly where I started with this thought).
Came to what Heikin described - the ratio of right/wrong answers in the learning sequence should correspond to reality.
I don't get it...
Well, if there is stilla "learning sequence" with answers, then you are still "imposing your solution on MLP".
I don't get it...
Well, if there is stilla "learning sequence" with answers, then you are still "imposing your solution" on the MLP.