Machine learning in trading: theory, models, practice and algo-trading - page 1126
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
So it remains to confess that ALL of us are here solely to f...tease each other yes to hilarious and why so I mentioned in my message .... :-)
In general I think to retire and take a student, to justify, so to speak, his honorary title "The Teacher" !!!!!!!
Let it be known to you teaching as well as healing is a good thing.
I almost got my ass kicked a couple of times for that. From friends, of course. They couldn't take it. Well kuli tortured mind is erudite, I do not know why, but I'm interested in life a lot, but in the field of MO I have surpassed myself!!!!!!
It is my top :-))) Or rather in it I have the most experience. I've looked at job openings for MO specialists, it's just about time for engineers. You don't need to create it as a rule. It is necessary to train it properly, and this requires knowledge of programming, etc... and the salary they have a lot more than mine, which surprised me......... that even I'm beginning to look at it with the thoughts yes to fuck these MOSCOW NEVER SLIPPP OOUUUUUUUUUUUOUUUUU!!!!!!
Now let's take a look atMihail Marchukajtes' version of the systemand see if it's that bad. For me, 50 trades for training is also not enough.
But it can be approached from the other side. I'll describe my experience in trading on a new instrument (sorry, on my own, because I have no other experience).
I play, say, Sberbank futures, and it occurs to me to switch to the RTS index.
So I did.
First day. Just stupidly looking at quotations, sometimes writing something down on a piece of paper, not doing anything.
Day 2. Rare virtual transactions. Opening and closing are written down on a piece of paper.
Day 3. Frequent virtual trades on paper.
The 4th day. That's it, let's go to the real market for small transactions in the 1st futures. Training is over).
Everything. It took 4 days for training and only 15-25 deals.
It turns out that even on 15-25 deals it is quite realistic to train. If we take into account that the NS does not stare into the monitor all the time and learns only from deals, then 50 deals appears to be quite a teachable sequence.
The conclusions are preliminary. I have doubts myself.
So it remains to confess that ALL of us are here solely to f...tease each other yes to hilarious and why so I mentioned in my message .... :-)
In general I think to retire and take a student, to justify, so to speak, his honorary title "The Teacher" !!!!!!!
Let it be known to you teaching as well as healing is a good thing.
I almost got my ass kicked a couple of times for that. From friends, of course. They couldn't take it. Well kuli tortured mind is erudite, I do not know why, but I'm interested in life a lot, but in the field of MO I have surpassed myself!!!!!!
It is my top :-))) Or rather in it I have the most experience. I've looked at job openings for MO specialists, it's just about time for engineers. You don't need to create it as a rule. Necessary to properly train it, and this requires knowledge of programming, etc... and wage-scheme they have not chet mine, which surprised me ......... cheto even begin to look at me with the thoughts yes yes to fuck these MOSCOW NEVER SLIPPP OOUUUUUUUOOOOOOOOOOOOOOOOOO!!!!!!
he's writing from an insane asylum...
If I understood the teacher correctly, it's like this
In general, the code from the idea is somehow transferred without hyphenation ((( I do not know why
SZY Misha you need to see this: dark-os.com/viewtopic.php?t=308103
And this is on the case. Thanks buddy, be sure to look, but somewhere in the week ....
Actually thought about metrics today and realized that they can be riveted an infinite number. Take a set of parameters from MT tester and start building relationships between them, eventually getting an overall score that includes this entire set. And voila.....
I have a moderate HFT, 30-200 trades a day, and with non-HFT, IMHO algo-trading is not compatible, firstly you can not check whether the strategy works, statistically reliable, and secondly the whole reason that the market in general can be predicted by price data and the like, The main of which is diffusion of information, disappears, that is, the price(s) do not matter at all, the market is effective, only macroeconomic forecasting is somewhat different, it's the highest level, the level of Soros and Buffett, you need to guess what the politicians and insiders will do.
Regarding the changes in the markets and the value of the data for 3 years or more, here, as usual all depends on experience, I tried - learned, I get so that even 5 years sometimes yuzayu, I do not theorize about it, I know people who and 10 years of data for HFT yuzayut and this is not a novice, to put it mildly.I'm not a theorist, either. My own experience shows that the lifetime of the system is about 2 years. And hand trading shows that you can't play now the way you did even 3 years ago. When trading hands there is a constant adaptation and there changes are not immediately noticeable.
I've tried resurrecting and adapting old systems to new data - they don't go either way. As for moderate HFT, you can compare the noise of the same RTS index now and 3 years ago. It's substantially different noise, and a completely different game.
And, in principle, the concept is the same - forecasting is only realistic on short intervals.
from an insane asylum, I guess...
I'm not... I'm sitting at home,you cardboard idiot....
That's right, 2 years of strategy with one configuration is a wow! Although I'm not thinking in terms of individual "strategies" for a long time, there is a large-scale forecasting system with hundreds of modules that digest thousands of scalar price and other time series and produce hundreds of vector series of forecasts for traded instruments, for different parameters (sign, ox, etc.).There is also a system of execution modules that quasi-optimally implements those forecasts into orders flow into the market, of course risk management modules and different "supervisor" modules controlling the quality of other modules, deflections and errors, signaling or stopping something abnormal. So, some "strategies" appear and disappear by themselves, while some modules are trained real-time, that is, in a sense, "strategies" are updated every second))) but of course this is a far-fetched comparison.
Otherwise it will be like LTCM. They say that they also configured the system based on several years of history, their AI didn't learn how to behave during a financial apocalypse.
My setup is much simpler. What you write about is another level, and, understandably, other concepts and design and testing. And these concepts cannot be transferred to systems with soldered-in logic, where even adaptation is done only within this logic and is controlled by 2 - 3 parameters.
Imho, the system you describe is not available to a single person. And I think even firms that have implemented such an approach are few and far between. One I know, and these guys have seen and heard. It is very similar to what you are writing about. It is not a mass product.
toxic:
That's right, a 2 year strategy with one configuration is wow! Although I'm not reasoning in terms of single "strategies" for a long time, there is a large-scale forecasting system where there are hundreds of modules that digest thousands of scalar time series of prices and other things and produce hundreds of vector series of forecasts for traded instruments, for different parameters (sign, ox, etc.) and time horizons.There is also a system of execution modules that quasi-optimally implements those forecasts into orders flow into the market, of course risk management modules and different "supervisor" modules controlling the quality of other modules, deflections and errors, signaling or stopping something abnormal. So, some "strategies" appear and disappear on their own, and training is real-time for some modules , that is, in a sense, "strategies" are updated every second))) but of course this is a far-fetched comparison.
If training is performed every second, then the corresponding increment in the length of the input will be the same.
And if you train on 864 thousand rows and feed them one by one, the increment of each of them will be exactly 10 days (60*60*24*10 = 864000), that is exactly the length for which you have long criticized me here.
Just don't tell me that unchanged, two-year parts of the series are re-fed to the input with each cycle of training, because that would simply be evidence that the models are inefficient.
PS:
Either admit the falsity, or come up with some other justification, such as the fact that many millions of series are used for training, obtained by transforming their timelines.
Here already exhibited, in your support, an example of profitable backwards and forwards tests on random quotes, now we can already move to the distortion of space and time:)
Look atthe course on youtube, where everything is explained in detail, even in excess:
The market is not the only one, but it is the only one, where it is possible to understand the market, and it is easy to guess how to apply it to the market. It's better not to read articles about the market, for obvious reasons, there is a lot of misinformation and noise. I can read books on ML and algotorgovale, which are very useful from my point of view.
I wrote to the L.S.
Throw me something to read too.
Van, not random quotes, but random. How random it is is another question, but it doesn't change the point. The point is different and not many people
understand it. Toxic no one defends, he's not small, and half of what he uses, I would abandon without loss of quality. About
"But if it's so, we need a tutor, not a teacher.)