Machine learning in trading: theory, models, practice and algo-trading - page 1122
![MQL5 - Language of trade strategies built-in the MetaTrader 5 client terminal](https://c.mql5.com/i/registerlandings/logo-2.png)
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
it is so )
I, not to lie, for all these years that I've registered for the forum, have tested about a thousand TS, have written and modified about 40-50 EAs since this year, all that I have long understood:
- beautiful flat chart in the tester means only excessive risks (usually TS without stoplosses or stoplosses purely for ... who knows why the SL is 3 times the TP), ie over sitting losses
- a nice flat chart ... means averaging or martingale with oversleeping
.... In the theory and practice branch there are the sets that have already appeared, but alas, without stop-losses but nice sets )))
And in addition... i.e. to enter safely with minimal risk the MO should be on the 2nd signal to enter the market - I'm writing about not large TF (M1-30, on higher TF, there is nothing, there contrendovyh TS is more effective, and for higher TF position holding time in the market is completely different)
... And if the MO uses all signals to enter the market, some of the signals will not be impulsive, but the studs with a lack of liquidity, respectively, increases the randomness of entries ... If you do not know what kind of market events there are in the Forex market, you cannot be sure that the market will go in one direction in 2-5 minutes and move as much as it did during half a day.
I used to earn good money on martech (about 1500% for 2 months, with high risk). But it was possible to make several attempts and still earn. It was during the crisis, when the Eurobucks was constantly falling, and only the lazy could not sell.
I liked arbitrage as well, but I got bored.
I have not yet mastered it to the end, but the potential is evident in terms of risk control.
Probably Mr.fxsaber just mixed up the terms, but "out of sample" is a test, that is forward, it MUST be obtained from the data BEFORE, otherwise you get absurd, the future predicts the past, it is very easy to do, but it makes no sense.
Why, the data is independent. I do tests that way sometimes too, I don't see the difference
You don't say. It's easier to take the trend in installments than all at once. And there's no need to watch).
Also an option (unnecessary care about the broker, and he has long worried about himself). But if you can do something worthwhile, he will (most likely) be grateful to you repeatedly.
I have not yet fully mastered it, but I can see the potential in terms of risk control
You have not mastered it? Don't be ridiculous, one third of those present here do not have even one third of your vast knowledge, just ... like me - they have picked up the basics for years and now they are too smart)))
I would like to achieve adaptivity in IR, either for different financial instruments or for different predictors, but ... I've got a lot to read, the IR study material is really huge, I'd like to deal with the basics for now
I want to achieve adaptability in MO, either for different financial instruments or for different predictors,
I don't think so. The basic methods of MO are similar to conventional logic - if... then..., a little more complicated. And not one step further.
Adaptive ones are definitely possible, but they are much more complex structures and I don't think that they are realistic under our conditions.
You haven't mastered it? Don't be ridiculous, one third of those present here don't have even one third of your knowledge, just ... like me - they have picked up the basics over the years and are very smart ))))
I would like to achieve adaptivity in IR, either for different financial instruments or for different predictors, but ... I still have a lot to read, the IR study material is really huge, I just need to master the basics
I'm hooked on books, otherwise some nonsense will come out of my mind :)
my mathematical background is weak, but i have an understanding of what comes from where and where it goes
I'm making an adaptive one myself, so that it eats everything in a row. A lib for any strategy, connects in 1 lineOh no... there is a difference, a very big difference. Forward ONLY after backward and of course they should not overlap.
Besides the peeking fics, the time mixing of samples from past and future is also one of the reasons why ML algorithms are unrealistically high in training, you can't randomly mix them.
Honestly, I thought it was a misprint, otherwise it's very crude mistake, or another "hyver-staker" style fake, like he likes to do to make the "meat" think))))
And still, what difference does it make that the forward is behind threin? If the picks don't overlap. They're just sets of points, not connected in any way (well maybe connected to a little depth because of the lag of the predictors)
you can ask him :)@fxsaber or how to make an invitation hereadaptive myself just doing what would chew everything. The lib for any strategy, connects in 1 line
If I've read enough books I'll start my experiments with 2 charts for 2 symbols. The same guy wrote that if my TS doesn't work with some symbols then don't bother with machine and look for other symbols or use synthetics. I checked some TS in the strategy tester, in most cases it is so - for some currencies there is no problem with optimization, but there is no positive expectation on the history of any currency I want to optimize. So I decided "it's not for nothing" (C), it means it is possible to find these differences with the help of MO
No one and nothing forbids you to work with minutes/seconds and write samples every minute/second even for long term trading, scale filter parameters accordingly and you will have hundreds of thousands of samples.
Although if you are not experimenting on your own, but are getting "inspiration" from "articles" of biased upstarts around the market, all these tips are irrelevant, there are hundreds, maybe even thousands of them, you need to be able to invent and check them yourselves, if not there is no point in it.
Mr. The Wizard said something similar to that to someone else.
Hear me before it's too late.
Well, let's talk..... It's highlighted in yellow. I do that, only I do it myself. The difference in the size of the training sample is directly related to the capacity of the neural network. All of you who are training on 1000 examples or more have gone the easiest way. A kind of shot from a tank at a sparrow. That is, you take the capacity of the network, the number of neurons, layers, etc. Cramming even what it doesn't need into the net in the hope that it will sort itself out and thanks to its capacity it gives out a good model. But this is nothing to say about the quality of training. For my part, I try to train just one neuron, but so well that it is enough. In my case, it's a sniper rifle shot. Then a reasonable question arises:
Can you reduce your AI to a single neuron and still run my data? Of course I respect Sanych, but the statistical analysis I could do myself in Rattle, and the networks there to twist... Not interesting :-(
I have not waited for an answer so far, and frankly, I don't expect it anymore.......
And on java I would not mind help, to start optimizing by my super-duper metric I have only to transfer an array from one class to another and the bomb will turn out. But I don't think I'll be able to do it in a year or so ...... :-(
P.S. The argument between us will go on forever, because these are two approaches and both have the right to be, as the yin and yang, black and white, wet and dry, AI developers and training engineer. Yes yes these are two completely different specializations in MO and if you think that you successfully combine two professions at once you are deeply mistaken.... Although it's not uncommon either...