Machine learning in trading: theory, models, practice and algo-trading - page 2737
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
As an idea-generator: maybe not a quote or bar colour should be determined ? but for example something less noisy and half known, and not a single event.
For example the position of LWMA 20 in 7 bars ?? simple mathematics can be used to outline approximately where it will be, and if ML/NN methods manage to narrow this area, then it's just a matter of coming up with a trading algorithm that will make a profit.
There is such a thing that we need to take into account the vector of price movement in the future, and we are trying to predict it, albeit with different discreteness. But, I am interested in solving an even simpler problem - determining the price of the muving if the price will move towards its crossing, ideally even calculate the crossing point.
Unfortunately, I don't know how to make regression models or multiclassification models in MQL, and without this there is no incentive to start solving the problem in any packages of other languages. Although, I have such a task, so I am ready to solve it together with someone with interest, including calculation.
Yeah I don't understand what you want to do and how you want to do it either, how to define what
I'm of the opinion that it's all useless, and the bigger the sample the better, but I'm ready to test it, and for this purpose we need a suitable tool. A tool that will determine the optimal areas for training a model, let it be at first on history, and then we will see if we can do it without looking into the future.
The thing is that we need to take into account the vector of price movement in the future, and we are trying to predict it, albeit with different discreteness. But, I am interested in solving an even simpler problem - determining the price of the muving, if the price will move towards its crossing, ideally even calculate the crossing point.
Unfortunately, I don't know how to make regression models or multiclassification models in MQL, and without it there is no incentive to start solving the problem in any packages of other languages. Although, I have such a task, so I am ready to solve it together with someone with interest, including calculation.
If the price will move towards SMA and we need crossovers, the problem is exactly the same (perhaps even more complicated because of the need for crossovers, despite the known direction in advance).
Knowing that dSMA slope=(price[N]-price[0])/N and statistical characteristics of the quote (how many ticks there are, how many ticks are translated into points for a given time) you can build probability fields "in the future" - here is the price field, here is the sma field, here_there_price_will_meet_SMA.
But it will be a statistical-analytical solution and you can't get money from it. Then you have to narrow these fields in the same way via ML/NN and then come up with an alg. how to get money out of it :-)
If the price will move towards SMA and we need crossovers, then the task is exactly the same (perhaps even more difficult because of the need for crossovers, despite the known direction).
knowing that the slope of dSMA=(price[N]-price[0])/N and statistical characteristics of the quote (how many ticks there are, how many ticks are translated into points for a given time), we can build probability fields "in the future" - here is the price field, here is the sma field, here_there_price_will_meet_SMA.
But it will be a statistical-analytical solution and you can't get money from it. Then you have to narrow these fields in the same way via ML/NN and then come up with an alg. how to get money out of it :-)
I know how to get money - trading on the channel waiting for correction. And the rest - yes, that's what multiclassification or regression is for. In fact, you need to build a model of probable bars, and then calculate the model.
Well, if you analyse not only time series, then sure. But not only time series can be analysed not only by everyone, so I settled on the obvious representation of the market in the form of a price series as an object of research by methods of MO.
Discretisation is necessary for analysis, without it there is no way. Usually, a time series is a discretisation at equal time intervals. But you can do it in another way - renko or zigzags, for example.
As I see it, the source of discretisation is a dumb basic TS, which we then try to improve by means of filters from our models. For example, the initial TS "buy at the beginning of the hour and sell at the end" or "open at the formation of the knee of a zigzag in its direction and close at the formation of the next one", and the final TS on the basis of some indicators-predictors rejects some inputs.
You consider fiddling with different ways of discretisation as an empty dabble and there is some reason in it, but there will always be those who disagree with you. This is another reason why constructive discussion of any specifics is impossible in this thread.
Discretisation is necessary for analysis, without it there is no way. Usually, a time series is a discretisation at regular intervals. But you can do it in another way - renko or zigzags, for example.
As I see it, the source of discretisation is a dumb basic TS, which we then try to improve by means of filters from our models. For example, the initial TS "buy at the beginning of the hour and sell at the end" or "open at the formation of the knee of a zigzag in its direction and close at the formation of the next one", and the final TS on the basis of some indicators-predictors rejects some inputs.
You consider fiddling with different ways of discretisation as an empty dabble and there is some reason in it, but there will always be those who disagree with you. This is one more reason why constructive discussion of any specifics is impossible in this thread.
You consider fiddling with different ways of sampling as an empty dabble, and there is some reason in it, but there will always be those who disagree with you. This is one more reason why constructive discussion in the thread of any specifics is impossible.
I am generally of the opinion that all this is useless, and the bigger the sample the better, but I am ready to test it, and for this purpose we need a suitable tool. A tool that will determine the optimal areas for training the model, even if at the beginning on history, and then let's see if we can do it without looking into the future.