Machine learning in trading: theory, models, practice and algo-trading - page 3739

New comment
[Deleted]  
There is no more professional approach, with validation on history. There is only speculation about it.
[Deleted]  
A useful theoretical extract from the introductory section:

Julia sets as an analogy of "specific market scenarios"

For each particular value of c (i.e., for each point on the "Mandelbrot landscape"), we obtain a unique Julia set. This set describes the dynamics of iterations for a given c: which initial points z0will remain bounded (belong to the filled Julia set) and to which attractors they will converge, and which ones will "run away" to infinity.
Each particular Julia set for a given c (which in turn represents the "current state of the market" or "set of prevailing parameters") can be an analogy of the market behaviour at that particular moment in time or in a given "market regime".
  • If c is inside a Mandelbrot set (e.g. in one of the "onions" of period 3), then the corresponding Julia set will be coherent and have attracting orbits. In the market, this could mean that despite the apparent complexity, there are certain "attracting" levels, prices or cycles to which the market continually returns or around which it fluctuates. These could be strong support/resistance levels, averages to which price tends to return (i.e. reversion to the mean).
  • If c is outside the Mandelbrot set, the corresponding Julia set is incoherent ("Fatou dust") and has no attracting orbits (other than infinity). This can be analogous to highly unstable or "explosive" market conditions, where price has no "anchors", tends to infinity (e.g. hyperbolic growth or collapse), and any small perturbations result in unpredictable and radically different behaviour. Any trading strategies based on "return to the mean" fail under such conditions.

[Deleted]  
Thus, no matter how much the marginalised promptus builds a "hierarchy of structures at different scales", there will always be a more global unaccounted scale factor. And since the promptus denies the importance of statistics, these factors will constantly slap it on the cheek.
 
Dare I ask, has anyone tried to express all these clever thoughts in mql?
[Deleted]  
Inquiring #:
Dare I ask, has anyone tried to express all these clever thoughts in mql?
I did and threw it in this thread in the archive, 50 thousand years ago, try to find it.
 
Inquiring #:
Dare I ask, has anyone tried to express all these clever thoughts in mql?
He's put together a crutch there

. It's kind of like a forward.



Lumpy, but at least it shows attempts to stabilise at a distance.
[Deleted]  
Ivan Butko #:
It's reinforcement learning.
Yes, it's probably a good method. What I was trying to say was a little bit different. Apart from the fact that the market itself can be divided into several sectors, or as Maxim says, modes.

For example, according to the place of the current price in the trend:
- on the edge of the trend,
- trend pullback,
- return from the pullback.

Also the deal itself can be opened and closed in two variants: in the direction of the trend or against it. And it can be closed either in profit or in loss.

Total we have 6 states when opening a deal and 12 states when closing it. Obviously unfavourable states can be reduced, depending on the strategy. And for each of these states there should be its own analyser based on the patterns found earlier. Whether this analyser will be its own brain, rigid branching rules in the code, neural network, or mixed - there are variants to choose from.

That is, conducting a deal is a more complex action than opening and forgetting. And training is correspondingly complex. Simple doesn't work, so let's make it more complicated)

.
 
Aleksei Stepanenko #:
That is, conducting a transaction is a more complex action than opening and forgetting.

The only question is whether it can be formalised in principle in the form of some numerical function. For example, as a dependence of the stop level on the price and/or values of some indicators.

If yes, it can be done within the framework of the MO. If we are talking about an unlimited flight of fancy (an indefinite set of used data and ways of closing trades), then MO is hardly applicable.

[Deleted]  
Of course, each pattern must yield yes/no with probability>0.5. Otherwise, it all makes no sense for long term trading
[Deleted]  
6 entry states and 12 exit states - this is a general scheme that describes all possible variants.
For a specific strategy this number can be greatly reduced, so as not to get bogged down in the curse of dimensionality.
For example, for trading on a trend pullback, one entry state:
- pullback of the current trend, the direction of the trade is according to the trend.
To exit a trade 4 states are suitable:
- movement on the edge of the trend in the direction of the trade, the trade is obviously in the plus,
- return from the trend pullback, which moves in the direction of the trade, we are interested only when the trade is in the plus,
- trend pullback, which moves against the trade, there are two options: the trade in the plus and the trade in the minus, to sometimes successfully jump out of a losing trade with a smaller minus.
Other states are not favourable, because they lead to a losing trade. In these states we do nothing and wait.
If something went wrong, a stop will be triggered.
And so for any strategy. There is a big advantage in this: we do not enter or exit where the strategy does not provide. It is much easier to process 1+4 states instead of 6+12, and it reduces the risk of adjusting to the history.
Like this.
1...3732373337343735373637373738373937403741
New comment