Machine learning in trading: theory, models, practice and algo-trading - page 2363

 
Alexander_K:

And I looked in my wallet - I didn't find any signs of cash either...

It's time to twist the gauges for scrap metal again...

 

Yes, another great discovery - the unsyndicator systems (in the sense that there are no indicators counting on each bar).

By the way, the simplest example is a zigzag, if you understand it not as a broken line, but simply as a list of tops.

 
Aleksey Mavrin:

That sounds reasonable. But whatever traits you take, they will still be the last n values, even if n-> infinity, simply because future values of the traits cannot be taken).

So in principle there's nothing new, just a selection of signs, even if you put Biden's lunar horoscope there.

Let's imagine that there are regularities in the market.


By regularity I mean a complex repetitive structure.

A pattern is composed of a sequence of "events": an event is a rule or a cluster.

An event has three parameters - price, time, and value.

It makes no sense to take into account the sequence of events in indexes , as the market is not stationary, so all that can be considered is the order of events and their parameters

Here is an example

The same regularity on the nonstationary market


So the task is to find this pattern...

I forgot to tell you that there is an event we don't know either, we need to find it

 
Aleksey Nikolayev:

Yes, another great discovery - the non-syndicator systems (in the sense that there are no indicators counting on each bar).

By the way, the simplest example is a zigzag, if you understand it not as a broken line, but simply as a list of vertices.

Find a solution))

 
mytarmailS:

Let's imagine that there is a pattern in the market.

So the task is to find this pattern.

The example is just the same as Alexei pointed out in the post above, while you were writing yours :).

I don't know if you think everyone works by indicators, no. Nothing new in your description, all MO-scholars have long been using this notation and many others and their mixes.

I'm not criticizing, just trying to understand what's new and benefit/ideas for myself.

 

The essence of it is to set TA patterns (one has to start dabbling with something :) ) as primary examples, run them through the history, the most crooked moments to show and 3-5 buttons to classify myself, so it should be fun) I don't think that they will be useful, probably everybody has done it before and I started writing in mql, ~55% maximum.

But the process itself should be addictive, and maybe more ideas will come up.

ap: here's an idea - to sell such toys on the market, like guess what pattern the neural network found here, etc.)
 
Aleksey Mavrin:

The example is exactly the same ZZ, as Alexei pointed out in the post above, while you were writing yours, I see :)

ZZ is one of a million types of events, it's not an example of searching for patterns, what Alexey said I did with ZZ and nets about 4 years ago.

Aleksey Mavrin:

Nothing new in your description, all MO-scholars have been using this notation and many others and their mixtures for a long time.

All MO-ers use a feature matrix "X" and a response "Y".

The feature matrix is a sliding window with a fixed size.

How's that? Do we find a lot of patterns? ))

Not a single MO-scholar has found one yet, and will never find one, even if you train GPT-6.


Also I haven't found in the whole web algorithm which can find such things out of the box

 
Aleksey Mavrin:

I'm not criticizing, just trying to understand what's new and useful/ideas to extract for myself.

The new thing is that there is no such algorithm (at all), but only he is able to find something on the market...

And I see two realizations, I am interested in ideas from experts))

 
Aleksey Mavrin:

The idea is to set TA patterns (one has to start dabbling with something :) ) as primary examples, run them through the history, the most awkward moments to show and 3-5 buttons to classify myself, and that should be fun to run, I don't think the benefit will appear at once, probably everyone has already run the patterns and I started writing this in mql, ~55% maximum.

It would be useful to make a calculation of the significance of the difference between the result and the SB. You could choose some parameter - for example, the total number of patterns found per segment and see how atypical it is for the saturation point. It is hardly possible to calculate the kind of parameter distribution for the SB analytically, so it is worth applying Monte Carlo. From the price many realizations of SB through random mixing of increments are constructed and for each realization the values of the parameter are counted. We get a large sample, in relation to which we look at the value of the parameter obtained for the initial price. If it is strongly shifted to any end of the sample, this pattern deserves a more detailed study.

 

I didn't even bother to reread what was discussed! Fellows, let's define the terms within this thread, otherwise we'll all be wandering around in our own fog!

1. Normalization of inputs? What is it, I have it simple - I managed to drive through the math in the limits (0,1) ,(-1,+1)

2. Factorization of outputs? This is quite difficult to do((((((((

What should I do? the syss tem has to trade!

it means either:

1. Predicts the PRICE - which, imho, is the most difficult!

2. Predicts EVENT - (-1)(0)(+1) - which is also questionable))))

3. Predicts EVENT - no case of a stopper of a specific level?????????

yes this is just the beginning!!!!

All of this is still ***, with HOW FITNESS IS ALL OF THIS?