Machine learning in trading: theory, models, practice and algo-trading - page 1639

 
Aleksey Vyazmikin:

I completely agree.

I have repeatedly wondered about this issue, and I think it is necessary to compare the results of the system to its potential in a certain area.

I was just thinking about this today, how to do it better and more universally. I imagine the learning process to consist of several stages, the first of which is the partitioning of a sample, and you can partition it based on some signal strategies. These strategies should be primitive but have potential, for example, MA crossing by the price will generate a signal to enter in the direction of such crossover or vice versa. Then training is just a way to filter out false signals. If such assumptions are accepted, we can calculate the percentage value of efficiency of such filtration at each time section. The simplest one is to calculate the accuracy and completeness of classification relatively to the basic strategy. There are other options - metrics. Then we can see how the effectiveness of the model changes, even if it has begun to lose money.

It also seems like a good idea to build a system based on a complete set of primitive but meaningful systems. Completeness means that you can pick profitable systems from that set for any piece of quotes. Meaningfulness is roughly what you call potential. From here I'm following the way of building a portfolio of this set with weights depending on time.

 
Evgeny Dyuka:
I have practice. I have not noticed any changes within a month since the last training, even after the bit has plummeted. The only thing that affects it is the period right after the manipulative movement of the asset, during this time the neuronet is completely lost and displays all sorts of nonsense, the further away from such a storm the more adequate the predicates become.

Practice usually shows that "trees never grow to the sky". Sooner or later the equity/balance of any EA/portfolio will start to decrease considerably and something has to be done about it.

 
Aleksey Nikolayev:

The almost complete ignoring of the non-stationarity problem in this thread is quite disconcerting. For some reason, it is assumed that the patterns found in the past will work in the future, and if they do not work, then retraining has occurred. But, it is quite possible that some patterns simply stop working over time - gradually or even by leaps and bounds (for example, as a result of a crisis like the current one).

The problem I see is that MO patterns are complex and poorly interpreted by humans. If they start to work badly, it is impossible to distinguish (within the models) the overtraining variant from the non-stationarity variant. In conventional thechanalysis it is always possible to say: "change of trend", "level/channel breakdown" etc.

If I want to guess, I should use the "physics" of symbols quotes. Their main property, in my opinion, is the change, sometimes very fast and cardinal, of statistical characteristics of a time series. In this sense it would be reasonable first of all to create a classifier that would sort the history into sections with similar statistical characteristics and give them numbers from 1 to 20, say. And then for each such similar type of market to create its own individual TS. But how to invent predictors for such partitioning of time series into sections with similar statistical characteristics - I do not have a good idea.

 
sibirqk:

Imho of course, but in my opinion, we should rely on the "physics" of quotations of financial instruments. Their main property, in my opinion, is the change, sometimes very fast and drastic, of statistical characteristics of a time series. In this sense it would be reasonable first of all to create a classifier that would sort the history into sections with similar statistical characteristics and give them numbers from 1 to 20, say. And then for each such similar type of market to create its own individual TS. But how to think up predictors for such partitioning of time series into sections with similar statistical characteristics - I can't really imagine.

I usually call such areas "market states. Each state can be matched with some portfolio of primitive systems. I suppose, to segment the market into states and compare portfolios to them, we can use some recursive networks.

 
Mihail Marchukajtes:
Where the hell do you find a chick who's so good at neural networking? I don't know where to find such a chick who knows the business of neural networking and who can rant on such topics after johnnshpohan. I think we should move to the capital. That's where they all seem to be concentrating.

Usually you have to choose between regular johnannshpohan or endless talk about lofty matters.

 
Aleksey Nikolayev:

Each state can be matched with some portfolio of primitive systems.

All right, the point is that the price series is not continuous, it is piecewise continuous - depending on the valotility, it usually corresponds to the running time of the session

and that is why the hope to train a neural network by simply presenting it with a price series tends to zero, imho

but if you divide the price series by the time of sessions and train it by -session, then the information about overbought .... will be lost and the circle has closed again? - Nothing works

 
Andrey Dik:

Usually you have to choose between regular johnannshpohan or endless talk about lofty matters.

Well, if she will be distracted during oral procedures to discuss katbusting, then of course I'll be against it. But if a woman is not stupid in general, then it's cool :-)
 
Well, how happy I get when I have a home Internet. I can't get enough of it. I'm gonna go get a beer. The rest of you, what's up? Any plans?
 
Igor Makanu:

All right, the point is that the price series is not continuous, it is piecewise continuous - depending on the valotility, it usually corresponds to the running time of the session

and that is why the hope to train a neural network by simply presenting it with a price series tends to zero, imho

but if you divide the price series by the time of sessions and train it by -session, then the information about overbought .... will be lost and the circle has closed again? - nothing works

You can get rid of session volatility fluctuations by switching to a zigzag or renko, right? Of course, the natural time structure will suffer. But it is possible to introduce the normal time as an indicator defined for each knee/brick.

 
Aleksey Nikolayev:

It is possible to get rid of session volatility fluctuations by switching to a zigzag or renko, isn't it? Of course, the natural time structure will suffer, but it is possible to introduce normal time as an indicator set for each knee/brick.

I don't want to go back to Renko, I've already wasted time on it, not only does it completely lose OHLC information, but in addition you get a lag of 2 heights of Renko bricks - it lags very much

The same will probably be true for ZigZag, but I did not use it.