Machine learning in trading: theory, models, practice and algo-trading - page 258
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
SanSanych Fomenko:
I've been nosing around the idea that in financial markets this intuitive description gives ZZ................
zig-zag foul....
I did a little experiment...
As usually the MI trains, we make a sample, make a target, then shift the target one step back that would help the MI learn to predict the future value of the target based on the current values of the sample. I shifted the target 10 steps forward from the sample and it turned out that the MI knew what would happen when I saw the future target. And what do you think? At new data maximum that showed the MO is 75% of correct answers, although the idea should be about 100%
Conclusions...
1) zig-zagbullshit....
2) a bad target can eat up even 25% of MO's error
3) found some way to test the target for lousiness
This is utopia, but R is so popular that it has no such targets.
My thoughts on how we can fight the non-stationarity of the markets.
1)We have a certain trading system - TS. TS is an object that takes in certain settings that depend on the market and generates trading signals.
For simplicity, let's take the most primitive model of TS, let this be the RSI indicator
The indicator has a period - a kind of setting, the optimal parameters of which depend on the market
2.we will also create a trading rule, if PCI crosses zero from the top down, then it is a sell signal on the contrary - a buy signal
Here we have such a TS, very primitive, but we don't need more for the example, in the real TS we may have hundreds of market settings, the same with the trading rules
2)You don't have to be a genius to understand that this TS is doomed to failure, because the market characteristics are constantly changing, and the period of the indicator is fixed and it just does not react adequately to different market fluctuations
This leads to two thoughts:
3)It is possible to extract the market characteristics with the help of spectrum analysis with the help of three parameters only
1.the frequency of oscillation (we can say the period which objectively dominates the market at the moment)
The amplitude of the fluctuation (this is also important to know if the price is fluctuating in the range of 30 or 300 points)
Phase (it means the position from which we calculate the oscillations).
That's all, with these parameters we can describe absolutely any time series and the market in particular
4)Now we have to find the optimal period for the indicator for each spectral characteristic of the market, in other words we have to find the optimal settings for the TS. To find such optimal parameters probably need the MO.
5)So the final scheme of working with real time data appears to me as follows:
We take spectral characteristics from the real time market.
We pass it to the trained IL, the IL gives optimal characteristics for the TS at the current moment.
Transfer the optimal characteristics to the TS.
Drain the money :)
Think, comment on how this can be implemented, who knows what who knows what?
4) Now all that is necessary is to find the optimal period for each spectral characteristic of the market - those are the optimal settings for the TS. Finding such optimal parameters probably should be done by MO.
But it is possible to find the best period for the indicator by the strategy tester - to create an Expert Advisor on this indicator, the parameter for the period in the settings of Expert Advisor, and then optimize it by genetics.
By the way, this kind of Expert Advisor is not profitable, I tried it thousands of times with different indicators.
I have not tried it, but I do not believe that it will bring rsi to life and will bring profit.
Finished this later -
In general, the ability to trade using RSI is based on the hope that this RSI reflects some internal market processes, that there really is overbought and oversold, and that the indicator detects them correctly.
I have no doubt that such phenomena in the market existed in the 80s on daily charts, otherwise the indicator would not have been popular. But now there is nothing left of those regularities.
It may as well be possible to invent a bunch of formulas like (H+L)/O, and try to use it to predict something. The rsi has no more power now than such random formulas.
Determining the spectral characteristics of the market is probably a good thing, it sounds strong. Already these characteristics themselves can be used as predictors for the forest or neuronics, and use their prediction for trading, without rsi and other indicators.
But we need to make sure that spectral analysis gives stationary predictors (it is believed that for forex the answer is "no", although I have not seen any proof or examples, maybe they just muddy the water).
It seems to me that since the Fourier decomposition takes raw data and transforms it according to certain formulas and operations, obtaining a new data set, the new data does not have some new patterns specifically for forex, and the effectiveness of the model trained on them will not be better than when training on price.
But I would very much like to get some new logical rules of price formation during preprocessing of data and model training.
As an example - cluster networks and picture recognition. By analyzing data from individual neurons, you can determine the type of surface in the picture, all sorts of object boundaries, etc. The network first finds simple lines, corners, etc., then uses combinations of such primitives to identify contours of objects, then a combination of objects, colors, etc. to determine the type of object (actually there are more intermediate steps, I just described it in general).
Something similar must be done for forex - first the model recognizes price rise and fall, then it forms trends; using trends it forms patterns, and already makes decisions based on patterns. It can be done with cluster or deep nets, but there are so many details in training that it is scary to try it and it is unclear where to start from.
Above I tried to express some thoughts, using ZZ for illustration. Unfortunately, all my thoughts were discarded on the grounds that ZZ is junk.
But the point is not ZZ.
The point is that it is imperative to initially, on a descriptive level, determine what we are modeling in the market.
Every time something is discussed, supposedly amazing, but it doesn't specify WHAT we are going to model, what nuance of the market
Going back to ZZ, which we use exclusively to somehow structure the problem. And very visually.
Let us draw a graph and what do we see?
1. There are trends
2. There are deviations from trends - noise
We can clearly see periodicity, but it is somewhat strange: distances between tops on BOTH axes are always variable. Fourier has never seen such periodicity even in a nightmare.
Having looked at all this it is necessary to define WHAT we are going to model and how it will be connected with profit of trade.
And only after that we begin to define with tools, preliminary studying for what, on what conditions these tools have been developed, i.e. we prove applicability of tools.
P.S.
And do not enlighten me any more about the indicator Gold. All right?
But you can look for the best period for the indicator by the strategy tester - make an Expert Advisor on this indicator, the parameter for the period in the settings of the EA, and then optimize it with genetics.
This kind of Expert Advisor, by the way, does not bring profit, I have tried it thousands of times with different indicators.
This is the optimization, it's not even relevant to this topic, it is clear that it will fail and I understand why.
I have not tried it, but I do not believe that it will bring rsi to life and make profit.
What are your doubts at all I don't understand?
Spectral analysis is designed to take the characteristics of signals, amplitude, frequency and phase - there is nothing else, do you understand?
In general, the ability to trade using RSI is based on the hope that this RSI reflects some kind of internal market processes, that there really is overbought and oversold, and that the indicator detects them correctly.
OMG )) Forget about that RSI (stupid) I've written twice that RSI is just an allegory of the trading system to simplify perception, the TS also has some kind of settings that depend on the market and has an output trading signal, right? I never use indicators (standard) myself and I've written about it more than 5 times.
If it's easier for you, then replace the RSI with a two-legged arbitrage, whose settings are not fixed, and are constantly changing over time, depending on the objective spectral characteristics that are present in the market here and now. It is better :) ??
Might as well invent a bunch of formulas like (H+L)/O, and try to predict something with it. The rsi has no more power now than such random formulas.
100%, but no one's arguing that )
Already these characteristics themselves can be used as predictors for a forest or neuronics and use their prediction for trading, without rsi and other indicators.
You can't, they are just characteristics and nothing more, MO will not understand what to do with them...
You need to create a simplified model of the market or TS and this model should be adjusted to the market by spectrum characteristics, so you get statsionalnost and then you can hang MO from the top, you know?
Remember video from internet when tank is rushing at high speed on big pits and slides and turret is motionless! If you draw parallels, "hills and pits" are market people trying to trade the market with fixed parameters of the TC so that the "barrel of the turret") was not moving - do you understand how idiotic this is?
Those between the market and the MO in the middle should be inserted some layer which will make the data statsyonarnymi (or make the barrel of the turret when driving stationary)
But you need to make sure that the spectral analysis gives the stationary predictors (it is believed that for Forex the answer is "no", although I have not seen any evidence or examples, maybe they just muddy the waters).
They don't make things unclear, 99% of them just stupidly approximated the market through Fourier, and this is not what I'm talking about, that's why they failed, what they did is exactly what you can compare with some mashka or rsi or other nonsense
zig-zag is a dud....
I did a little experiment...
As usually the MO trains, make a sample, make a target, then shift the target one step back that the MO learns to predict the future value of the target by the current values of the sample, I took and shifted the target 10 steps forward relative to the sample, so it turned out that the MO knew earlier what would happen by seeing the future target. And what do you think? At new data maximum that showed the MO is 75% of correct answers, although the idea should be about 100%
Conclusions...
1) zig-zagbullshit....
2) a bad target can eat up even 25% of MO's error
3) found some way to test the target for lousiness
This is utopia, but R is so popular that it has no such targets.
Really "the human mind is limited, stupidity is limitless".
Where does this narcissism of your stupidity come from. You formulate your thoughts correctly: "I don't understand, I can't, I don't understand".
And your categorical evaluations sadden the readers with your adolescent maximalism.
Be modest...
It seems to me that since Fourier decomposition takes raw data and transforms it by specific formulas and operations, obtaining a new data set, the new data does not contain any new found regularities specifically for forex, and the effectiveness of the model trained on them will not be better than when training on the price.
It's not better, it's the same information in a different form, but that's not what I suggest.
As an example, cluster networks and picture recognition. By analyzing data from individual neurons, you can determine the type of surface in a picture, all sorts of boundaries of objects, etc. The network first finds simple lines, corners, etc., then uses combinations of such primitives to see the contours of objects, then a combination of objects, colors, etc. - determines the type of object (actually there are more intermediate steps, I just described it in general).
It should be something similar for forex - first the model detects rises and falls in prices, then it forms trends, then it forms patterns based on trends and makes decisions based on patterns. It can be done with cluster or deep nets, but there are so many nuances of training that it is scary to try and it is unclear where to start from.
All this will not work until the data is stationary, the market history is not repeating itself in the future
I apologize for my abrupt statement if I have offended or have grieved someone, but I have not changed the point of view, I have made experiment and have received result.
I am ready to take back my words if you will convince me in a contrary ...
Above I tried to express some thoughts, using ZZ for illustration. Unfortunately, all my thoughts were discarded on the grounds that ZZ is junk.
But the point is not ZZ.
The point is that it is imperative to initially, on a descriptive level, determine what we are modeling in the market.
Every time something is discussed, supposedly amazing, but it doesn't specify WHAT we are going to model, what nuance of the market
Going back to ZZ, which we use exclusively to somehow structure the problem. And very visually.
Let us draw a graph and what do we see?
1. There are trends
2. There are deviations from trends - noise
We can clearly see periodicity, but it is somewhat strange: distances between tops on BOTH axes are always variable. Fourier has never seen such periodicity even in a nightmare.
Having looked at all this it is necessary to define WHAT we are going to model and how it will be connected with profit of trade.
And only after this we begin to define with tools, preliminary studying for what, on what conditions these tools have been developed, i.e. we prove applicability of tools.
P.S.
And do not enlighten me any more about the indicator of the phase. All right?