Machine learning in trading: theory, models, practice and algo-trading - page 464
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
GARCH
I have my own understanding of forex with practice and garch does not solve the problems I need.
It may turn out that garch easily solves many problems and provides results comparable to a complex MO using simple methods. But input data for garch is limited by price, and I think this is not enough. I.e. if we take MO model, give it prices, and compare predictions with those obtained with garch, it may appear that they are not inferior to each other. But if you give the MO model much more input data, it will gain a significant advantage.
In general garch and other machine learning models have a lot in common, working with garch you are much closer to MO than you think.
In both cases (MO and garch) you take a price, do all sorts of transformations with it to create predictors (indicators for normal models, and arima transformations for garch), and use them to model the price and try to predict gains.
Then in MO there is crossvalidation to make sure everything is ok, and for garch there are similarly some internal statistical checks.
In general the approach is very similar in both cases (do something with the price to satisfy all the model requirements and predict the future), just slightly different ways of doing it.
But apparently the price is too random, and the patterns found are very small. No matter how much time is spent training the model, its estimate will still be far from perfect, only a couple of percent higher than the estimate compared to a random guess.
Sometimes, they are lucky and these found dependencies, though small, exist for a long time and bring profit. But they always suddenly disappear. This is the main problem for me now - to find some external indicators, like financial reports, by which the model could be trained with a very good estimate, so that the found dependencies would be stable and would not suddenly disappear.
I have my own understanding of forex with practice, and garch does not solve the problems I need.
It may turn out that garch easily solves many problems and with simple methods gives results comparable to complex MO. But input data for garch is limited by price, and I think this is not enough. I.e. if we take MO model, give it prices, and compare predictions with those obtained with garch, it may appear that they are not inferior to each other. But if you give the MO model much more input data, it will gain a significant advantage.
In general garch and other machine learning models have a lot in common, working with garch you are much closer to MO than you think.
In both cases (MO and garch) you take a price, do all sorts of transformations with it to create predictors (indicators for normal models, and arima transformations for garch), and use them to model the price and try to predict gains.
Then in MO there is crossvalidation to make sure everything is ok, and for garch there are similarly some internal statistical checks.
In general the approach is very similar in both cases (do something with the price to satisfy all the model requirements and predict the future), just slightly different ways of doing it.
But apparently the price is too random, and the patterns found are very small. No matter how much time is spent training the model, its estimate will still be far from perfect, only a couple of percent higher than the estimate compared to a random guess.
Sometimes, they are lucky and these found dependencies, though small, exist for a long time and bring profit. But they always suddenly disappear. That's the main problem for me now - to find some external indicators, such as financial statements, by which the model can be trained with a very good estimate, so that found dependencies are stable and do not disappear suddenly.
I form my own understanding of Forex with practice and garch does not solve the problems I need.
It can turn out that garch easily solves many problems and provides results comparable to a complex MO using simple methods. But input data for garch is limited by price, and I think this is not enough. I.e. if we take MO model, give it prices, and compare predictions with those obtained with garch, it may turn out that they are not inferior to each other. But if you give the MO model much more input data, it will gain a significant advantage.
In general garch and other machine learning models have a lot in common, working with garch you are much closer to MO than you think.
In both cases (MO and garch) you take a price, do all sorts of transformations with it to create predictors (indicators for normal models, and arima transformations for garch), and use them to model the price and try to predict gains.
Then in MO there is crossvalidation to make sure everything is ok, and for garch there are similarly some internal statistical checks.
In general the approach is very similar in both cases (do something with the price to satisfy all the model requirements and predict the future), just slightly different ways of doing it.
But apparently the price is too random, and the patterns found are very small. No matter how much time is spent training the model, its estimate will still be far from perfect, only a couple of percent higher than the estimate compared to a random guess.
Sometimes, they are lucky and these found dependencies, though small, exist for a long time and bring profit. But they always suddenly disappear. That's the main problem for me now - to find some external indicators, such as financial statements, by which the model can be trained with a very good estimate, so that found dependencies are stable and do not disappear suddenly.
Inputs in financial markets are always formed by people who have their own interests, and different at different times under the same conditions.
It is especially strong on the TF up to H1. We get a random nonstationary process formed by the desires of relatively small groups of individuals.
Above D1 the law of large numbers starts working and also the direction of the movement is formed by the investments of the state, large corporations and politics.
On small TF it is meaningless to look for patterns - there are none. What was formed and we defined as a pattern is actually circles on the water: there were and there aren't. We can catch the beginning of a wave, cut it down and wait for the next occurrence of a similar wave, and there may be no such wave at all.
On large TF it is necessary to take into account a large number of factors, because initially there is no clarity about their influence on the target variable. But the fundamental point is that the source of movement is not a random process, but a deterministic one, long produced, purposefully maintained until fulfillment. We may not understand/know, but these movements can be attempted to be fished out with MO in the form of patterns, which are not eternal, but whose life can last for years.
It follows from this that MO and GARCH are mutually complementary tools and need to be combined in one trading system.
Inputs in financial markets are always formed by people who have their own interests, and different at different times under the same conditions.
It is especially strong on the TF up to H1. We get a random nonstationary process formed by the desires of relatively small groups of individuals.
Above D1 the law of large numbers starts working and also the direction of the movement is formed by the investments of the state, large corporations and politics.
On small TF it is meaningless to look for patterns - there are none. What formed and we defined as a pattern is actually circles on the water: there were and there aren't. We can catch the beginning of a wave, cut it down and wait for the next occurrence of a similar wave, and there may be no such wave at all.
On large TF it is necessary to take into account a large number of factors, because initially there is no clarity about their influence on the target variable. But the fundamental point is that the source of movement is not a random process, but a deterministic one, long produced, purposefully maintained until fulfillment. We may not understand/know, but these movements can be attempted to be fished out with MO in the form of patterns, which are not eternal, but whose life can last for years.
It follows from this that MO and GARCH are mutually complementary tools that need to be brought together in one trading system.
I'm tired of writing that there is no difference which timeframe to take, the risks are the same everywhere, there are no differences except in spread and commission for scalping, but that's another kind of negative factor. Any time series is described by only a few variables, maybe even by 2, such as persistence (heurst) and volatility (fractal dimension) and, say, the duration of these non-periodic cycles that should be separated from each other, everything. From this it is possible to build simple and reliable models with MO or with something else. From this it follows that it is never possible to build a continuously working model, because the quotes are random and non-periodic cycles are formed spontaneously, therefore the moments of transitions from one cycle to another are not predictable under the same conditions, for this purpose it is necessary to change the scale of representation to see these cycles as a constituent of a larger one. I think that understanding is the basis without which it's impossible to think of the market as a probabilistic system in the right way. Within non-periodic cycles even simple models work well, things will always break down when the cycles change. From cycle to cycle the market is always new and different. And what does the TF have to do with it?
Where do you get all this nonsense about that on small TF the desires of small groups of people prevail, and on large ones everything is determined. It has never been so and it's a finger in the sky, the markets are another mechanism, where everything is interconnected with everything, hence the scale invariance of all TF (investment horizons).
If we are to talk about any determinism, it is the dependence of all the TFs from each other, because the quote is the same, we only change the scale of the event, it's like moving along a fractal from less to more and vice versa.
Inputs in financial markets are always formed by people who have their own interests, and different at different times under the same conditions.
It is especially strong on the TF up to H1. We get a random non-stationary process formed by the desires of relatively small groups of individuals.
Above D1 the law of large numbers starts working and also the direction of the movement is formed by the investments of the state, large corporations and politics.
On small TF it is meaningless to look for patterns - there are none. What formed and we defined as a pattern is actually circles on the water: there were and there aren't. We can catch the beginning of a wave, cut it down and wait for the next occurrence of a similar wave, and there may be no such wave at all.
On large TF it is necessary to take into account a large number of factors, because initially there is no clarity about their influence on the target variable. But the fundamental point is that the source of movement is not a random process, but a deterministic one, long produced, purposefully maintained until fulfillment. We may not understand/know, but these movements can be attempted to be fished out with MO in the form of patterns, which are not eternal, but whose life can last for years.
It follows from this that MO and GARCH are mutually complementary tools to be combined in one trading system.
On any TF is a random non-stationary process. Ithas a non-stationary deterministic component and non-stationary random component also on any TF.
Using somebody's "desires" and other nonsense as "explanations" for randomness and nonstationarity is nonsense.
All information about the process is contained in the process itself (in the chart-history, "the price that accounts for everything"). The goal is to get the right information out of the process, which can then be used for our purposes. The task is difficult, but solvable.
The economy of any country is the production of goods and services, which is a very deterministic and highly inertial process. Examples can be found where such processes have been perfectly stable for decades. A change in the gross product of any country in a year by units of percent is either a success or a disaster. A 10 percent drop in a country's output can cause a social explosion.
What we are witnessing is a froth on this deterministic process. As of today, this froth is becoming more and more detached from the real economy, but at the level of the macroeconomy, at the level of quarterly natural economic indicators, of which there are tens of thousands, everything remains the same.
http://my-files.ru/9iyzd6
There was a topic on the forum somewhere where somebody used similar data to predict some index once in a couple of months.
I found this.https://www.mql5.com/ru/forum/40739(also FRED). The topic is interesting, maybe someday I will try to predict eurusd on monthly timeframe using this data.
Pay attention, indicators may be revised (redrawn).
The economy of any country is the production of goods and services, which is a very deterministic and highly inertial process. Examples can be found where such processes have been perfectly stable for decades. A change in the gross product of any country in a year by units of percent is either a success or a disaster. A 10 percent drop in a country's output can cause a social explosion.
What we are witnessing is a froth on this deterministic process. As of today, this froth is becoming more and more detached from the real economy, but at the macroeconomic level, at the level of quarterly natural economic indicators, of which there are tens of thousands, everything remains the same.
And there you go again with your barefaced rhetoric: "very deterministic," "extremely inertial..." You say "frothy"? and "everything remains the same"? It's as if you drew a picture in your head, and you describe that picture without caring at all about how out of touch it is with reality.
You are a kind of "econometrician" and should be able to understand... but no.
Try to read an article by Sergey Glazyev. It will be useful for understanding the situation.
Today my opinion is that we should trade deviations and not trends due to the doubtfulness of these trends in the near future.
To trade deviations we have our own device - GARCH, which is widely used in the financial markets starting from high-frequency trading. Dream limit is up to 100 pips. Thus, on TFs up to H1 we catch 15-20 pips and wait for the next signal. The less time in the market the better the Expert Advisor.