Discussion of article "Multiple Regression Analysis. Strategy Generator and Tester in One" - page 2

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
First of all thank you for your article which I read with great attention.
I was wondering how much your method differs from the built in MT5 optimizer (take som indicators, play them on past data, this give them some weight and apply the result to the "futur").
Is the difference between your Multiple Regression method and the built in MT5 method so great ?
Any comment from you on this question would be appreciated.
Thank you.
By the way I have almost zero background with statistics but your article was quite clear and very good quality.
Hi ArtemGaleev
Thank you for your awsome article, I have read it many times. And I have some questions:
1) In Fig.12, It show a good result but I think It's not fair because the EA is run on the trained data. The EA parameters were calculated from 6.30.2011 to 9.1.2011 and test from 7.1.2011 to 8.26.2011 (Fig.1). So, it was test on the data it was trained.
2) And I wonder how to optimize p-level. In this article, you said "removed the insignificant parameter with the highest p-level". I removed I run analyzed again, but all of p-level was changed and the parameter which p-level significant turn to insignificant. And in Fig.10, the table show five parametes: dDeMarker, dAC, DeMarker, Bulls, Bears. Are you run analyzed only on 5 params or more ? Some parameters be hiddend.
Maybe I wrong in some steps. Anyway, thank you very much for this article.
Другое, что нужно учитывать, это выбросы в данных. Редкие, но сильные события (в нашем случае скачки цены) могут внести ложные зависимости в уравнение. Например, после выхода какой-либо неожиданной новости на рынке произошло сильное движение, продлившееся несколько часов. В этом случае значения технических индикаторов имели малую значимость в прогнозе, но регрессионный анализ припишет им высокую значимость, поскольку было сильное изменение цены. Поэтому желательно фильтровать данные в выборке или проверять наличие выбросов в данных.
That's a very good point. One should not look for regularities where there are none. This applies to almost any method of analysis used.
Unfortunately, the most important component (analysing residuals is not the same thing) of such studies - the extent to which the "regularities" found are survivable - is not addressed.
"Regularities" are more survivable the longer the found weight coefficients (in terms of MR and any NS) "hold" (minimally change). It does not cost anything to find remarkable coefficients on the SB (where there are no regularities by definition) on any window. But it is the dynamics of their changes that will allow us to say with a high degree of truth that we are dealing with SB and not with something else.
I.e. the dynamics of the weights changes is a certain criterion of the presence of regularities in the initial BP.
To give an example, here is a video where you can see (upper right graph) how the weights of one of the research methods change linearly over time:
There is some smoothness at times (which is very good), as well as situations of jerking weights (which is extremely bad).
Moreover, market regularities depend on the time of day, seasonality, etc. Therefore, time zones should be considered separately when searching for them. Night trading of some crosses, which has been profitably used for several years, is a vivid example of a REAL regularity, which is present only in a certain interval of the day. And it would never have been found, if the whole original BP was investigated, without time zone filter.
P.S. Something carried (apparently, the atmosphere of the "Day of Knowledge" affects) on a lot of letters, so I cut off the post.
In many TSs, a trading signal is formed on the basis of a linear combination of indicators.
This means that if the task is to unravel a TS from the results of its trades (its steutment), the MR is a good tool for such purposes.
We can look at the task from another side:
A trader arranges trading signals on the history. This can be the result of his informalised manual trading, or some of the ZigZag tops. In general, anything.
Then the same task is solved: automatic formalisation of the TS by its trading signals. I.e. finding some linear regularities in the TS through MR.
P.S. You can also apply NS to such a task. But interpretation of NS results is much more complicated than MR.
We can look at the task from another perspective:
A trader places trading signals on history. This could be the result of his informalised manual trading, or it could be some of the ZigZag tops. In general, anything.
Then the same task is solved: automatic formalisation of the TS by its trading signals. I.e. finding some linear regularities in the TS through MR.
Yes, the idea is interesting. Even suggested writing an article that does the initial processing for this.
The other day published an article Visualise a strategy in MetaTrader 5 tester, which shows the processing of optitmisation results "on the fly". But the topic is not fully disclosed, of course. There is a whole layer of possibilities here, and there are topics for articles in this regard:
But no one has responded so far.
...
Then the same task is solved: automatic formalisation of the TS by its trading signals. I.e. finding some linear regularities in the TS through MR.
...
Gentlemen, how do you imagine it? Something like factor analysis. And what factors? All indicators with all possible sets of parameters from all possible symbols?
Let me say up front that these are just some thoughts out loud and the result of my complete ignorance of mathematics.
To begin with, let's assume that trading signals are formed exclusively by a linear combination of some indicators. Moreover, we have non-zero assumptions (wishes) about the TS under study. For example, that most likely MAs, Fibo, PriceChannel, etc. are used there. Or on the contrary, we want to see the formalisation of the TS with marked inputs through a certain list of indicators.
Further, as it was quite rightly stated above, the matrix method is applied to the table of all kinds of indicator values (including the price itself) and trading signals of the TS.
The trading signal itself represents three values: BUY, SELL, LOVE. For this (and other) reasons, it makes sense to consider separately the TS with only one type of signal. Let it be BUY.
The BUY signal itself is formed when some threshold is crossed in a certain direction. For convenience, this threshold is taken as zero.
Let's summarise the above:
We take only BUY-signals of the considered TS, in these points we take a set of values of the indicators we are interested in. And through the studied mathematical method we try to express a linear combination:
K[1] * Value[1] + .... + K[n] * Value[n] = 0, while imposing the restriction Sum(Abs(K[i])) on the weights. = 1. MR does not quite solve the problem at hand, since it is designed to express Value[j] through the others. Therefore, for each j the vector of solutions will have a different direction - not colinear. However, it will still allow to obtain solutions, even if not perfect, but not perfect. In addition to MR, of course, other mathematical methods can be used.
After finding the weights, we will need to plot the residuals: for each signal, calculate the value R[i] = K[1] * Value[1] + .... + K[n] * Value[n]. We will get a kind of graph of formalisation of the investigated TS, which will be a curve oscillating around zero. Of course, there will be outliers. It is desirable to throw them out - in simple language it would mean ignoring certain trading signals.
After that, we should apply the matrix method again. As a result, ifmost of the TS signals can be formalised through the selected list of indicators, we will get the final answer in the form of weights, where R[i] will "narrowly" fluctuate around zero.
But let's not forget that even if all of the above has worked out, it cannot be called formalisation of the TS. Because it is necessary to run the found combination further not only on the places of trading signals, but also on the whole price BP. And here it may happen that in our points the combination gives an excellent result, but in other points it gives false signals, which can hardly be filtered.
These are briefly the initial naive ideas about automatic formalisation of some (widespread) types of TS.
Everything is individual. For example, it is difficult to assess the usefulness and necessity of formalising the strategies of championship leaders, which is carried out by many enthusiasts every time.
My thoughts were simply about a tool to make this process much easier and faster. Some kind of support.
From a research point of view, it is another way of studying the market.
very similar to neural networks
SZY: and even if it is not NS, the result of such a tool will be similar to the work of NS - on the history of positive results