Alternative and common approaches in the construction of TC - page 8

 
hrenfx:
Cram all the FIs you are interested in into Recycle2 at once and see the level of correlation. If you find it too high, remove the ones that contribute too little.

hrenfx , I haven't got into the hang of using these indulators yet. I have already put them all in mt4 though. I am going to do it tomorrow.

If you have the time, please advise the optimum size of European bonds for the input variant at mf=m30:

FGBMH1 - vs (FGBLH1+GFBSH1),

i.e. MIDDLE vs (LONG+SHORT)

 
hrenfx:
Yeah, spoke ironically about equality and fair price. The indices are rubbish.

Some kind of fair dollar index can be looked for in low liquid currencies, e.g. the same ruble when the price of oil stands ;)
 
hrenfx:

The alternative implies a very different approach. You think about what kind of financial instrument you need to have to make your TS work the way you want it to. After you have thought it over, you start thinking about how to create such a synthetic out of those FI that are available. Having created such a synthetic, you will perform the same optimization operations on us. The results of optimization, Out of Sample and all the other tricks aimed to get rid of the fitting will be better than those of the original instrumentation. What FI will you run your TS on?

Hmmm... Mathematical intuition tells me to think in the following direction. Suppose we have found a synthetic that is optimal by some criterion, let it be the maximum volatility (dispersion) for certainty, since it has already been discussed in this thread. Are you saying that the results of, for example, a breakout system on such a synthetic will be better? Pardon me, I take it rather to suggest that they will definitely not be better than the best of the FI included in the synthetic. Why not? Quite simply. Practically all FI on the market have not the same, but very similar statistical characteristics - shape of probability distribution, autocorrelation, frequency composition, and so on. Well, I have no grounds to assert that any linear combination of such FI will have statistical characteristics which will essentially differ from characteristics of each FI separately (most likely on the contrary). Do you have such grounds?
 
 
hrenfx:
Check.

No, it doesn't work like that. So, I've just throwed a few instruments into some synthetic, put it into Statistic - and what do I see? The distribution is still with the same exponential tails, as each of them separately (slightly cut the top, which agrees with the theory - TFT works just in the vicinity of zero /drawing attached for reassurance/). No significant autocorrelations (as well as there were none, second figure)... other characteristics are also normal. What changes? I will never believe that the key word in my reasoning is "at random". Can you show me at least one live synthetic instrument with "optimal" characteristics in the form of a history file? I will personally analyze it and if I find something really different, I will eat my hat in public (which I will buy with the money I earned on this synthetic:).


 
Virtually all FIs on the market do not have the same, but very similar statistical characteristics - the shape of the probability distribution

About the distribution. Strange as it may sound, but it seems that its estimation (histogram) cannot be built correctly in principle (or is very difficult), if the distribution of the general population, to which the sample under estimation belongs, is unknown a priori.

I addressed this question to the forum of Moscow State University:

http://www.mathforum.ru/forum/read/1/33909/ (no one said anything);

NSU forum (there are econometricians there):

http://www.nsu.ru/phpBB/viewtopic.php?t=22051 (one person said you can't build it);

recently found information about the same problem in the textbook as well.

It is mentioned that Kendall and Stewart have some information in "Theory of Distributions", but I haven't gotten to it yet.

There is no way to determine what number of intervals to take to construct the histogram, equal or equal probability, what should be the shift parameter that determines the position of the first central interval. Depending on all of these, the histogram score floats strongly (fatally) (i.e. depending on them in principle a different law test can be successfully passed). Does the test done by "Statistics" solve all the issues mentioned?


P.S. I also wanted to build a histogram estimator of the distribution, but could not yet.

 
-Aleksey-:

About the distribution. As strange as it may sound, it seems that its estimation (histogram) cannot be correctly constructed in principle (or it is super-complex) if the distribution of the general population, to which the estimated sample belongs, is not known a priori.

I addressed this question to the forum of Moscow State University:

http://www.mathforum.ru/forum/read/1/33909/ (no one said anything);

I asked on the forum of NSU (experts in econometrics there):

http://www.nsu.ru/phpBB/viewtopic.php?t=22051 (one person said you can't build it);

recently found information about the same problem in the textbook as well.

It is mentioned that Kendall and Stewart have some information in "The Theory of Distributions", but I have not got to it yet.

There is no way to determine what number of intervals to take to construct the histogram, equal or equal probability, what should be the shift parameter that determines the position of the first central interval. Depending on all of this, the histogram estimate will float strongly (fatally). Does the test run by "Statistics" solve all the issues mentioned?


It may float at some point, but certainly not when everything is so obvious. Here, it may go sideways or backwards, but the histogram is the same. So the test conducted by Statistics in this case is designed to give a visual picture, after which the desire to check the figures disappears for lack of use.

Econometricians, by the way, say so because they try to use normal distribution everywhere to apply their mathematical apparatus - but it does not fit...

 
It may float at some point, but certainly not when everything is so obvious. Here, it may go sideways or backwards, but the histogram is the same.
I would only assume this once the test has been run with different values of the parameters mentioned. It is quite possible that other laws might work for compliance (if they are among those tested, of course, and if not?). That's the problem.
 

An example of a stationary synthetic:

In general, stationarity is not necessary.

 
hrenfx:
By the way, you can put the majors in Recycle2 together with USDLFX and immediately see the exact formula by which LiteForex calculates its dollar index.

We see that only USDLFX, AUDUSD, EURUSD, USDJPY, GBPUSD, USDCAD and USDCHF are correlated.

From these ratios we see how USDLFX is calculated:

It is now clear that it is calculated by an absolutely meaningless formula:

USDLFX = ((USDJPY * USDCHF * USDCAD) / (EURUSD * GBPUSD * AUDUSD))^(1 / 7)

P.S. Even this version of the index makes more sense (there would be degree = 1 / 6).