Why is the normal distribution not normal? - page 15
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
If your figure is to be believed, the bigger the timeframe the higher the Correlation.
There are even positive correlations on large...?
Look again at the Fig. It shows in red the correlation coefficients (CC) between neighbouring samples for the first EURCHF difference series as a function of the selected Timeframe. Remember, the correlation coefficient is considered to be large, and the relation between the samples is "strong", if the correlation coefficient is close to 1. On the contrary, if KK tends to zero, then there is no (or weak) connection between samples, or it is non-linear (for Price VR like this, we can forget about it - if there is any connection, it is linear and simple). Positive QC means that there are group effects in the initial price series, i.e. "if the trend has started, it is more likely to continue", "trend is frend", etc. Negative QC characterizes the market as a "rollback", this is when the trend is likely to turn at the next readout (bar) rather than continue its movement in the chosen direction.
You can see from the fig. that CK for the price RPR is noticeably different from zero in all TFs and decreases (gets close to zero in absolute value) with the growth of the TF, which suggests loosening of relationships and, consequently, decrease of the forecast reliability for large TFs. This fact is in contradiction to your "...If you believe your figure, the larger the timeframe the higher the Correlation."
To give you an example, I have blue colored QC values for white noise created by a Random Number Generator (RNG). As you might expect, its RNG QRs hang around zero, signaling to us that there is no meaningful relationship between the samples. The fact that the RNG QC is not exactly zero, but takes small, but different values, is of statistical nature and is related to the finite length of the initial BP and has nothing to do with the "non-ideality" of the RNG used. It is well visible, how with growth of TF, the amplitude of noise "zero" grows because of reduction of number of samples of VP received from one and the same initial minute series. Thus the BP with TF500 m is 500 times shorter than the BP with TF1m.
Now about statistical significance of the obtained result. Comparing correlograms of real price series and synthetics of the same length, the error range related to the limited length of input data is clearly seen. Since there are no other sources of errors that can affect the result, we can reliably say that within the "noise" corridor clearly visible on the chart, the AC values are statistically significant.
Grasn wrote >> Well yes, that's a somewhat bold statement by Sergei about significant stat-value for increments. It's objectively speaking below the plinth there.
This is me being an artist trying to explain what is statistically correct by looking at the picture. To be precise, we don't need to compare the behavior of two QFs in order to estimate statistically correctness. We need to slice the original BP into 100 identical pieces and construct a QC(TF) for each one. We will get 100 non-matching QC values for each TF. We find the average value for each one and plot this point on the graph. Then we need to plot the confidence interval within which our values are located (for a certain TF). For this purpose, knowing the MI we find the standard deviation (SD) and plot the vertical whiskers for each point on the plot. These whiskers correspond to the confidence interval of 1/e level, the probability of being in this interval for CB is 68%. That's it. If the obtained QC values do not touch zero with their whiskers, then we can say that there is a statistically significant relationship between adjacent samples in the RRR of the price series. See figure below:
The QC(TF) for the LFG within statistical limits is "zero", which is to be expected. The KK(TF) for the price series is statistically significantly different from zero, indicating the presence of obvious dependencies between bars for different TFs. For large TFs, the length of the series used is clearly insufficient.
This material is proof of my assertion above. A separate question is about the significance of the dependencies found. Unfortunately, for us as traders, these relationships are of practical value only if the product of an instrument's volatility on IR for the selected TF exceeds its transaction costs (spread of a brokerage company). This is not observed and to find an algorithm for profitable trading it is necessary to use additional methods of BP analysis, in particular to find time "windows" with high values of the product of volatility on KK and evaluate their ergodicity (stability).
>> How's that?
Ugh.
"Long thought, read the pager." (your last posts on the mql forum).
Summarized roughly as follows: your question is too practical for such an academic branch... :)
// In general - it seems to me to enter into such an intimate relationship with the market in front of the public somehow unseemly.
// The people won't understand us and they'll be right. And if they do... - we'll still be wrong. ;)
Go ahead and poo-poo in private. I'll report back tonight when I get home from work.
As an artist, I tried to explain what statistical validity is by looking at a picture. To be precise, we don't need to compare the behaviour of two FCs in order to assess the plausibility. We need to slice the original BP into 100 identical pieces and construct a QC(TF) for each one. We will get 100 non-matching QC values for each TF. We find the average value for each one and plot this point on the graph. Then we need to plot the confidence interval within which our values are located (for a certain TF). For this purpose, knowing the MI we find the standard deviation (SD) and plot the vertical whiskers for each point on the graph. These whiskers correspond to the confidence interval, the probability of being in this interval for CB is 68%. That's it. If the obtained QC values do not reach zero with their whiskers, we can talk about statistically probable relations between adjacent readings in the BPR of the price series.
There could be another kind of unreliability in the EURCHF Minutes. The instrument is low-liquid, and in the night time it is drawn by DC according to its own not always clear logic. It obviously filters out a part of possible combinations of its EURUSD and USDCHF major during the current minute. It is probably possible to use (all the same night pipsers), but they will not allow earning normal money. They will raise spread, change filters, etc. This is because this is a direct loss for brokerage companies and there is no overlap at such moments.
Therefore more trust in minutes of liquid majors at liquid times of the day. This will weed out the dependencies of the quote filters that will not let you use them anyway.
Lots of bukafs.
The given material is a proof of my statement made above. A separate question is about the significance of the found dependencies. Unfortunately, for us as traders, these relations are of practical value only if the product of an instrument's volatility on KK for the selected TF exceeds its transaction costs (spread of a brokerage company). This is not observed and in order to find the profitable trading algorithm it is necessary to use additional methods of BP analysis connected with searching for time "windows" with high values of the product of volatility on KK and evaluation of their ergodicity (stability).
Are you writing all this on purpose so that you can end by saying: sadly, it's all rubbish?) It's not the first time I've noticed your post. But with pleasure I read from the beginning. Now, of course, not every one, but still interesting.
IMHO, any dependency can be used in TS, as the market cannot be described by a single formula anyway.
Таким образом, Сергей, статдостоверность определяется вовсе не "близостью" или "дальностью" КК от нуля, а статистическим разбросом полученного результата.
And where did I write that stat validity is defined by "range"? Determined by error evaluation (and not quite a spread, if you carefully reread my post) Which means that 0.4 - no significant value simply by definition, and an error in its calculation will eat another 20%-30% of its "body". Sergey, there's no correlation there, there just isn't. Accept it as a fact and take it easy. Exactly for these reasons (or rather, for these reasons as well) it's useless to predict for a few counts, it's useless to do it in any way, including NS, you have to predict in a fundamentally different way.
As an artist, I tried to explain what statistical significance is by looking at a picture.
Sergey, I'm not a professional mathematician (I guess you're not either :o), but it's obvious, that a part of your scientific conclusions about the correlation and especially the statistical significance are questionable, to put it mildly. But that doesn't prevent us from sticking to our opinions. :о)
EURCHF minutes may have a different kind of unreliability. The instrument is low-liquid, and in the night time it is drawn by the DC according to its own not always clear logic. It obviously filters out a part of possible combinations of its EURUSD and USDCHF majors within the same minute. This may be possible to use (all the same night pipsers), but they will not allow earning normal money. They will raise spread, change filters, etc. This is because this is a direct loss for brokerage companies and there is no overlap at such moments.
Therefore more trust in minutes of liquid majors at liquid times of the day. This will weed out the dependencies of the quote filters, which we will not let to use anyway.
Trade not in DC and there will be no such problems. I think it makes sense to talk about real prices and not DCs.
Sergei, there is no addiction there, there just isn't. Accept it as a fact and take it easy.
Well, it's akin to religion... No, that's not our way!
But it doesn't prevent you from sticking to your opinions. :о)
Trade with a non-DC and you won't have such problems. I think it makes sense to talk about real prices and not DCs.
We are talking about the adequacy of the study. On non-DC data the picture may be different, and the problems are different. On the "real" market (ECN for example) there may be no deals at this time for this instrument and the spread will be crazy.
Well, it's akin to religion... No, it's not our way!
That's what it's all about. If we changed our opinions every time without a REAL reason, the history of Mankind would be very different :-)Well how can I say, so far you've proved (but done it brilliantly :o) that the correlation of the first lag increments is different from zero. In this case (from the outside), religion is the holy belief that 0.4, is somewhere - 0.6, and 0.6 is already, you know, not more than 0.99, in the sense that it's about 0.97, but not less than 0.9 for sure :o) :о) That's how it turns out so far :o)
Sergei, as an artist to the artist, take two series with correlation 0.4 and look at them, even with the naked eye.