Put in a good word about the occasional wanderer... - page 24

 

It would probably be better to change the task for the market - add that there are also 1000 princesses and they choose at the same time when the next bridegroom is shown. If two brides like the groom, he will only be half of them)). Once a bride has a total of 1 groom, she no longer participates in the show. What is the bride's optimum strategy? :)

The problem is clearly not solved if the brides are not interested in each other's strategies.))

 
Mathemat:

The prince is the accumulated paper profit from when the princesso took a pose (opened).

+ hundred and fifty

And if we imagine that princes are stamped like on an assembly line (as soon as someone dies off, a new one is added to the queue), then the "desire" of the princess to constantly have (or at least strive for) the best groom available is akin to the task solved by the Risk Analysis departments.

 
MetaDriver:

You've got it all wrong.

Forex is a princess. When she thinks she's already "THERE HE IS !", she makes her choice and turns back around. :)

That's the way it is, as the market always has the last and decisive word.

The whole ugly thing is that if the deal is presented as a groom, then the market can also have several grooms in a row as a princess. T h e conditions of the problem are quite different.

 

Still, I don't really understand why we don't want to apply the condition as close to the original problem as possible?

We have 1000 price readings, which come one after another, we need to choose one of, say, m=10 maximal ones with maximal probability. The problem is fundamentally decidable, the necessary and sufficient condition for it is to know the probability distribution of the next quotes for each moment of time, albeit arbitrary (if we assume that the quotes are dependent). And this question can be easily solved, estimating parameters of the conditional distribution when its form is known is a quite standard task.

The method is further described in the article, except that calculations will be more complicated... but who prevents them from doing them numerically, without piling up formulas?

The problem here is not so much in this, but in the fundamental possibility of solving the problem at the "profit" level, because in order to use the result in practice, we need to get with a probability of over 50% at both maximum and minimum of the interval, ie.i.e. Pmax*Pmin>=0.5, where Pmax and Pmin>=0.7071, i.e. it is necessary to choose m in such a way to provide no worse than 71% guess of the maximum-minimum, which may be practically unrealizable.

But on the whole, in my opinion, the problem in this formulation deserves close attention. I'll probably do it anyway.

 
alsu:

Still, I don't really understand why we don't want to apply the condition as close to the original problem as possible?

We have 1000 price readings, which come one after another, it is necessary to choose with maximal probability one of, say, m=10 maxima. The problem is fundamentally solvable, the necessary and sufficient condition for it is to know the probability distribution of the following quotes for each moment of time, albeit arbitrary (if we assume that the quotes are dependent). And this question can be easily solved, estimation of parameters of a conditional distribution with its known form is a quite standard task.

The method is described in detail in the article, except that the calculations will be more complicated... but who's stopping us from doing them numerically, without a bunch of formulas?

The problem here is not so much in this, but in the fundamental possibility of solving the problem at the "profitable" level, because in order to use the result in practice, we must get with a probability of more than 50% at the maximum and minimum of the interval, ie.i.e. it must be Pmax*Pmin>=0.5 where Pmax and Pmin>=0.7071, i.e. it is necessary to select m in such a way as to provide no worse than 71% guess of the maximum-minimum, which may turn out to be practically unrealizable.

But on the whole, in my opinion, the problem in this formulation deserves close attention. Probably, I will be doing it after all.

if we know the type and parameters of the future conditional distribution, is that not already enough to make money? How do we know this, or how do we get it?
 
Avals:
If we know the type and parameters of the future conditional distribution, isn't it already enough to make money? How do we know this, or how do we obtain it?

Not enough. The presence of dependence does not mean that the expectation of the conditional distribution will be different from zero. Moreover, I will tell you, as far as I have researched it deeply, the MO of conditional distributions is exactly 0 or very close to it for the depth of dependence to be at least 3 bars. The main content of statistical dependence, therefore, is the effect of previous quotes on the variance of subsequent ones.

We obtain the conditional parameters quite simply. The conditional density of the current value x0 of the difference series depending on the previous x1 is sought as W(x0/x1) = (a0+a1*x1)/2 * exp{-(a0+a1*x1)*|x0|} - it is an exponential distribution with the variable linearly depending on the previous quote. I've been investigating the form of this function and I can say that this form of writing fits the market very well. And then we adjust the parameters a0 and a1 to the current series by any known method and use them.

 
alsu:

Not enough. The presence of dependence does not mean that the expectation of the conditional distribution will be different from zero. Moreover, I will tell you, as far as I have researched it deeply, the MO of conditional distributions is exactly 0 or very close to it for the depth of dependence to be at least 3 bars. The main content of statistical dependence, therefore, is the effect of previous quotes on the variance of subsequent ones.

We obtain the conditional parameters quite simply. The conditional density of the current value x0 of the difference series depending on the previous x1 is sought as W(x0/x1) = (a0+a1*x1)/2 * exp{-(a0+a1*x1)*|x0|} - it is an exponential distribution with the variable linearly depending on the previous quote. I've been investigating the form of this function and I can say that this form of writing fits the market very well. And then we adjust the parameters a0 and a1 to the current series by any known method and use them.


But the distribution depends on the input data and knowledge of the process. That is, you have investigated the known dependencies and found various effects in the volatility memory and based on that you can build conditional variance distributions. Based on this model mo=0 on a large ensemble of data. But that doesn't mean that there really is no memory in the direction of the increments, only in the magnitude of the increments. So princes may well not randomly go to the bride, but those who are worse and those who are better may go first)) Or in some other non-random sequence. And that fact will confuse all the cards. The scheme works if the princes go in at random, regardless of their goodness and the goodness of those who went in before them. Of course, if there are only dependencies like very good prince will be followed by very good or very bad prince (value dependencies), then yes, the problem can be solved considering the type of these dependencies
 
Avals:

But after all, the distribution depends on the input data and knowledge of the process. I.e. you have investigated known dependencies and found different effects in volatility memory and from that you can construct conditional variance distributions. Based on this model mo=0 on a large ensemble of data. But that doesn't mean that there really is no memory in the direction of the increments, only in the magnitude of the increments. So princes may well not randomly go to the bride, but those who are worse and those who are better may go first)) Or in some other non-random sequence. And that fact will confuse the cards. The scheme works if the princes go in at random, regardless of their goodness and the goodness of those who went in before them. Of course, if there are only dependencies like very good prince will be followed by very good or very bad prince (value dependencies), then yes, the problem can be solved considering the type of these dependencies

I'm not saying that I know _all_ the dependencies, but I do know some and can estimate probabilities with them. As far as I know, parameters a0 and a1 float very slowly, with a period of a few hours on a minute chart, and fluctuate in a fairly narrow range, so you can calculate and use them.

The presence of a correlation does not mean that the princes do not come in by chance. For example, it could be that the probability of a "slightly better" prince coming after a "bad" prince is slightly higher than the probability of a "much better" prince coming, i.e. in this case there is some positive autocorrelation (as opposed to the classical scheme, where the probability of getting "much better" and "slightly better" is the same). Dependencies of this kind do not affect the performance of the scheme.

 
alsu:

...the probability of a "slightly better" prince coming in after a "bad" prince is slightly higher than the probability of a "much better" prince coming in, i.e. in this case there is some positive autocorrelation (unlike the classical scheme, where the probability of getting "much better" and "slightly better" is the same). Dependencies of this kind do not affect the performance of the scheme.

They do. The correlation here is not "some positive" correlation, but close to unity and on a very large lag space.
 
alexeymosc:
And there is an impact. The correlation here is not "some positive" correlation, but close to one and over a very large lag space.
I meant influence from principle-methodological point of view, and not from the point of view of practical achievability of result. The algorithm works if it is possible to estimate probabilities, and how it works - allowing it to trade profitably or not - is a subject of research, I can't answer this question in advance.