Hodrick-Prescott filter - page 6
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
I used to buildneural networks, about 20 years ago. It wasn't even called that yet. They were R&D on recognition, and it was then that the mathematics of these procedures were written. I later saw these procedures in funder, the ones we were counting and creating, hoping that they would replace the human brain, but no, they did not. To think in advance that others are smarter than you, is a loss.
Regarding prediction, one argument so far is that it's easier to predict the direction, yes I agree, easier. But that doesn't mean it's better.
In my opinion - if it's easier, then definitely better (with Zhirik's accent :) ), especially in such a difficult market as Forex....... >> Why go for a detour through the gullies if you can go straight on - along an asphalt road?
I think you are talking about the same thing.
As I understand it, Leonid, predicting the direction of movement, considers
Close[0] - Close[fcastbars],
if more than 0, then upwards, if less - downwards.
I.e. solves the classification problem.
Prival talks about predicting the magnitude of this increment. Even if it doesn't use NS, it's a regression problem.
Both problems can be solved on the same architecture. I, for example, classify on PNN, regression on GRNN, they are not fundamentally different.
It may be easier to classify, but, Leov, sometimes it is not enough. I remember the first VNS in mql when I was testing it, I got more than 70% of correct directions on a couple of inputs and a hundred of neurons.
But when I inserted the classifier into an Expert Advisor it turned out that the average profit trade was much less than the average loss trade, reducing it practically to zero.
This is a simple example that the right direction may not be enough to make a profit.
-------------
By the way, Close[0] - Close[fcastbars] is the swing of the fcastbars period from the first price difference. There are retuns all around.
Neutron, can you write how the muving difference equals the muving derivative? EMA do you mean?
on a couple of inputs and a hundred neurons I got more than 70% of correct directions.
But when I stuck the classifier into the Expert Advisor, it turned out that the average profit trade was much smaller than the average loss trade, reducing m.o. to almost zero
This is a simple example that the right direction may not be enough to make a profit.
It must be overtraining or what is more likely a common effect on time series - "tomorrow will be like today, today will be like yesterday" .... )))))
I don't know how anyone here thinks or thinks, but in my mind and concept, predicting price series or price range increments was abandoned 10 years ago - due to the futility and low profitability of the activity...... ))))
Absolutely agree! I gave the formula out of general considerations.
If we are talking about the relevance of a forecast for a market-type price series, it makes sense to forecast ONLY the sign of the expected movement of the quote. This has already been mentioned many times in scientific literature on trading and my personal experience.
Neutron, can you write how the muving difference equals the derivative of the muving? EMA you mean?
Yes, in three ways, with varying degrees of mathematical rigour. Starting from "...just look for yourself..." and ending with synthesis of bandwidth (AFC) of an ideal differential operator from two LPFs, but a little later - I need to find literature on the subject. Let me remind you that we are talking about the difference of two muvings differing in the smoothing parameter by an extremely small value. If it is EMA, the difference a1-a2 must tends to zero; if it is a standard moving average, the difference in smoothing periods must equal one.
That's odd.
You talk about autocorrelation. But I don't understand much of it. Maybe it's different from the known 'autocorrelation function'.
Where do you get that 0.9-0.6 or negative 'always' from ? Cross your heart, I think it will help.
The ACF of a price series and its first difference has a very nice form that shows that the series is predictable (not a delta function).
Here, ACF is built from EUR/USD 2006 minutiae. The algorithm is given.
You can see that ACF is small on average and negative (almost) for any TF. The TF is plotted on abscissa in minutes.
And how do you keep MM in this state of affairs? After all, you also need to know how much the currency will change by.
I'll go to the entresol (for books))) - I have not seen a corollometer in the form of a device for 20 years(((
Let me remind you of old autocorrelation tasks set back in the days of the World War II:
1. We have a locator, the ICO has a group target, i.e. all merged, the illumination is a big spot, as the diagram is 3 degrees)))
The question - what is going on there? This is still Raid or already Air Combat?
2. At night a plane flew over us(c) How many engines it has even if it fell into the Ocean.
Let me remind you the "formula" of autocorrelation - it is a convolution in which the kernel is the function itself taken on a segment (window).
Therefore, the statistical treatment of ACF is somewhat different in the physical sense and in the sense of getting the output correlogram,
which correlogram in turn has the width of the window, i.e. in discrete counts the width of the correlogram is equal to the width of the core))
i.e. the correlogram is an accumulation of the sums of the autocorrelational multiplication segments)) of that window.
So if the counts are positive, then the correlogram is always positive, and is a function (curve, graph, diagram) .
-One of the properties of a correlogram diagram - as a sort of display of similarity of a segment of the window to the whole data stream)))
for trading in rough approximation it is something like a pattern search,
and also roughly we can say that correlogram is a graph of similarity of BP to itself in the window.
..
P.S. I put smiles on the explanation on my fingers.
P.P.S If we set the task of predicting the sign of motion, then the initial BP should be transformed accordingly,
otherwise we will get a degenerate result.
P.P.S If the task is to predict the sign of movement, then the initial BP should be transformed accordingly,
otherwise we will get a degenerate result.
Share your experiences, if you don't mind, I wonder what kind of conversion you think is "appropriate".
I'll go to the entresol (for books))) - I haven't seen a Corellometer in the form of a device for 20 years((
Let me remind you of old autocorrelation problems from World War II times:
1. we have a locator, on the ICO a group target, i.e. all merged, illumination by a big spot, as the diagram is 3 degrees)))
Question - what happens there? This is still Raid or already Air Combat?
2. At night a plane flew over us(c) How many engines it has even if it fell into the Ocean.
Let me remind the "formula" of autocorrelation - it is a convolution in which the kernel is the function itself taken on a segment (window).
Therefore, the statistical treatment of ACF is somewhat different in the physical sense and in the sense of producing an output correlogram,
which the correlogram in turn has the width of the window, i.e. in discrete counts the width of the correlogram is equal to the width of the kernel)))
i.e. the correlogram is an accumulation of the sums of the sections of the autocorrelational multiplications)) of that window.
I don't understand so many clever words at once - I need time to comprehend:-)
So if the counts are positive, then the correlogram is always positive, and is a function (curve, graph, diagram) .
In a series of first differences, the counts are cognate, What are you writing about?
And how do you observe MM in such a state of affairs? After all, you also need to know how much the currency will change by.
This is where you come in.
Now, as for the synthesis of the first derivative, by crossing two muwings.
Suppose that all possible moving averages in the first approximation smooth the price series no worse and no better than a conventional moving average. Let us define the moving average as an average of n values of quotes to the left of the current value (upper equation).
The first derivative of the moving average, let's define it classically (second equation). Then, the difference of two moving averages that differ in the smoothing window width by 1 will give us the first derivative accurate to the smallest term of order 1/n (the third equation).
Drawing two moving averages, one with period 100 and the other with period 70 (see figure on the left, the red line represents a quotient, and the blue line represents a moving average), you can graphically show the validity of this statement.
On the right, the red line indicates the derivative of the MA with a period of 100. Blue shows the difference between the MA with periods of 100 and 70 samples, the green shows 100 and 99.
We can see that the agreement is not bad. Which is what I needed to prove. All other cases involving different realizations of muvings, do not change the essence of the statement, only the method of proof is varied.
to Neutron
....If one is looking for a sign prediction, the BP should be converted from a series of counts to a series of slopes without wool,
e.g. a series of regressions from n bars, or another option = output of the HP considered in this thread,
....
Would share if I had, but otherwise, I'm from old memory. Corellometers from Elecronics 60)))
However, under present conditions I even pulled down Matlab so as not to distract from my trading tasks,
I took Matlab down for the following reasons:
ACF as I understand it is accumulation, i.e. only after passing the 100th window can I trust the result,
in trading it is interesting only for large depots in the classical game, duration of position holding=months,
but for me personally as a rabbit trader((( unsuitable.
I can take appearance frequencies from autocorrelogram, i.e. use autocorrelogram as a frequency distribution of something,
but whether it is useful for a particular TS I doubt it.
P.S. although the convolution looks suspiciously like a perseptron (it seemed)))