From theory to practice - page 69

[СанСаныч Фоменко]

Yuriy Asaulenko:
SanaSanych, I am not talking about history, but about the principles of real-time signal processing. In this case, history is ancillary and we are more interested in the here and now.

I am only interested in history in the sense of the predictability of kotir in the future.

[Yuriy Asaulenko]

Nikolay Demko:

Oh, how many wondrous discoveries we have... )

You're half a step ahead of me).

Just wanted to talk about fast minded Neutons, and already...

And before that about Bollinger, about real-time.

[Mykola Demko]

Yuriy Asaulenko:
Sana-Sanych, I'm not talking about history, but the principles of real-time signal processing. In this case the history is auxiliary and we are more interested in the here and now.

In case you don't understand SanSanych, when you calculate some statistics you generalise within a window, so non-stationarity means that the statistics will change before you go out of the window.

In practice the next bar will have different statistics, because there is a half period delay in averaging. And exactly for this half period the forecasting power of the model is spread.

In other words, you can explicitly consider only those bars that have already entered the window and been calculated.

[Yuriy Asaulenko]

СанСаныч Фоменко:

I'm only interested in the story in the sense of the predictability of kotir in the future.

I don't think that's even possible, at least not systematically. I try to do without predictions. Only at the top-bottom level. Beyond that, life will show.

[Mykola Demko]

Yuriy Asaulenko:
I don't think it's possible, at least not systematically. I try not to make any predictions. Just up and down. And then life will show.

Hmmm, even upwards downwards is already a prediction.

Once we argued on the fourth forum, one of the authors said that one shouldn't predict anything, but one should follow the market.

But that's the problem, if you want to follow the market, you have to understand where it will go next, because it is so volatile that you cannot keep up with it.

[[Deleted]]

Nikolay Demko:

Now take the differential from the wrecker and get the same graph of the average.

And if you subtract the Mach Mach from the BB band and take the differential, you get the RMS graph.

You don't need to take the differential from the RMS chart, because it is derived from the scale. Just take the mach mach from BB and you get the same series as the RMS from the average differential.

so does bollinger use a sliding skew?

i was wondering why the upper and lower bands do not resemble the average.

)))

---

and what have we achieved in this thread after 70 pages? ))

[СанСаныч Фоменко]

Yuriy Asaulenko:
I don't think this is even possible, at least not systematically. Trying to do without predictions. Only at the top-bottom level. And then life will show.

Not only necessary, but possible.

If we are talking about regression models (as opposed to classification models), the main enemy is NOT stationarity. The main idea is to remove this non-stationarity.

• First we take increments instead of prices.
• Then we model the mean of this increment
• then model the variance of this increment
• then take into account the distribution of this increment (due to thick tails)

The purpose of all these tricks is to get a STATIONARY residual, and if the residual is not only stationary but has a normal distribution, then we have a proof of predictability of financial series' increments.

This predictability can arise in the first step is the ARMA model. I've seen a practical application of this model for data from some US agency - it turns out that their increments are stationary.

Each of these steps has its own nuances, but all of these nuances were killed by the author using the stationarity assumption - the financial series increments that do NOT happen are investigated.

[[Deleted]]

Nikolay Demko:

In case you don't understand SanSanych, when you calculate some statistics you generalise within a window, so non-stationarity means that the statistics will change before you go out of that window.

In practice the next bar will have different statistics, because there is a half period delay in averaging. And exactly for this half period the forecasting power of the model extends.

What is important is "how fast the changes will occur".

Nikolay Demko:

In practice, the next bar will have different statistics, because we have a half-period delay during averaging. Predictive power of the model is spread exactly for this half period.

In other words, you can unambiguously consider only those bars that have already entered the window and been calculated; you can't say anything about uncomputed bars.

this is for sb only

[Mykola Demko]

What settings did you use?

Note: the EA based on current candle MA crossover! So the MA can repaint on current candle!!!

Please check these wrong" trades with"Visual mode" (with slow speed and two MAs)?

Then you can see everything(actually there was MA crossover)!!!

[Yuriy Asaulenko]

Nikolay Demko:

Still, logically we understand that memory is present. It's EA's job to detect it, to bring it out into the light.

Let us imagine a black box (with traders and a marketplace inside) without any external influences, and living a life of its own - the output of the box: a certain flow of quotes. Even without external influences it will change somehow.

Now let this BS receive random (for the observer) delta-functions of different sign and intensity (news, for example). The NM starts to react somehow and we observe not the effects themselves, but the response of the NM to them + independent life of the NM itself.

The memory is there, but the output is a superposition of response to many events and the self life of the NM. Even in the case of a simple control system (ACS) the problem of division of what is not very solvable.