The market is a controlled dynamic system. - page 64

 
It is possible to determine at each step the value of n that minimises the prediction error of the previous step, and use it for the current step. With this choice, of course, the error will not be minimal at each step. But we should see what happens. So I will make this choice today.
 

The problem is that the bigger this n (you will search it at optimization stage, right?), the fewer NEW (last, i.e. newer) steps can be subject to optimization process. I mean that it will turn out that, say, to see, for example, what it will be at n=10, you can take for optimization from 1 only to "NOW" minus 10. I.e. the bigger this n, the less trust to optimization variant (due to information obsolescence, so to say).

 
alsu:

The deterministic component of the input signal and the structure and parameters of the system (blind deconvolution problem) is estimated using some optimization method on a chosen interval using the chosen criteria, then the input estimate is run through the model; the difference between the obtained output and the real process is thus a noise estimate.


i.e. even a positive forecast error is bad? If the forecast was +30p movement and the reality is +50p, why would that count as an error?

I.e. we are initially predicting only the direction (without takeaways, stops, etc.). An error then is an increment when the direction is opposite. Stupid system with entry at the opening and exit at the closing. The 2 values are profit and loss. Profit and error. That is what we are working with. In other words, the concept of error should be based on trading as a bid on directional movement imho

 
Avals:


i.e. is even a positive forecast error a bad thing? If the forecast was +30p movement and the reality is +50, why should this count as an error?

I.e. we are initially predicting only direction (without takeaways, stops, etc.). An error then is an increment when the direction is opposite. Stupid system with entry at the opening and exit at the closing. The 2 values are profit and loss. Profit and error. That is what we are working with. In other words, the concept of error should be based on trading as a bid on directional movement imho

Does it make sense to route the "positive" error through the positive feedback circuit to tune the system?

In general, this technique is used quite often - the main thing is not to break the overall stability of the system....

 
Avals:


I.e. even a positive forecast error is bad? If the forecast was +30p movement, but in reality it was +50p, why should it be considered as an error?

I.e. we are initially predicting only the direction (without takeaways, stops, etc.). An error then is an increment when the direction is opposite. Stupid system with entry at the opening and exit at the closing. The 2 values are profit and loss. Profit and error. That is what we are working with. In other words, the concept of error should be based on trading as a bid on directional movement imho


This may be good from a wallet point of view, but from a model point of view it means that we have understudied it. The logic is this: if there is a pattern that allows us to account for the possibility of exceeding the forecast in the positive direction, then this pattern is worth highlighting and adding to the model, it will not lose at least from this. If there is no such a pattern, then we can reformulate this fact as follows: it is impossible to determine beforehand in which direction the outlier exceeding our estimate will be. Thus, there is no way to make extra profit from it either.

I am a supporter of the fact that the model should be scolded for both "unprofitable" and "profitable" (well, maybe less, but still) errors. Otherwise we'll simply teach it to give lower forecasts: it's more advantageous for the system to give a "bad" forecast, because then it's less likely to be scolded: if the bad forecast comes true, then it was right, and if not, then it won't be scolded anyway. This is good from the point of view of the model, but not so much from the point of view of the result.

PS The analogy with planning in different organisations is quite clear: if the boss has a sufficient IQ, he will scold his subordinates for under- or over-fulfilled plans. This is the case in most successful companies in the west and is a real success factor. But here... Five years in three years, with all that implies.

 
alsu:

This may be good from a wallet point of view, but from a model point of view it means that we have understudied it. The logic is this: if there is a pattern that allows us to account for the possibility of exceeding the forecast in the positive direction, then this pattern is worth highlighting and adding to the model, it will not lose at least from this. If there is no such a pattern, then we can reformulate this fact as follows: it is impossible to determine beforehand in which direction the outlier exceeding our estimate will be. Thus, there is no way to make extra money out of it either.

I am a supporter of the fact that the model must be scolded both for "disadvantageous" and "advantageous" (well, maybe less, but still) errors. Otherwise we'll simply teach it to give lower forecasts: it's more advantageous for the system to give a "bad" forecast, because it's less likely to be punished: if the bad forecast comes true, it means it was right, but if not, it won't be scolded anyway. This is good from the point of view of the model, but not so much from the point of view of the result.

PS The analogy with planning in different organisations is quite clear: if the boss has a sufficient IQ, he will scold his subordinates for under- or over-fulfilled plans. This is the case in most successful companies in the west and is a real success factor. But here... Five years in three years, with all that implies.


The question is what are we forecasting. If the exact price after 1 bar (actually at a fixed interval of time), then yes, an error is any deviation from the forecast. But we can set the objective of the forecast differently - reaching the level in time, not reaching it, just the direction. The latter is more realistic. imha. I.e. the criterion depends on the objective of the forecast, and it may be different
 
Avals:

the question is what are we forecasting. If it's the exact price after 1 bar (actually a fixed period of time), then yes, an error is any deviation from the forecast. But we can set the objective of the forecast differently - reaching the level in time, not reaching it, just the direction. The latter is more realistic. imha. I.e. the criterion depends on the objective of the forecast, and it may be different

Quite right. The prediction objective can be formulated in many different ways.
 

And here, at each step, the value of nthat minimises the prediction error of the previous step is determined, and this value is used for the current step.

GBPUSD H4

and video

Files:
pp2.zip  1821 kb
 

;)

 

The previous picture(OpenSELL) was quite confirmed by the market --- by 1000 5-digit points going down.

And this is what is seen now on the H4:

Here the forecast value is one step ahead.

And of course, the forecast cannot be viewed as an absolutely accurate value, down to the penny. It is correct to view the forecast as the centre of some limited area of reach.