Machine learning in trading: theory, models, practice and algo-trading - page 3561
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
We don't know the prices, but we know the average for N bars (N>> SMA period), which is surprisingly equal to the average of MA values.
We can take anything as price 0 (e.g. 1.0 or SMA "on the inflection"), build a "similar" curve from 1.0 by increments, then scale and position it.
In the general case, SMA requires a large number of bars (at first glance, it is a period squared in half, somewhere like that). But we do not have a general case - we have strict ranges and periodicity from nature.
And with large periods - SMA is close to sine-weighted (monotonic ma-mas are interchangeable), which again Fourier and inverse methods are known there
you are slightly confusing the mean value and the moving average.
you can get the original from the moving average, the accuracy will depend on the duration of the fixed "slip".
you cannot get an average from a one-step average, alas :-)
yumayo
Have you tried training a model from MS yourself?
why not just round up the prices and get pretty much the same thing without a tonne of terms and confusion?
The simplest example is shown to understand the process, also when processing more complex data.
The purpose was to show the possibility to accentuate the attention of the model to different data using quantisation.
Well, who didn't understand - it's not my fault :)
I somehow gave up dealing with any kind of price smoothing, as it smears and simply erases information about the chart.
But the price position is interesting here
I got the best results with the price position feed. And only on the hourly chart. Apparently, the pattern lies between sessions
You can say no to anything, in my philosophy I rather rely on the popularity of methods, and MAs are very popular both among traders and CBs, so it is foolish not to use them.
A simple example is shown to understand the process, also when processing more complex data.
The purpose was to show the possibility to emphasise the model's attention to different data using quantisation.
Well, who didn't understand - it's not my fault :)
If we have N prices and a moving average of period K, it will be known at N-K+1 points. Obviously, at K>1 the averages are simply not enough to reconstruct the prices. But if we know the first K-1 prices in addition to the values of the moving average, we can restore all the subsequent prices.
Naturally, we are talking about restoring those prices for which the average was calculated, not all OHLC.
Naturally, we are talking about restoring those prices for which the average was calculated, not all OHLC.
Not necessarily the first ones. Just different in modulus.
and they (and all "restored" prices) will be in close neighbourhoods of the average, and when N is large, their average value is equal to the average value of MA and so on (there are many additional conditions and properties).
in our conditions, when the value fluctuates in a small range, does not change much at each step, MA has reversals, all K points can be "restored". I wrote above - it's a question of feasibility, data availability and computational complexity. Actually the whole dialogue is about the fact that information is not lost. Just some periods are strengthened at the expense of weakening others, we get a different representation of the same price.
If instead of uniform weights (SMA) there is a "piece of sinusoid", then N is surprisingly small - a few periods of K.
not necessarily the first. Just different modulus.
and they (and all "restored" prices) will be in close neighbourhoods of the mean, and at large N their mean value is equal to the mean value of MA and so on (there are many additional conditions and properties).
in our conditions, when the value fluctuates in a small range, does not change much at each step, MA has reversals, all K points can be "restored". I wrote above - it's a question of feasibility, data availability and computational complexity. Actually the whole dialogue is about the fact that information is not lost. Simply some periods are strengthened at the expense of weakening others, and we get a different representation of the same price.
If instead of uniform weights (SMA) there is a "piece of sinusoid", then N is surprisingly small - a few periods of K.
You've made something mysterious here. I was writing about a very simple thing - if the value of K-means and K-1 of previous prices are known, then from the definition of the mean you get a linear equation to find the price. And if you know the first K-1 prices, you can find the price K, then K+1 and so on.
It is also possible to further define the average in the first K-1 points as the average of the maximum possible period - at the moment 1 just the first price, at the moment 2 - the average of the first two, etc. For such an average, the price is reconstructed by itself without knowledge of the initial prices.
You have made something mysterious here. I was writing about a very simple thing - if the value of K-mean and K-1 of previous prices are known, then from the definition of the mean a linear equation for finding the price is obtained. And if the first K-1 prices are known, then you can find the price K, then K+1 and so on.
It is also possible to further define the average in the first K-1 points as the average of the maximum possible period - at moment 1 just the first price, at moment 2 - the average of the first two, etc. For such an average, the price is reconstructed by itself without knowledge of the initial prices.
once again - not necessarily the first ones. Just K-1 points, and their indices modulo K are different.