Machine learning in trading: theory, models, practice and algo-trading - page 660
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
Because the price change is not in times, but less than 10%
Yes, it's not interesting in that form.
But the logarithm of differences is curious, maybe, as a feature for classification - there the graph is stuck on the trends in one state, like buy/sell will not fail :)
Somebody advised it. Yes, they may have a completely different task.
I personally from the tails in general enjoy: the more - the better.))) And you fight them with the whole world.))) It's funny.
When normalizing with outliers, the center will stray a lot.
For example, you took a sample of 10,000 bars and found the center between max and min. Next training: we added 1000 bar and this 100 bar showed a stronger kick than the previous one. You get a new center. As a result the data will be incompatible.
For example the last bar of the first sample turned out to be =0, when normalizing by the first sample. In the second sample it is already 1000th from the end and in the new normalization it can shift to 0.2, for example. And the forecasts from the first sample with its normalization and from the second sample will be different, because even the input data for them are shifted vertically.
If the data is deleted or underestimated to the level zero will either be in place or walk, but not so much.
I'm cutting off 1% of the amount of data at the top and bottom for now. This way the zero, although it staggers, is much weaker. If you rigidly set the trim levels, the zero/center will always be in the same place. But I don't know what levels to choose, and for dozens of predictors to each his own.
But the logarithm of differences is still curious, maybe just as a feature for classification - the graph is stuck on the trends in one state, like buy/sell will not fail :)
Maybe... But does it predict at least one bar further? Or it shows the past as usual?
Maybe... But does it predict at least 1 bar ahead? Or it just shows the past as usual?
No way it predicts by itself... but if there is some combination of different lags then mb
If I normalize it with outliers, the center will fluctuate a lot.
For example, you took a 10,000 bar sample, found the center between max and min. Next training: we add 1000 bars and this 100 bar has a stronger overshoot than before. You get a new center. As a result the data will be incompatible.
For example the last bar of the first sample turned out to be =0, when normalizing by the first sample. In the second sample it is already 1000th from the end and in the new normalization it can shift to 0.2, for example. And the forecasts from the first sample with its normalization and from the second sample will be different, because even the input data for them are shifted vertically.
Yeah, I get it, all right. But there is a detrend! And the center will go there along with the distribution.
I have a distribution estimated at only 600-1000 points Tf 1 min. And the detrend is very short, and the center shifts very quickly. Yes, but this is at the stock futures. I still can't decide to try it on forex.
By the way. For weeks the variance is almost the same -+/- some points.
Yeah, I get it, all right. But there is a detrend! And the center will go there along with the distribution.
I have a distribution estimated at only 600-1000 points Tf 1 min. And the detrend is very short, and the center shifts very quickly. Yes, but this is at the stock futures. I still can't decide to try it on forex.
By the way. My aim is to try to change dispersion in some weeks.
But maybe I'm wrong... I'll have to experiment some more, if I don't forget to do other things.
It seems to me that it is desirable for the center not to move anywhere, especially quickly).
But maybe I'm wrong... I will have to experiment some more, if I do not forget about other things.
Well, this is already different approaches to the projectile.)
Once again I publish a picture.
On X - time, on Y - price, normalized from 0 to 60 - the usual time-price graph. On Z - the probability density of price distribution in time.
We can see that the distribution is formed around a certain price, then the price "jumps" to another level and the distribution is formed around another price.
If we keep up with it, we are always in the vicinity of the center of the distribution.
Yes, I forgot to say, in between directed movements from one price to another.
Well, this is already different approaches to the projectile.)
Once again I publish a picture.
On X - time, on Y - price, normalized from 0 to 60 - the usual time-price graph. On Z - probability distributions of price in time.
We can see that a distribution is formed around a certain price, then the price "jumps" to another level and the distribution is formed around another price.
If we can keep up with it, we are always near the center of the distribution.
Are you exactly the prices in the NS?
Elibrarius:
Well it's the prices themselves need to analyze to get the probability of prices. And most seem to indulge in increments, and working only with increments you can only get the probability of increments.
Are you feeding prices into the NS exactly?
For God's sake, shift the center of distribution to zero, and you get increments. There's no difference.
In the NS, yes, prices are relative to rent. I.e. Price is detrend. Plus rationing.
Geez, shift the center of the distribution to zero, and you get increments. There's no difference.
In NS, yes, price is relative to rent. I.e. Price is detrend. Plus rationing.