Machine learning in trading: theory, models, practice and algo-trading - page 1497
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
Please try to play with the indicator I told you about, I see you clearly know more than me
As for data thinning/pre-processing, I stand by my opinion: it is necessary. But, let it be a matter of discussion.
I thinned M1 according to the formula of raising bar number to a degree. If degree =1, no thinning, if =2, then by parabola.
If I thin by parabola, then 8 bars will remain out of 50. Bars 0, 1, 4, 9,16,25,36,49. Something similar (sampling by bar numbers) was described by the creator of this discussion in his blog.
But I'll still experiment...Tried different degrees from 1 to 2. For example one of variants at degree=1.6 gives good result, at 1.7 not so much, at 1.8 - bad. In my opinion too sharp peaks and the system is unstable.
In my opinion there is one more parameter for overfitting.
Alexander, how does your thinning differ from the thinning of a bar removed from the current point along a parabola, for example?
Or from the zigzag thinning as described above?
Is there an indicator or formula for obtaining bar numbers?
We are testing slowly. It was not bad predicted growth on the Eura. The system cautiously bypassed the flat and entered just before the start of a powerful upward momentum.
Try tick bars or volume bars
i.e. price channel on ticks. Why exactly in ticks? Because in the neighboring topic I showed screenshots that there is no information in the closing prices, i.e. the graph is not different from SB
or bars by the number of ticks but not by their amplitude
Another advantage is getting rid of undersampling when the sample is unbalanced at standard bars. For example, sharp movements occur in 1 bar, and flat shifts may last for tens. As a result the strong movements appear as spikes. If you build by one of the suggested methods, then strong movements (i.e. abs changes in price) will be given more value, i.e. there will be more samples. Because of that the distribution of increments will be more like iid (normal). Basically, it is similar to a zigzag, only from a different bell tower and, most importantly, by ticks and not by claws.
If you start from these new bars, you can then play with thinning. I.e. people knew about it long time ago and about thinning and all that stuff, and it kind of makes sense. If I look at codes of fxsaber, there seems to be the similar idea, for example the price-channel on ticks. I look at codes of fxsaber - it seems that the analogy is obvious. That is, everyone is doing the same thing and talking about the same thing, but in different languages.
If that won't help, there are still some Markov chains to spin and meditate on them, if nothing at all will help, then you can calmly powder your head with ashes and go smoke bamboo
If it helps, then we'll have an advantage bordering on spread, i.e. we can kill such a stratum with spread widening. That's what Doc wrote about.Try tick bars or volume bars
i.e. price channel on ticks. Why exactly in ticks? Because in the neighboring topic showed screenshots that there is no information in the closing prices, i.e. the graph does not differ from SB
or bars by the number of ticks but not by their amplitude
Another advantage is getting rid of undersampling when the sample is unbalanced at standard bars. For example, sharp movements occur in 1 bar, and flat shifts may last for tens. As a result the strong movements appear as spikes. If you build by one of the suggested methods then strong movements (i.e. abs changes in price) will be given more value, i.e. there will be more samples. Because of that the distribution of increments will be more like iid (normal)
On the basis of these new bars you can then play with thinning. That is, people have known about it long time ago and about thinning and all that stuff and it kind of makes sense. If I look at codes of fxsaber it seems to be the same idea. I would not call it so, but it seems that the analogy is obvious.
If it doesn't help, then I'll use Markov chains, and if nothing helps, then I can safely powder my head with ashes and go smoke bamboo.
The calculations are slow enough as it is...
If we switch to ticks, I'm afraid the process will slow down many times, just by switching the tester from the opening prices to real ticks.
In addition, I always see messages about corrected quotes.
By the way, I already do all the tests in channels between High and Low. Close and Open are random values between them. As an option, you can test in the middle between H and L, to reduce the dimensionality of inputs.
We are testing slowly. It was a good prediction of growth on the Eura.
What exactly is being tested? What kind of model?
And on what predictors?
The calculations are slow enough as it is...
If you switch to ticks, I'm afraid the process will slow down many times, just by switching the tester from the open prices to real ticks.
In addition, there are constantly messages on the subject of editing quotes.
By the way, I already do all the tests in channels between High and Low. Close and Open are random values between them. As an option, you can test in the middle between H and L, to reduce the dimensionality of the inputs.
Well, using opening prices is like cutting off your legs and trying to run a marathon from this position
any variations on the theme are a work with SB, with all that followsNot yet, I'm into theory
because in RL it works a little differently
Can you please tell me what you are reading?
Andreev, Ioffe "These Remarkable Circuits" - introductory science pop
Kelbert, Sukhov "Probability and Statistics in Examples and Problems" Volume 2. Markov chains as a starting point for the theory of random processes and their applications
Thinned out M1 using the formula for raising the bar number to a degree. If degree =1, no thinning, if =2, then by parabola.
If we thin by parabola, then 8 of the 50 bars remain. Bars 0, 1, 4, 9,16,25,36,49. Something similar (sampling by bar numbers) was described by the creator of this discussion in his blog.
But I will still experiment...Tried different degrees from 1 to 2. For example one of variants at degree=1.6 gives good result, at 1.7 not so much, at 1.8 - bad. In my opinion too sharp peaks and the system is unstable.
I think I got one more parameter to retune.
Alexander, how does your thinning differ from the thinning of the bar removal from the current point, for example along the parabola?
Or from the zigzag thinning as described above?
Is there an indicator or formula for obtaining bar numbers?
Ticks and methods of thinning them are too vast and conceptual topic to tell everything in one post.
I can only say that Maxim, stating that on OPEN/CLOSE M1, M5,... we, in fact, have SB, and the emissions have nothing to do with memory - is not far from the truth.
For, in this case, we lose the most valuable information - the time intervals between quotes, which are uneven and form a Pascal distribution. This fact indicates that the quotes flow itself has a certain consequence, i.e. it is a periodic process in time.
So, it can be stated that events in tick VR have some periods and the trader's task is to find them, to calculate. It is more clear with the thinning.
If we define these time cycles, we find out that this is the process memory, that time is the only parameter that distinguishes real BP from SB.