Machine learning in trading: theory, models, practice and algo-trading - page 1479
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
Well look at the picture I posted on the previous page, there's a hypothesis that crossovers can predict reversals, my thoughts on this subject are there too
I don't know why I'm in such a hurry, it's better to choose optimal and more accurate parameters ((...).
I've seen the picture. The system with crossovers is as old as the world.
I think it will not work. Everything is beautiful only in pictures. I have always worked with MAHs, and continue to do so, and have seen it all. Crossover is not the best way to work with MAshki. But, it's up to the owner.
:)) Lana. This is all nonsense.
What's not nonsense is that you're wrong. There is a certain cyclicity in the market and the selection of the average period (figuratively speaking) is extremely important.
Well, you are going to burn yourself with this selection. The +/- 30% does not affect anything. So, a number of orthogonal MAs is enough and you can do what you want.
Now, think about what orthogonal MAs mean. ))
Well, you'll burn yourself out with this selection. +/- 30% does not affect anything. And, therefore, a number of orthogonal MAs is enough, and do what you want.
Now think, what does orthogonal MAs mean? ))
Well, the integral of the product of which, on some period, = 0. Well, this corresponds to the topic of intersection of MAs... So?
Well, the integral of the product of which, on some period, = 0. Well, this relates to the topic of intersecting MAs... So?
Nothing. With the selection all parameters walk, while with a number of fixed MAs everything is as it should be and their parameters unambiguously determine everything.
That's okay. With the selection you have all the parameters straying, and with a number of fixed MA all stands like a rock and their parameters unambiguously determine everything.
I think, Yuri, having one and the same graphical idea, we are absolutely different in the methods of its realization :)
In my opinion, Yuri, having the same grail idea in hand, we absolutely differed in the methods of its implementation :)
Only I have the implementation of November-December '17. I did not disagree with anyone.))
Only I have the realization of November-December '17. I didn't split up with anyone or anywhere.))
I'd be interested to hear what role your neural network plays in TC. That's it! And so... Not interesting.
I'd be interested to hear what role your neural network plays in TC. Now that's a yes! But otherwise... Not interesting.
It is used as a teachable programmable logic matrix (PLM), I wrote about it some time ago.
And interesting or not - there is no information anyway).
Well, it's a classic of the genre, I already wrote above about predicting the optimums of the TS properties and results (equitability\pnl...).
If we are going straightforward, the principle is the same as with returnees or a wave, for each sample we divide the sample into "before" and "after", by some moving point of price(t), calculate any chips on {price(t-N),price(t)} and target {price(t+1),price(t+K)}, and run t through the whole series. In this case, the targets will be the waving optimums on {price(t+1),price(t+K)} on some window in the future, and the features can be basically anything from Stochastics or momentums of different periods, to waving optimums or other TS on the previous period{price(t-N),price(t)}.
But it doesn't make much sense.
No. What I'm saying is that there's no need to adjust anything. The choice of MA period has little effect on the strategy. The strategy can take any periods within a wide enough range, and it will not affect anything. I.e., if you chose MA 12 and 16 or 10 and 14, it makes no difference for the strategy itself, everything will stay in its place and the entry parameters will be different, but it's like measuring with an inch or a centimeter ruler - the figures are different, but the result is equivalent.
If you want to use a large range of frequencies, then you need several MAs with a fixed period, and they will cover the entire range, and again you don't need to adjust the parameters.
For other indicators, it's the same.
Exactly so, the output of indicator strategies depends not on the period of the indicator, but on the phase of the market - trend or flat, the range of effective values is wide, but to find out when the trend starts/ends - flat is a challange, then dumb to change the impulse trading and reverse trading, and only the frequency of trading depends on the parameters of smoothing.
But it doesn't make much sense.
This is exactly right, the output of indicator strategies depends not on the period of the indicator, but on the phase of the market, trend or flat, the range of effective values is wide, but to find out when the trend/flit begins/ends is a challange, then stupidly change impulse trading and reverse trading, and only the frequency of trading depends on the parameters of smoothing.
My experiments with scaffolding show the same situation.
Training for 6 months 2017, tes for Oct-Dec 2017 - err on the test/forward 42%. If you go by the valking forward method, when you shift to 2018 (Training for 6 months 2018, tes for Oct-Dec 2018) the error is 55% and a quick drain.
If you look at the annual chart, in 2017 there was a global rise with pullbacks up to 40 days, both on the tray and on the test. And in 2018 some more growth and the beginning of a global decline.
So we can conclude that if the 2 month pullback hasn't stopped, it's a new reverse trend. Les did not understand this and for a third of the learning curve, while there was still growth - trained on this growth, and therefore lost on the test.