Machine learning in trading: theory, models, practice and algo-trading - page 821
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
A classic example of using ML in algotrading is using returns with exponential window from the past as a feature and return(s) from the future as a target(s). Fiddling with "tricky" indicators makes no sense, only returns of other price series or stationary series of other data adds quality.
By the way, you don't have to take returnees, you can use any transformations of the price or time series bundles synchronized with the forecasted price, however it depends on the accuracy of future returnee forecast. How it depends, try to deduce it yourself, it's an interesting intellectual task, it's not in the public domain.
What is VERY IMPORTANT - chips and returnees should be taken from non-intersecting data, i.e. we should just cut the row window into two rows, which will be used to make chips and which targets, and then work with them separately as you want, cut initial prices (volumes, etc.), NOT INDICATORS, for example ZZ looks ahead on an undefined window, at any point it mixes the future and the past in a non-linear way, and then predicted not the future, but the past is mixed into the future, so such a high not real accuracy, cut ZZ, from the past contains not clearly future, this has been said many times in this forum and I and toxic, andrey and wizard and J.B, if my memory serves me right, but all to one side, which is probably for the best.
Target should not see the past, you can very simply check if the target is not looking into the past, run your set-up on random walk, if everything is correct then on SB should be accurate 50+-0.3% on hundred thousand samples, on million +-0.1%, if on SB you still have the same 70-90% or even stable >51% well you understand yourself....
And on ZZ as a target it is easy to get 90% accuracy on SB, like SB you can predict))))
PS : <75% down the drain... Don't make me laugh)))
can you elaborate on the exponential window? are the increments taken at exponential time intervals? erlang flow? or what. Or better yet, a code example. I keep hearing about it, but I still don't understand the meaning.
Also about returnees - does it make sense to use a p-order autoregression with an "exponential window" instead?returns with windows e~2 -> 1,2,4,8,16,32, 64.... 10 pieces, some take more meaningful 1,2,5,10,20,30,60,120,300,.... But it is not essential, at the expense of regression and other filters, it is also considered that it does not affect markedly, it's a matter of taste, the main thing that the data were synchronized and did not wander in relation to each other when there are a lot of series (and there always are), otherwise everything breaks, so you have to write them yourself, when there are a lot of sources, for example returns from the past {-1,-2,-4,-8,-16,-32,-64,-128,-256,-512} and target ruturns from the future {+1,+2,+4} as a "stylized" example, more volatility changes there, there are also changes in volatility, delta of cup, OI, etc. BUT the MAIN thing is to use features with targets NOT TRANSFERABLE DATA, otherwise it's a fake, which can be easily improvised by any peeping turkey like ZZ
Ah, I see, just a lot of returns with different windows exponentially. I did so, it didn't improve much :) but as a target I took a returnee with a different period, not included in the fiches
There is also a variant to sift the window, remove every 2nd element, etc., like a transformation to the Markovianity by Alexander k2 :)
You're lucky, Max. This thread is slowly gathering a bunch of old timers. With grains of knowledge, which also need to be "sifted" exponentially:)))
Yeah, a lot of people have been here a long time... it seems like from the beginning of forex :) Such petrified mammoths and dinosaurs who can't be moved by anything))
It remains to find a Hero (where is Misha?), who will take the price tick series and start smashing it:
1. to Erlang's 1st order flow (price input)
2. Second order (first input - price, second - returnee).
2- Third order (1 - price, 2 - return, 3 - return of returns)
No more.
Record the results.
Then the best one will be improved with the active participation of well-wishers.
And it is easy to get 90% accuracy on ZZ as a target ...
Can you list the predictors that would give decent accuracy on ZZ WITHOUT overtraining?
You're grown men, but you're talking about nothing!
Don't you understand that without specifying the method of estimation followed by an example, which necessarily includes this method of estimation, with a proof that the model is not retrained, all "blah, blah in the yard of poplars", hundreds of pages....
In a week I will post BP already reduced to the Erlang flow of order 1. But I'm unlikely to do 2 and 3 - I don't need it.
Alexander, do you want your idea to be pumped by neuronics?
Well, there is freelancing, grail hopefully helps the wood.
Alexander, do you want your idea to be pumped by a neuron?
Well, there's a market, a grail hopefully helps the wooden ones.
YES! I'm thirsty!
Now I need to tap these rows with neuron, and then only give them to coders.
I don't have the time right now. I'm telling you, our "partners" in Germany are forcing me to do a project while I've abandoned the Grail. There it is - in a corner...