Machine learning in trading: theory, models, practice and algo-trading - page 1575
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
This is bullshit for sure. There will be no adequacy of sampling at this frequency of transactions
Do you have a different opinion?
Would you consider it a problem to share it?
do you have a different opinion?
If it's a long term strategy, don't you think it's worth sharing?
I am of the opinion that in games with incomplete information and probabilistic outcomes, and hence probabilistic strategies, this very probabilistic nature MUST be taken into account to determine the quality of strategies.
In other words, to determine how close the player's result on the current sample is to his result on the infinite sample, the size of the current sample must be such that the deviation of the result with a probability of say 95% was less than 10% (whatever numbers you want to set). There is no other way to estimate it. This is counted with the help of statistical.methods.And so if the transactions are such long-term, then the forward to the year can not be estimated. Some time ago I was counting, if I need to look for it.
I came to the conclusion that if the strategy is based on stable patterns, then it should work in other markets of similar nature. This way we increase the forward rate at least several times. If a long-term strategy cannot work in other markets, then it is not a long-term strategy, but a fitting for something. Another thing is that this fitting can make a profit for a long time, as luck would have it.
I am of the opinion that in games with incomplete information and probabilistic outcomes, and hence probabilistic strategies, this very probabilistic nature MUST be taken into account to determine the quality of strategies.
In other words, to determine how close the player's result on the current sample is to his result on the infinite sample, the size of the current sample must be such that the deviation of the result with a probability of say 95% was less than 10% (whatever numbers you want to set). There is no other way to estimate it. This is counted with the help of statistical.methods.And so if the transactions are such long-term, then the forward to the year can not be estimated. Some time ago I was counting, if I need to look for it.
I came to the conclusion that if the strategy is based on stable patterns, then it should work in other markets of similar nature. This way we increase the forward rate at least several times. If a long-term strategy cannot work in other markets, then it is not a long-term strategy, but a fitting for something. Another thing is that this fit can make a profit for a long time, as luck would have it.
We have a sample for 13 years, from 01.01.2007 on one TF for all 28 currency pairs
On the basis of this sample according to a strictly defined pattern we make synthetics (combinations of these currencies), which, being observable for some time, after the observation period behave, say, quite a certain way some time later
this "some time" turns out to be quite long, say - up to 10 months
in this case it turns out on the average only 1,5-2 points of entrance per month, or nearly 2 times more, if we take the timeframe 2 times less )) etc.
Of course, you can wait for 2 times 10 months to make a real forward test
more precisely 10 months, since previous 10 months simply weren't sampled due to sampling conditions
then we would actually have 2 periods of 10 months - 1st from "minus 10 months to now" and 2nd "from now to plus 10 months"
Why the question?
Because, it seems to me, even such a forward test will not give an answer to the question about the overtraining of the system
It's not like we adjusted the results for the "last part" of the sample
although it is likely that the market at such a period of 10-12 months may well behave in accordance with some global trends, which can easily change, even after the end of the observation period, say, after Trump's election (just in November of the 20th)
Of course, you can try to make the same random forest find a similar solution, but not the fact that it will find it, since it is still unclear "what to shove into it"?
we have a sample for 13 years, from 01.01.2007 for one TF for all 28 currency pairs
On the basis of this sample according to a strictly definite pattern we make synthetics (combinations of these currencies) which, being observable for some time, after the observation period behave, say, quite definitively after some time
this "some time" turns out to be quite long, say - up to 10 months
In this case, on average, there are only 1.5-2 points of entry per month, or almost 2 times more if you take the TF 2 times less )), etc.
of course it is possible to wait for 2 times 10 months each to make a fair forward test
more precisely 10 months, since 10 months previous to this message simply did not fit into the sample according to the sampling conditions
then we would actually have 2 periods of 10 months - 1st from "minus 10 months to now" and 2nd "from now to plus 10 months"
Why the question?
Because, it seems to me, even such a forward test will not give an answer to the question about the overtraining of the system
It's not like we were fitting the results to the "last part" of the sample
Although it is quite possible that over such a period of 10-12 months the market may behave in accordance with some global trends, that can easily change, even after the end of the observation period, say, after Trump's election (just in November of the 20th).
of course, you can try to get the same random forest to find a similar solution, but not the fact that it will find it, since it is still unclear "what to shove into it"?
It depends on how many free parameters the system has. For example, if there are 2-3 parameters and all signals are independent, then you can do it without any forward, i.e. there will be some statistical confidence. But if there are very few signals, you can not make any conclusions, then you can think about whether there is a real pattern behind these signals, and analyze it.
Forest is just a logger with a large number of parameters, it will not give any answers.
we have a sample for 13 years, from 01.01.2007 for one TF for all 28 currency pairs
On the basis of this sample according to a strictly definite pattern we make synthetics (combinations of these currencies) which, being observable for some time, after the observation period, after some time behave in a certain way
this "some time" turns out to be quite long, say - up to 10 months
In this case it turns out to be on average only 1.5-2 points of entry per month, or almost 2 times more, if you take the TF 2 times less )), etc.
of course it is possible to wait for 2 times 10 months each to make a fair forward test
more precisely 10 months, since 10 months previous to this message simply did not fit into the sample according to the sampling conditions
then we would actually have 2 periods of 10 months - 1st from "minus 10 months to now" and 2nd "from now to plus 10 months"
Why the question?
Because, it seems to me, even such a forward test will not give an answer to the question about the overtraining of the system
It's not like we were fitting the results to the "last part" of the sample
Although it is probable that during a period of 10-12 months the market may behave in accordance with some global trends, that can easily change, even after the end of the observation period, say, after Trump's election (just in November of the 20th).
of course, you can try to get the same random forest to find a similar solution, but not the fact that it will find it, since it is still unclear "what to shove into it"?
Something tells me that this is how you find the fundamental trends. And as you yourself have concluded, yes, it can end in a year or two, or tomorrow. It's business as usual. If your system gives such rare entries, then there's nothing to do but rely on it, having previously (as Maxim has already noted) analyzed and determined the patterns, whether they are contradictory and random.
https://www.youtube.com/watch?v=aX887X3lAc0&t=677s
Useful video and the channel itself, maybe someone will get the message...
Useful video and the channel itself, maybe someone will get the message...
why do we need to advertise courses?
It depends on how many free parameters the system has. For example, if there are 2-3 parameters and all the signals are independent, then you can do without any forward, i.e. there will be some statistical certainty. But if there are very few signals, you can not make any conclusions, then you can think about whether there is a real pattern behind these signals, and analyze it.
The forest is just a memory, with a large number of parameters, it will not give any answers.
Well, in general we can say that yes, there are only 2-3 parameters, one of which is some widely known mathematical quantity ))), the second is the distance from the observation end point (length of a deal - open order) and the third is the value of observation period plus one more parameter but it is constant and does not depend on the first three
the second one is constant and does not depend on the third one, but it may be different at different timeframes
The third one is of variable length (varies in some range "from" and "to")