Machine learning in trading: theory, models, practice and algo-trading - page 1687
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
When smart people talk to each other, and you can only understand that you don't understand a thing.
And you want to say something clever in response. You want to say something intelligent in response...
But nothing but YA KREVEDKO comes out. Ehhh.
+++
There is some intuitive feeling that this model can only give something between SB and flat.
To describe trends we need some other game model. Perhaps we should ask Savvateev.)
In any case, it is unlikely to obtain a model that gives both trends and fluxes and transitions between them (as it always is in reality).
No, my task is not to build TS, but to look for a technique to evaluate TS
We need to find a fine line where TS works, and where it does not
As I wrote above - stat.indices from the Strategy Tester are tied to the beginning of observations and the total number of observations, incidentally, like all statistics
and want a technique that could evaluate the work of the TS in the future
ZZY: probably everything is simpler, I remembered about the error function, may be it is enough to estimate the divergence of profits-loss series between the test and forward-test and / or balance areas (equities)
ZS: probably everything is easier, I remembered about the error function, maybe it is enough to estimate divergence of profits-loss series between a test and a forward test and / or balance plots (equities)
The task from the opposite direction, not by external methods to determine the flat trend, but by the Expert Advisor on the history, good, not very good, not very bad, bad, watch the parameters of the series. There are a lot of them, too many, we must find the significant ones. I do not like the idea of working with simple algorithms to process certain parts of the series, that fit these algorithms, but the logic is working. All that remains is to identify the segments and determine the parameters that should be used to determine them. Parameters should take into account all that we can calculate for the series, ticks, averages, speeds, accelerations, volumes, volume increase rates, thinning and maybe something else. That's kind of a separate job, and what row parameters can be taken into account. MO and GA are there to help in identifying the significant ones. But the task is precisely in determining the significant parameters of the initially tick series. Which parameters are sensitive and correlate correctly with changes in the TC indices. Averaging of series thinning... there is no way to do without them today.
They told at the nuclear power plant at school, the operator monitors 19 parameters and he is trained to lead a stable state corridor, depending on the correlation of the parameters, selecting significant ones and changing them.
We have long enough tick series on the final number of instruments, the field of analysis is big enough, not sufficient for stable forecasts maybe, but I do not see another way.
In general the goal is to find parameters that correctly characterize the state of series. By a state we mean growing, not growing, these are stable states and starting to grow, finishing, not stable states. )))))
No, my task is not to build TS, and I am looking for a methodology for evaluating TS
we need to find a fine line where TS works, and where it no longer works
As I wrote above - stat.indices from the Strategy Tester are tied to the beginning of observations and the total number of observations, however, like all statistics
and want a technique that could evaluate the work of the TS in the future
SZZY: probably everything is simpler, I remembered about the error function, may be enough to estimate the divergence of series profit-loss between the test and forward-test and / or balance areas (equities)
Evolve something like a neural network(some phenotype of scales) this way -
the whole section of the sample, let it be three months, is divided for calculation into three sections - one month each
the phenotype with the minimal divergence of profit-loss series between three parts is the winner/leader and the worst series in a month is better, as in your suggestion.
This approach produces with enviable frequency rabast solutions on the forward test. I.e., the forward test is better than the worst series in one of the three areas of the sample.
SZY: probably everything is simpler, I remembered about the error function, maybe it is enough to estimate the divergence of series profit-loss between the test and forward-test and / or areas of balance (equity)
I was telling you that for two pages of the forum, and you were evasive... And now, all of a sudden, he remembers the error function !))
It is sufficient to evaluate divergence of profit-loss series between a test and a forward test and/or balance (equity) areas
Not the best variables to analyze, too much lag. When you see that the forward and the test have diverged, it is too late.
hmm... hard, I will try again, but once again: there is no task to search for TS, there is no task to determine the trend-flat
1. there is a set of strategies which showed good results at the test
2. there is a subset of this set of strategies which showed good results on the forward
3. there is a statistical estimation of the strategy tester
what is the difference between pp. between pp. 1 and 2 ?
is it possible to analyze items 3 and find differences between items 1 and 2 ?
how to evaluate pp1 and pp2 from the point of view of .... what the hell is the difference between them ? - how do they differ?
hmm... hard, I will try again, but once again: there is no task to search for TS, no task to determine the trend-flat
1. there is a set of strategies which showed good results at the test
2. there is a subset of this set of strategies which showed good results on the forward
3. there is a statistical estimation of the strategy tester
what is the difference between pp. between pp. 1 and 2 ?
is it possible to analyze items 3 and find differences between items 1 and 2 ?
how to evaluate pp1 and pp2 from the point of view of .... what the hell is the difference between them ? - how do they differ?
In my opinion - to analyze pp.3 and find the difference between pp.1 and pp.2 with this approach is a waste of time, because most likely you will be comparing, although in pp.3, incomparable, in a lot of strategies are from "flies" to "elephants". And flies and elephants still successfully live statistically, but the sense of comparing them?
hmm... hard, I will try again, but once again: there is no task to search for TS, there is no task to determine the trend-flat
1. there is a set of strategies which showed good results at the test
2. there is a subset of this set of strategies which showed good results on the forward
3. there is a statistical estimation of the strategy tester
what is the difference between pp. between pp. 1 and 2 ?
is it possible to analyze items 3 and find differences between items 1 and 2 ?
how to evaluate pp1 and pp2 from the point of view of .... what the hell is the difference between them ? - how do they differ?
If not, then pp1 is no different from pp2.
hmm... hard, I'll try again, but once again: there is no task to search for TS, no task to determine the trend-flat
1. there is a set of strategies which showed good results at the test
2. there is a subset of this set of strategies which showed good results on the forward
3. there is a statistical estimation of the strategy tester
what is the difference between pp. between pp. 1 and 2 ?
is it possible to analyze items 3 and find differences between items 1 and 2 ?
how to evaluate pp1 and pp2 from the point of view of .... what the hell is the difference between them ? - how do they differ?
Now the question is more clear. But now the answer is not clear at all) And it's even not clear why there should be such an answer for a completely arbitrary set of systems.
The problem is that any superprocedure with tests, forwards and other horseplay will eventually boil down to the usual optimization of some complicated system by some complex criterion and in a set of parameters of large dimensionality on the same limited history.OK, some of the answers already give at least something in the direction of the search for information
but the problem is as always in the question - a well posed question is 50% of the answer ;)
---------------------------
to the first part of my question - I will add more terms of search for TC:
TCs are all the same - literally
The principle of constructing a TS: we set the rules and the strategy tester generates the result by GA oversampling. The rules are simple - open an order, place a pending order.... place an order relative to a previous order ..... repeat placing an order at the same level ...... and other similar primitive rules for working with orders - only your imagination limits here )))
But no matter how you think about it, these are the same TS with an active rule of working with an order or this rule is disabled, the total number of simultaneously opened orders is 1-5.... in general, very primitive and in fact this is how all TS work, if no neat terms are used
well, we add an arbitrary indicator to these rules for testing - we obtain a lot of TS which satisfy the optimization parameter
well, back to the same question .... why part of the TS may pass forward testing, but some of the TS does not pass - we need a comprehensive assessment of statistical indicators, though.... I suspect that now we will have to unload statistics for each TS and on the basis of statistical indicators make a choice of TS and optimization - and see what will come out - i.e. another "rule of thumb")))