Activity Spectrum and AFC of MTS using the Moving Average advisor as an example - page 4

 
FION писал(а) >>
Something similar has been realized long ago. Self-training EXPERT (lsv)'.

No, there's a different song there - patterns and training.

We do not train the expert, but allow him to trade only at certain quantum frequencies. This will not make him smarter.

 
jartmailru писал(а) >>

That is, there is some MTS that contains a period dependency. In fact, you are saying that the behaviour of the MTS (trading system) relative to the condition that the period is defined correctly is invariant... That is, if I have defined a period and entered a number into the MTS, then the MTS will work *right* and the same as on past data?

Even if your MTS trades not by signals but by time - for example: trades open every Tuesday at 14:00 - frequency filtering is still possible. Only in one case the system is powerless - the signals are generated by a random number generator, because the repetitiveness of such signals is random.

===

And if you are talking about whether there will be profit from those frequencies, well, "where the money lies" in the future? It is unlikely for all of them, but for most of them, yes. Some previously loss-making frequencies may become profitable, while others may become profitable. Everything is in development.

 
DC2008 >> :

{...} Unlikely on all of them, but on most of them, yes. {...}

Counter-question: on average, how many of the profitable frequencies

remain profitable? And on what time interval on what average size of window were these statistics collected?

The point is that you don't even need to draw any frequencies in general.

Here we are talking about switching to a different type of thinking - not a specific MTS with parameters, but a set of MTS.

And how you number them - by frequencies, indexes - in general it makes no difference.

At some point they give a result - then a run on OOS - then a run on forward (real).

The best ones are chosen, of course. Do you have statistics of such runs on your system at least for a year?

 

jartmailru wrote >>.

  • "...on average how many times how many profitable frequencies continue to be profitable...?" - This is too specific a question. To answer it you need to analyse the specific mts. Spectrum of activity at quantum frequencies is qualitatively the same for most of Mts (i.e. resonance bands are at the same frequencies), but amplitudes are different for all of them. As for AFR, it's like fingerprints: each MTS has its own and unique.
  • "...on what time interval and what average window size were the statistics collected..." - that's a good question! It would seem that the longer the period of historical data analysed, the more statistically likely the result will be. But there is a nuance here: I believe that data before 2006 can not be used. Why? Only lazy EAs will not show great results on this period. Therefore, I take the 2006-2008 period for analysis. And I check the ready Expert Advisor for 2009.
  • "...We're talking about switching to a different type of thinking - not a specific MTS with parameters, but a set of MTS..." - BRAVO!!! That's exactly what needs to be done. Actually this whole quantum analysis thing was designed to create just such an EA, not the other way around. I don't have the statistics as the research is not yet complete. It takes me about 16 hours to study one quantum, while 512 frequencies - that's how much work it is. I'm trying to speed it up, but no breakthrough yet.
 
DC2008 >>:

jartmailru писал(а) >>

  • "...идет речь о переходе к другому типу мышления- не конкретная МТС с параметрами, а набор МТС..." - БРАВО!!! Именно так и нужно делать. Собственно весь этот квантовый анализ был разработан для создания именно такого советника, а не наоборот. Статистики у меня нет, поскольку исследования ещё не завершены. На исследование одного кванта уходит примерно 16ч, а частот 512 - вот и считайте какой объём работ. Пытаюсь ускорить, но прорыва пока нет.

16 hours is a bit long. Do you really collect statistics manually? Here's my data: collecting 3000 results of different combinations of input data (3000 trading systems :-) ), then taking the 20 best ones and running them through "optimization", and then running the 5 best ones through "forward" - it takes exactly one minute! This was me solving the problem of selecting inputs to the probability network. To tell the truth, I have not showed proper imagination neither in a choice of data, nor in a choice of the teacher, therefore I have nothing to boast except speed of calculations ;-)

 
jartmailru писал(а) >>

16 hours is a bit much.

And that's on an i7 965 and 6 simultaneous optimisers.

 
Perhaps you are testing in "all ticks" mode ? I have a tester running "on bar opening".
 
jartmailru писал(а) >>
Maybe you are testing in "all ticks" mode ? My tester works "on bar opening".

I compared the results of the two methods: the time is almost halved, but the quality of the calculation was not satisfactory.

 

Source Expert Advisor Moving Average

Strategy Tester Report
Moving Average
(Build 225)


Symbol EURUSD (Euro vs US Dollar)
Period 1 Minute (M1) 2009.01.02 05:01 - 2009.10.09 11:49 (2009.01.01 - 2009.10.11)
Model All ticks (most accurate method based on all smallest available timeframes)
Parameters Lots=0.1; MaximumRisk=0.02; DecreaseFactor=3; MovingPeriod=12; MovingShift=6;
Bars in history 262484 Modelled ticks 8743398 Simulation quality 25.00%
Chart mismatch errors 0
Initial deposit 1000000.00
Net profit -9178.80 Total profit 24196.20 Total loss -33375.00
Profitability 0.72 Expected payoff -1.58
Absolute drawdown 9210.80 Maximum drawdown 9552.60 (0.95%) Relative drawdown 0.95% (9552.60)
Total trades 5798 Short positions (% win) 2909 (29.87%) Long positions (% win) 2889 (30.18%)
Profitable trades (% of all) 1741 (30.03%) Loss trades (% of all) 4057 (69.97%)
Largest profitable trade 241.00 losing transaction -136.10
Average profitable deal 13.90 losing deal -8.23
Maximum number continuous wins (profit) 8 (74.00) Continuous losses (loss) 27 (-232.00)
Maximum Continuous Profit (number of wins) 241.00 (1) Continuous loss (number of losses) -256.30 (6)
Average continuous winnings 1 Continuous loss 3

Same Expert Advisor, but after applying frequency filter

Strategy Tester Report
Moving Average
(Build 225)


Symbol EURUSD (Euro vs US Dollar)
Period 1 Minute (M1) 2009.01.02 05:01 - 2009.10.09 11:49 (2009.01.01 - 2009.10.11)
Model All ticks (most accurate method based on all smallest available timeframes)
Parameters Lots=0.1; MaximumRisk=0.02; DecreaseFactor=3; MovingPeriod=12; MovingShift=6;
Bars in history 262484 Modelled ticks 8743398 Simulation quality 25.00%
Chart mismatch errors 0
Initial deposit 1000000.00
Net profit 1810.50 Total profit 7835.00 Total loss -6024.50
Profitability 1.30 Expected payoff 10.23
Absolute drawdown 384.00 Maximum drawdown 884.80 (0.09%) Relative drawdown 0.09% (884.80)
Total trades 177 Short positions (% win) 73 (45.21%) Long positions (% win) 104 (54.81%)
Profitable trades (% of all) 90 (50.85%) Loss trades (% of all) 87 (49.15%)
Largest profitable trade 453.80 losing transaction -260.70
Average profitable deal 87.06 losing deal -69.25
Maximum number continuous wins (profit) 7 (663.60) Continuous losses (loss) 5 (-471.90)
Maximum Continuous Profit (number of wins) 759.00 (4) Continuous loss (number of losses) -471.90 (5)
Average continuous winnings 2 Continuous loss 2

 
DC2008 >> :

Initial Moving Average Expert Advisor Same Expert Advisor, but after applying frequency filter

I.e., we took the results of the initial Expert Advisor and used them to build the filter.

And then, knowing a priori the distribution of trades, we obtained the result?

How is it different from gathering statistics by a number of instances?

in a series of successful and unsuccessful trades?

.

Why don't we use a probability network which simply stores

all the times to open - and not run it?

Or you could filter trades with a perceptron from Reshetov's EA :-).

Although if you know the perceptron coefficients - why do you need the original EA :-).

The results might be much better.

.

If you play by the rules - then be kind enough to measure the spectrum.

on the first half of the year, and apply it on the second half.