Yoghurt systems and canned systems or The relationship between trading tactics and the reliability of historical test results - page 7
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
I would like to propose a discussion of the relationship between trading tactics and the reliability of historical testing results.
When we create a trading system, we rely on some regularities that depend on the input variables. We often optimize the system according to these parameters at the testing stage and rejoice when we get good results. And then it turns out that the system starts to fail after a week. Some systems only start to fail after a few months. I also suggest discussing factors that affect the "shelf life" of systems. Why some systems are yoghurt, others canned tuna, and is it possible to create a perpetuum mobile or at least come close to it?
First we have to decide which systems we are talking about. I assume that we are talking about systems operating on the same timeframe and having approximately the same frequency of transactions. In this case "the shelf life" will depend on how much the pattern used in the strategy corresponds to the real price change pattern.
By the fact that if you take the statistics for an infinite number of years ago, or at least for 10 years, I think that your TS is unlikely to work well because the market rules (candle body, scope, volatility) have changed several times and the data for such a period is unlikely to be relevant today. But if you take these statistics for the past 2-3 months or a year (all depends on the TF in which the TS works), then I think the TS will work well, because these data are relevant to today's market. On the other hand, if you reduce this period to a few bars, then the TS will also work bad, because the data will not be enough for a complete picture. That's why your TS has a parameter - the period for which these statistics are collected. This parameter is very important for the TS. And it is impossible to say that your TS will work equally well with values of this parameter from - bezcon to + bezcon. That's actually what we were talking about.
Firstly there is no relevant and not relevant data (my opinion), if your data for 3 years ago is not relevant it's likely to fit the current situation.
The system works on D1.
The period for which the statistics is gathered, this is the optimization. The period so to speak, the adjustment (for example) 1990 - 1995, the run on the found patterns in 1995 - 2000, etc. Not a single forward makes it clear what such a system is worth. And it works the same way whether in 1990 or in 2008.
LeoV Are you dealing with NS?, as far as fitting there is much more complicated than in conventional TS... ... If you're a BS trader and you can't do it with usual forwards, look at D.Katz, D.McCormick Encyclopedia of Trading Strategies - there is something interesting about this problem ...
First of all there is no data that is current and not current (my opinion), if according to you the data from 3 years ago is not current, it will most likely be a fit for the current situation.
The system works on D1.
The period for which the statistics is gathered, this is the optimization. The period of the so-called adjustment (for example) 1990 - 1995, the run on the found patterns in 1995 - 2000, etc. Not a single forward makes it clear what such a system is worth. And it works the same way whether in 1990 or in 2008.
LeoV Are you dealing with NS?, as far as fitting there is much more complicated than in conventional TS... ... and you can't get away with ordinary forwards, look at the book by D. Katz, D. McCormick Encyclopedia of Trading Strategies - there's something interesting about this problem ...
Well, I wrote - it all depends on TF you're working on. For example. 3 years for a minute - is a lot, and 3 years for D1 - is very normal timeframe.
I meant that a good TS should work with any parameter - bezcon to + bezcon (in your case - a period of statistics), that the optimization of this parameter is unnecessary. I wrote that TC cannot work equally well with values of this parameter from - bezcon to + bezcon.
I haven't seen TC, and even more so on a neural network, which works equally well in 1990 and in 2008. Not even on daytrips.
Yes, doing the NS.
Testing on quote history is not the only method of testing system reliability. Here is an outline of an alternative testing methodology: An article on tossing a sandwich. But if you're allergic to mathematics, it's best not to read it.
Well, I wrote - it all depends on the TF you're working on. For example. 3 years for 1 minute is a long time, and 3 years for D1 is a very normal timeframe.
I meant that a good TS should work with any parameter (in your case - the period of statistics), that the optimization of this parameter is unnecessary.
I haven't seen a TS, much less on a neural network, which works equally well in 1990 and in 2008. Not even on daytrips.
Yes, I do the NS.
You know in my case is not so much a choice of optimization parameter (sampling interval), but through many intervals to choose a stable pattern. And the more sample is better for statistics...
I can't agree with that. No way )))
I can't agree with that. No way )))
1. Pichimyu? )))
2. What kind of sample would be sufficient for you? In terms of number of trades, in terms of testing time.
Testing on quote history is not the only method of testing system reliability. Here is an outline of an alternative testing methodology: An article on tossing a sandwich. But if you are allergic to mathematics, best not to read it.
Once again, a slight substitution of concepts. Statistics of a system with a limited number of parameters and searching for patterns of price movement when it seems that there are not many parameters.
And I personally am not allergic to mathematics (the first paragraph of the article). :)
It would seem - run the script on minutes from the times of "king Gorokha", unload "everything you can", get a terrrrabyt file and ... find a pattern. And they will be found - "it depends on the theory, but the facts will fit" (c).
And with the "window" so with OOS too - immediately after the end of the "window"/OOS ... we will have to start collecting new statistics.
`
It's only apologists of "indicatorless systems" who think that they don't use "indicators". In fact, there are as many "optimisable parameters" as in any other system, and, consequently, the graph of "Profit as a function of MA_Period" turns into ... A multidimensional blot, which "our statistics" can't get close to. And judging by the fact that I haven't heard "Statisticians" :) conquerors of Forex, and "not with our statistics" either. :(
But I don't mind being called "allergic" as well :)
`
SZY. Although in this thread, too, jumping from "search for patterns" to "evaluation of the Trading System".
SZY. Vot that TS with a probability of 95% sell I know, I know that the price at least "go" (of course in a limited time period) with a probability xx% do not believe. ;)
I can't agree with that. No way )))
I agree. The more the better for processes with few or no process-impacting inputs, but not for markets with macroeconomic inputs.
The more inputs, the better.
"The search for patterns" is an eternal and inexhaustible topic, for which no answer will ever be found (in the perpetuum mobile sense). And testing and evaluation of TC is quite a practical problem that can be solved for a given TC, even if the allegedly found patterns are not logically understood or completely unknown. It seemed to me that the topicstarter's question was asked precisely about how to adequately assess the behaviour of the TS in the future...
The search for patterns is pointless if we cannot justify with any degree of reliability that the TS built on the found pattern will behave acceptably in the future.