Statistics, optimisation and "lucky coin" ....
CONCLUSION: If getting an acceptable trading result is a consequence of a simple coincidence (obtained by flipping a coin, using indicators, working with price, using algorithms used in radar (and there are some), etc.), then it turns out that the price by itself or in the form of some functions expressed in the indicators for generating trading signals does not differ from a simple coin flip. Since it is clear that there is no "lucky" coin (here I am appealing to probability theory experts), so there is no "lucky" or "long-lasting" GTS based on price / price function. This means that any optimization of the TS or statistics used to evaluate this TS is meaningless, because at any time, as a "lucky" coin will stop showing "positive" results, and the TS built on a simple coincidence will stop "working", the only question is time....
The second paradoxical conclusion: if VTBs showing good results on the history in the tester/real, do not guarantee maintaining such results in the future, why can't those VTBs that have poor results on the history perform well in the future? So, it turns out that there is no difference between selecting a "co-optimised" and "non-optimised" GTS..... in real trading Either way, it's still a guessing game: "lucky or unlucky" ....
Conclusion: given all of the above, I am not specifically raising the question : 'how to live on? I am interested in the goals of optimisation and statistics, of those who are actively doing it...
Both conclusions are absolutely correct. Optimization of TS and statistics of results of such TS - meaningless. The answer is "those who have been actively engaged in it" during the last years, applying, in particular, the method described by R. Pardo in his fundamental work on optimization and testing.
...... The answer is "those who have been actively doing so" over the last few years, applying in particular the method described by R. Pardo in his seminal work on optimisation and testing.
Pardo wrote in his book what conditions/results of the TS should satisfy the trader (it may be exaggerated, but nevertheless ....). My idea is that any results of TS can be obtained, even the most "grail", the only question is the number of takes necessary for this. The problem is that the "graality" of the results does not increase confidence in the TS itself.
A good TS has some peculiarities.
Both conclusions are absolutely correct. Optimization of TS and statistics of results of such TS are meaningless.
Both conclusions are absolutely delusional. Optimisation and random coin flip statistics are meaningless (although some autumn and spring personalities don't think so, hehe), but what about trading systems that use patterns, i.e. the opposite of randomness?
To the topic-starter. If a TS with a stable positive expected payoff is found, the purpose of optimization is to find the optimal set of parameters (and not to find the only profitable set for some abstract coin on a certain piece of history). The purpose of a set of statistics is to find the system itself, which is worth optimizing. In the Trade Idea - Statistics Set (idea verification) - Finding Optimal Parameters cycle, you release the most essential part, the first. The second and third make sense only if it is available.
Topikstarter. If a TS with stable positive expected payoff is found, the goal of optimization is to find the optimal set of parameters (not to find the only profitable set for some abstract coin of a coin on a given piece of history). The purpose of a set of statistics is to find the system itself, which is worth optimizing. In the Trade Idea - Statistics Set (idea verification) - Finding Optimal Parameters cycle, you release the most essential part, the first. The second and third make sense only if it is available.
What is a trading idea? The topikaster's trading idea is to open positions with a coin.
Congratulations to him.
thank you
You have to think hard about whether a random number generator or a TA indicator is more "scientific".
A good TC has some features.
Like what?
Well, first of all, it doesn't open positions like the topicstarter's from scratch,
and second of all, it doesn't have a grail.
Since the topic was about R. Pardo and his book "Developing... ".
After reading it, I still don't understand the Criterion for selecting the "perfect" OPTIMISATION MODEL? What is it? In each individual period (approach to) optimising the values of the variables.
Is it really just to beat everything by the average and that's it...
I would be grateful for clarification of my question.
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
You agree to website policy and terms of use
On the eve of the weekend.... The topic is inspired by the reasoning in neighbouring threads about the value of TC optimisation, statistical performance of TC, etc. I'm not trying to address any practical aspects, I'm more interested in the conceptual approach.
So, thesis #1: There is a TC: TP - X pips, SL - Y pips; entry - at the close of the last trade (i.e. at the TP or SL); entry direction - determined by flipping a coin. If we have 1000 physical pennies, we can assume 100 pennies will show acceptable results on history, 10 of them will show acceptable results on demo account in real time, 1 penny of them will show acceptable results on real cent account (... after that, proud of the obtained result, we open a standard account, filling it to XXXXX dollars, and recruit investors through PAMM-accounts, showing them the statistics of previous transactions ...; Question 1 (rhetorical): how this story ends?) - Note: if mathematicians suddenly have objections to the possibility of getting acceptable results from 1000 coins, you can take 20 000 coins (the question is not about the numbers, the idea itself is discussed) ....
Since it is clear that the coin, in this case, acts as a random number generator (RNG), then instead of 1000 (or 20 000) coins, you can insert this RNG into Expert Advisor (I'm not considering software capabilities of MQL to accomplish this task, I consider the idea itself ...), and after 1000 (or 20 000) runs, we will get "rough" tester result ... I.e. my thought: with a large enough number of attempts at a test, forward, etc., it is always possible to REMEMBER results that satisfy the most demanding statistical lovers of Trading Systems....
Thesis #2: there are 100 different MT-4 indicators. With all possible combinations of two, three, four, five different indicators with different customizable parameters, we obtain 1 000 (but if we want, taking indicators of these same combinations from other symbols and/or TF, we can obtain 20 000) trading signal generators (GTS). We use the above-mentioned TS, but instead of a coin, the direction to enter Buy or Sell will be shown to us by the STS. If we have 1000 (or 20 000) STS, we can assume that 100 STS will show us acceptable results on history, of them ..... (further, to save space - similar to the above...)
Question ¹2 (rhetorical): What is the principle difference between the combination of indicators as GTS, and coins as HGS...? - I foresee the answer in advance: the coin is pure "case" and the indicators are price functions, which...., therefore..... Consider Thesis #3...
Thesis #3: there is the same TS... This time the trading signals are generated as follows: Option 1: if the close of the last bar is above its opening - BAY, if the close below the opening - SELL; Option 2: if the close of the last bar is above its opening - SELL, if the close below the opening - BAY... We can suppose that the tester will show better results for Variant 1 and for Variant 2, but in general, both results will be quite sluggish "on aggregate of instruments and timeframes"... Now, we add Variant 3: to the ratio of the last bar's close/open we add the close/open of the penultimate bar, Variant 4: to the close/open of the last two bars we add the high position of these two bars,.... Variant 97: we consider Variant 4, but we take data from the "neighboring" instrument, etc..... Thus, we get 1,000 (or 20,000) possible entries. If we have 1000 (or 20,000) choices of inputs available, we can assume..... (I hope my thought
understandable...).... So, Question #3: if in one case, prices generate a rather weak signal, but in the other case, a very good signal (I insist that out of a large (!!!) set of different signals, at least one "grail" result will necessarily be obtained), then what is it: discovery of a fundamental market pattern or a simple coincidence?
CONCLUSION: If obtaining an acceptable trading result is the consequence of a simple coincidence (obtained by flipping a coin, using indicators, working with the price, using algorithms used in radar (and there is such a thing), etc.), then it turns out that the price itself and some functions expressed in the indicators, in order to generate trading signals, do not differ from a simple flip of a coin. Since it is clear that there is no "lucky" coin (here I am appealing to probability theory experts), so there is no "lucky" or "long-lasting" GTS based on price / price function. This means that any optimization of TS or statistics used for the evaluation of this TS makes no sense, because at any time, as a "lucky" coin will stop showing "positive" results, and the TS built on a simple coincidence, will stop "working", the only question is about time....
The second paradoxical conclusion: if TCPs showing good results on the history in the tester/real, do not guarantee maintaining such results in the future, why can't those TCPs showing bad results on the history perform well in the future? So, it turns out that there is no difference between selecting a "co-optimised" and "non-optimised" GTS..... in real trading Either way, it's still a guessing game: "lucky or unlucky" ....
Conclusion: given all of the above, I am not specifically raising the question : 'how to live on? I am interested in the goals of optimisation and statistics, of those who are actively engaged in it...