Signs of a REAL system - page 19
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
And the eighth attribute is very reasonable, by the way. The number of transactions from which one draws conclusions about the stability of a system must be statistically representative.
But this one parameter is not enough. It must be combined with the profitability of the system shown in the test. Roughly speaking, it can be as follows: there are 1000 trades and the profitability of 1.03 in all other "correct" conditions, which is not enough to consider the system as steadily profitable. However only 100 trades at profitability 4 is a good enough base (given all other "correct" conditions).
Where should we put it - probably, into logical errors. By the way, this error is rather frequent here, even too frequent. The programmer displays a test that has a good profitability (say, 3.5), but with 15 trades, and asks to evaluate the system. How can you evaluate it?
Unfortunately, for the most part, all of these are just beautiful words that give you a reason to philosophise, rather than cutting out the wrong blocks from the code:
what do you mean "sensible"?! for someone, the idea that there should be a maximum of indicators on a chart can be very sensible, because this is the only way to see "all" signals, to find their confirmation among each other. and this, in general, is not without its logic ;)
again, what do you mean "random" and what do you mean "provisions"?! is that: there should be a rule that the ADX should always be in the topmost sub-window and the RSI should be placed strictly below it, second?!! nonsense of some kind....The discrepancy: the first part of the sentence says about a cluster, then it should obviously be that it cannot be changed to another one or trade on the other ones, if it doesn't give you a signal. But then comes the phrase completely unrelated to the first one
we had that too :)
great! i decide to use this rule! and..... ???? what are the limits "reasonable"???
a completely useless rule for practical application
wellooooooo.... let's say :( ... added to Logical Errors of Algorithms
I would also add this "and do not exceed the number of transactions performed by a single trade object for the whole array of statistically reliable data present in the sample with the probability determined by the ultimate trading performance of previous periods...." - just like that :))))))))))))))))
In short - useless rubbish, which does not give a guide to action: how to check what is wrong in your system, what you need to discard/redo.
All signs are subjective. There are no finite formulas to calculate the probability of fitting, it's all subjective/expert judgement. Like in your first post the word "radical" is very subjective.
Imha, the main criterion for robustness is statistical validity. This is mainly the number of trades, but there are features of systems that require more or less trades for the same level of certainty. For example, the system has no stop or the take is much smaller than the stop. This is overhosting and it has been written that it is bad. But it is not over siting that is bad, it requires more trades for statistical validity. If for example take 10 points and stop 100 points, then 100 trades will not be enough for statistics because loss will be too small for validity. I.e. system statistics must be such, that both losses and profits would be enough for stat validity.
The same concerns multicurrency and multiframes - these conditions are not necessary, they simply increase statistical reliability proportionally to the increase of the number of deals.
The same applies to the sensitivity to the indicator parameter. The wider the optimal zone, the greater the statistical validity. In fact each set of options can be considered an independent system the profitability of which confirms the profitability and robustness of the idea in general.
And also I still have not ignored the fact that the system must have logic. It is not an empty phrase. If someone uses the indicator and good results are obtained with logical time parameters equal to the length of trading sessions, days, weeks, seasons, etc., it can be understood, checked and make additional hypotheses, which also can be tested and this will serve as additional statistics to confirm or deny the robustness.
Signs of fitting:
Logical errors in the algorithms:
I would also add:
Signs of fitting:
And Svinozavr's remark about the fact that in the test, the TS should pass all market phases (flat, up-trend, down-trend), which is consistent with the buy/set skew in the test results, was very correct.
About the fact that the tests with a number of deals up to several hundreds of deals are in fact not trustworthy, I agree completely. But we must formulate a quantitative criterion such as:
- up to 100 trades: fthopu adnat ("unsatisfactory"),
- up to 300 trades: already something ("satisfactory"),
- up to 600 transactions: there is reason to trust ("good"),
- over 800 trades: reliable ("excellent").
The opinion is purely subjective.
Of course, this is not an evaluation of TS, but only one of the criteria according to which we can trust its tests. On the other hand, it doesn't agree much with the "Signs of a Fitting System" topic.
Nevertheless, it may be useful to someone.
And the eighth attribute is very reasonable, by the way. The number of transactions from which one draws conclusions about the stability of a system must be statistically representative.
I criticised the unsuitability of that formulation for practical use: simple declarations without quantitative criteria are nothing.
Criteria that say more about the RIGHT strategy..:
- Increase in profit and number of deals (the same parameters of the Expert Advisor) while reducing the spread.
- Filtering quotes decreases the profit and number of trades. Adding noise - increases.
- Relatively small number of explicit and implicit (any possible) input parameters (important for logic).
Anti-fitting criterion:
- Smooth n-dimensional surface of profile change during optimization (coordinate 1 - input parameter1, 2 - input2, ..., (n - 1) - input (n - 1), n - profile).
If it is not clear, I will explain.
Recommendation:
- When MO is low, it is desirable to work via pauses (including SL and TP).
I almost completely agree with faa1947 about the importance of systematic strategies and the lack of importance of the forward test.
An example of a perfectly correct system (tester only) is given in CodeBase. The example allows you to look at the characteristics of the correct system on the tester.
I would add more:
Signs of fitting:
Here's a strange thing... As for me, it is the opposite of everything - it is a sign of ability to distinguish the direction and move in the right direction, rather than flailing about (and how else to determine the situation 50/50...)
here's the strange thing... As for me, it is the opposite of everything - it is a sign of ability to distinguish directions and move in the right direction, rather than flailing about (and how else to determine a 50/50 situation...).
If the TC is able to distinguish directions and the directions were changing during the test, then there should not be a noticeable misalignment.
If TS is able for example only to buy and it was taught to trade on a profit at the pronounced up-trend,
>> How else can we call it other than a fitting?
HEY!!! There's no one-size-fits-all! Take it any way you like! Then what kind of "sign" is this...
.
The same goes for most of the other "signs" mentioned here...
Well, guys, I just get a kick out of you all. Any TS is a fit for a certain type of quotation and as soon as such an area is encountered the TS makes a profit. And the question is how often, if at all, the area to which the TS is fitted will be encountered in the future. If you have a loss-making forward test, then just another market, and a negative result does not mean anything. The market is constantly changing and forward testing gives no guarantee that the future (for real) market will not be the same as the test one, i.e. the high time will come again. Utter nonsense this fitting and all the fuss and foreshadowing of it. Smart guys from NS have fooled everybody's heads - it is their problem of fitting (retraining) that is central and it exists for a simple reason: NS training is based on a representative sample while trying to turn the initial BP into a stationary one. But the whole problem is that BP is non-stationary and for it there is no concept of representative sampling. That is why TSs based on NS are not fundamentally different from TSs based on other principles.
Regarding testing criteria the author of the thread would do well to acquaint himself with the forum where he opens the thread. This issue has been discussed many times, although I, in my opinion, am the first who does not recognise fitting as such and all talk about it I consider harmful.
That's weird. We were sitting there, talking normally. And then it turns out it's not healthy! It's crazy...
"Buddy, I know it all - I'm one of them, bro,
But you didn't understand anything in the song" (Shevchuk)