How do I identify the patterns on which the ready-made TS shows a profit? - page 4
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
The terms of the question are rather abstract and streamlined. I think, to be specific, you need some details about the TS: what is its input (indicator, set of conditions... etc.) and what pair, what time interval to test?
Otherwise, we can say the same abstract: the comparison of the logic of position opening and the price behavior on the time interval will give the answer to your question.
...
The question of the branch is how to reengineer any of its TCs
Ward 6 ?
Imagine that you deliberately wrote a flat trading system knowing that it will fail on trends. The only hope was that it would be less losing than earning. But in the end it works great through some trends and on the contrary, it fails in some flat patterns. Obviously, its profit is not due to initial flatness. And the reason is that the logic you have laid down, you have not fully understood.
If averaging is used, by any algorithm, the regularity "a trend always ends in a correction" is exploited. The predisposition to corrections is different for different Per's, but the axiom is immutable.
I faced the phenomenon of the unknown regularity when I made a filter just for the sake of checking visual observations - I fixed the МАС position and the result considerably improved on all pairs. I called the effect "price settling".
Quite abstract and streamlined the terms of the question. If you want to be more specific, you may need more details about the TS: what do you use as input (indicator, set of conditions, etc.) and what pairs, what time interval is it tested on?
It was important for me to outline the problem and to see if I am aware of it at all. Who is confronted with it and how many consider it "ward number 6". Certainly the author himself will dig in. And there have already been thoughts about correlations and other attempts at comparison. Certainly there is a need to make comparisons. It was interesting to know who does it and how. My particular case is a particularity of this thread to begin with.
It's a bit of a surprise that I didn't find confirmation:
Interestingly, most of the TCs in the world that have made profits are written on this very principle - random TCs. That is, people get profits but don't fully understand what patterns they are exploiting. Circumstances have simply worked out in such a way that profits are gained in the tester. They tune up even more with various filters, etc., getting even more profit in the tester. But they do not understand the basis of the pattern. However, they do not need to delve into it when they already have profit. It is better to spend efforts on primitive improvement of blackbox.
The maximum reasonable explanation I've heard from creators of the TS that brought profit - I exploit the night flat on this pair. As a rule, the creator has found this pair by accident or picked it up from someone else's monitoring. He does not know why the algorithm is so precise. The result is positive in the tester. Sometimes it is even better than the monitoring benchmark. So I am satisfied. What to think!
And there may be explanations that Hirst is very different from 0.5. That is why I have a profit. But this is nonsense as an explanation of the basis of the profit. Because there are two reverse-TC, but it is like day and night according to the results.
Still, there are people who write similar things:
-Aleks-:
Faced the phenomenon of an unexplained pattern when I made a filter purely to check visual observations - I regularised the position of the MAs and the result improved considerably on all pairs. Called the effect "price settling"...
SZ
show the miracle graph, MO if you can.
For a long time I was complexed by the fact that I have never understood the reasons of profitability of this or that TS I had written. Instead of a systematic approach I was always defeated by Chance. Then I gave up on it and stopped trying to understand why a particular thing works. I just write and use. I think it is useless to look for the reason.
It is still frustrating that always complex, but seemingly reasonable and even well OOP-implemented logic is inferior in profitability to dumb, like a hammer, logic. I.e. you try to think up adaptivity, twist and turn, cumbersome, but in the end a hatched variant with immutable parameters (in the optimizer, you just trim it and that's all) excels. And such failures are, to put it mildly, frustrating. Because it is many times more expensive to write complex logic than simple one. And you expect a payoff accordingly. But there is none.
Does it mean that the basis of a profitable TS must be simple? Or the author has reached the ceiling of his intellectual capabilities and it is useless to keep trying?
It still frustrates me that logic which is always complex but seems reasonable and even well implemented in OOP is yielding to dull as a hammer logic in profitability. I.e. you try to think up adaptivity, twist and turn, cumbersome, but in the end a hatched variant with immutable parameters (in the optimizer, you just trim it and that's all) excels. And such failures are, to put it mildly, frustrating. Because it is many times more expensive to write complex logic than simple one. And you expect a payoff accordingly. But there is none.
I know this all too well. How much effort and time was put into the cleverest research - and then some waving(moving average) turned out to be more effective!
It was also noticed that TC almost always fails to improve. I.e. you write a TS and get a good result. I think I'll add this and that again and it will be absolutely good, then you do all that and run it and get a bummer. The result just gets worse and worse.
OK, let me try again. You are doing machine learning on some data. Let's say you don't have much time, and you don't have time to do a thorough statistical study of the input data. You build some model based on your experience, train it and get a decent result on OOS. What you have in the end:
Many developers will say that pattern == this model. I am against this formulation. It is the model itself that sometimes falls into a given pattern. But that is why this hitting occurs - no answer.
Now to our price rams. Again, we have a flat TS that is positive on flat patterns and negative on trends. But suddenly, such a pair appears on which it is showing profit even on trends, and on some of the fluxes it is losing. But in the end it is much better than on other pairs: heaven and earth.
You try to understand what is so great about this pair. That is, what is the pattern in it, which is absent in other pairs. I could call it a pattern like "my TS is earning cool". But not in kindergarten. I want to understand what peculiarities of the pair make money. I hope it's clear now. I cannot give any other explanation.
It has also been observed that TC almost always fails to improve. That is, you write the TS and get a good result. You think, now I'll add this and that and it will be absolutely good, you do all this, run it and get a bummer. The result just gets worse.
The only improvement, which is difficult to call in such a big word, but still it gives results - limitation of trading by time of day and even day of week. Sometimes disabling some Tuesday and then over-optimising gives amazing results. But such tricks can be just explained by the fact that the patterns depend on both the time of day and the day of the week.
But the addition of tweaks, indeed, almost always does not bring improvement, apart from the obvious cases of tweaking.
Sometimes shutting down some Tuesday and then over-optimising it produces surprising results...