TSR - resuscitating trading systems - page 2
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
Constructive? Who's looking at the OOS from behind? That's from the obvious.
Me, I'm not complaining, the effect is the same, another autotrader I know, here's Reshetov more. Sometimes in the back, sometimes in the middle. Try it.
Yuri, this is obvious. Well who can disallow the test / get rid of it? That's not the point, the point is (my opinion) that all these checks are akin to... pre-optimization! That is, getting a positive result on OOS we get TC figuratively (and actually) optimized at another site (which now includes OOS as well). What is the probability that it will be robust further into the future or into the past? What are the objective criteria for assessing future robustness, other than "I think so"?
How can I explain it to you more popularly?
There are various methods of culling trading systems, such as the one outlined by R. Pardo, the one I stated above, and some more. Which one suits you better, choose that one.
I am not going to agitate, persuade or discourage anyone. I put the method on public display and that's it, then everyone can drive it, compare and choose. Want to take it, want to see. I do not give guarantees of reliability in the future, because I am not a market director, and I am not even the ringleader of a shitty central bank of a third country. If tomorrow someone spreads a rumour on the news or starts a joint intervention and TC will fail despite the success of all the tests, the developers of methods of TC rejection are not involved in any way.
Therefore, regarding the probability of a loss and other guarantees, do not come to me, but to the Fed's boss, Ben Chopper.
The market is changing, that is obvious and there is no arguing with this fact.
However, I would argue that causality is evolving (say the market is changing) in the direction of the future, not the past.
Speaking about regularities, we should not forget that they do not exist by themselves as an iron law of gravitation once and for all approved by the Almighty, not at all. They are present because of looking back to past events when making decisions about the future, taking into account additional factors external to the system (price chart). Thus, the market changes (we mean the function which we see on the chart and try to study) under the controlling influence from the outside.
It follows that there is no point in checking the TS on OOS, located before Sample. If the TS will drain on such an OOS, it will mean nothing. Nothing. Just as nothing as if the TS shows positive results.
Yes, there are "great hindsight" to the distant past, which lies on OOS before Sample, but the point is just the fact that TC was trained on smaller Sample (time distance from end of Sample to at least end of OOS), and such distant in time patterns TC simply does not know.
The only correct solution, in my opinion, is to integrate OOS into Sample (to use for optimization stopping criterion), spreading evenly. But there is a chance that such OOS may be found at unstable restless parts of the history and the result on such OOS will be very poor (however it will not mean that the TS is bad), therefore OOS must be big enough - not less than 20%.
Test on the OOS that lies behind Sample. If the results are positive, re-optimise on Sample ending as close to the current point in time as possible.
PS I objected not to the method, but to the practice of using OOS BEFORE Sample. (remark made for especially gifted nerds, not for Reshetov)
Apparently, I am a special gifted one. The enthusiastic and the author please explain how the method differs from crossing two different TCs fitted at different intervals?
Let me remind the very talented ones that the filter is tuned to coherent trading signals and the entire method is based on this. And now try to answer how consistent the signals of two different TS can be?
joo:
PS I did not object to the method, but to the practice of using OOS DO Sample. (the remark is made for especially gifted nerds, not for Reshetov)
Actually Figar0 suggested a good alternative, namely OOS between two Sample. I'll give it a try.
And what about "before" or "after", in some cases results look better "before" and in other cases "after". The truth is somewhere in the middle.
The method, on the other hand, is elementary generalisable:
A kind of diversification, which is not really a diversification. But also a method. Why not?
OK, I have described a more general case where TCs can be inconsistent. We take the same TS, adjust it at different intervals. And then we cross it. Is this exactly what was suggested in the first post?
This is not crossing, but filtering.
I.e. during optimization we use TS, fitting it to the history, pre-filter trade signals. Using the same TS we filter the signals on the other section. We get convinced that they are not filtered enough, then we filter additionally according to the consistency.
I.e. in the process of optimization the filtering at the TC level can confuse true and false signals. In the second stage, all true signals that have not been previously cut out by an error TS, successfully passes through an additional filter (because if the signal is true, then it must have an agreed causal relationship in both areas), but with them trickle and some false, as they may also agree and may not agree (as false signals cause-effect relations are not burdened).
That's actually the whole principle of operation.
In fact, it is possible to break down into more than 2 fitting areas. But because of the limited shelf life of trading signals, moderation must be observed.