Discussion of article "Extract profit down to the last pip" - page 5
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
An example on the topic of regularities in price history. Let's take EURGBP.
Over the years of this pair's history, there was a fundamental relationship between EUR and GBP based on the EU treaties. This allowed sometimes to find long-lasting patterns, on which many people made money.
At the moment, the fundamental connection is broken. But it is not properly reflected in the price history.
In this way, you can get a cool result on history. For example, you optimised for six months, and then on OOS for many years the result is the same.
But the bummer is that there is no preparation for Brexit in that history. So the collapse of the "pattern" found is highly likely. And this problem cannot be solved by any machine learning techniques. There is simply no data to train on. None at all.
It is not for nothing that some trading floors have been setting high margin requirements for GBP-pairs for a year now.
Do you teach fundamental patterns? It seems to me that it is the technique of movement that should be taught.
However, I agree that strong shocks can make it impossible to train on old data, like USDRUB (Si) before 2014 and after.
Do you teach the fundamental patterns of the model? It seems to me that it is the technique of movement that should be taught.
If the result from the article was obtained on EURGBP, the probability of robustness would be considered much lower.
If the result from the paper had been obtained on EURGBP, the probability of robustness would have been considered much lower.
The question is how to determine early on that the market has changed, rather than just being in an area where the results were poor during training. If you solve this problem, you can significantly reduce losses.
But the bummer is that there is no preparation for Brexit in this story. So the collapse of the found "pattern" is highly probable. And this problem cannot be solved by any machine learning techniques. There is simply no data to train on. None at all.
The question is how to determine at the early stages that the market has changed, and not just being in the area where the results were weak during training. If you solve this problem, you can significantly reduce losses.
I revisited selectively Game Theory on Savvateev's video lectures, I think I managed to "glue this information together" ))))
According to game theory, we play a positional game with incomplete information, in other words: we play a card game where we can see the moves of the opponent, but we have an infinite pack of cards, i.e. even knowing all the previous moves of the game, we can not guess the future outcome.
In this formulation of the problem I again remembered and reread my memo from Max Gunther:
Auxiliary Axiom #5. Beware of the trap of historical parallels.
Auxiliary axiom #6. Beware of the illusion of recurring figures.
Auxiliary axiom #7. Beware of the delusion that correlation and causation exist.
Basic axiom No. 8. On religion and the occult.
Subsidiary axiom #12. If astrology worked, all astrologers would be rich people.
Auxiliary axiom #13. There is no need to shrug off superstitions. They can be fun if they have their proper place.
revisited selectively Game Theory on Savvateev's video lectures, I think I managed to "glue this information together" ))))
According to game theory, we play a positional game with incomplete information, in other words: we play a card game where we can see the opponent's moves, but we have an infinite pack of cards, i.e. even knowing all the previous moves of the game, we can not guess the future outcome.
In this formulation of the problem again remembered and reread his memo from Max Gunther:
I think it all fits again ((((Is algotrading doomed? Or is it still necessary to determine when a new pack of cards starts?
Or, do we still have to determine when the new pack of cards starts?
once again: we are playing a positional game with incomplete information.
in any analysis of market data, even if you consider the market from the position of economic cycles, even if it is wave theory, even if it is the mentioned MO..... in any case we have an "infinite pack of cards", it is "infinite" because you can assume that you have found the "beginning" of the cycle / wave / data for the MO, but there is nothing else to rely on except the assumption, perhaps you should have taken the data even deeper in history?
and to the point of the article, or as far as I understood it: we do not know the source of market patterns, but we can develop a profitable strategy (in our opinion) and quickly run this strategy through all possible data (trading symbols) and thus the task of finding a profit is reduced to developing a strategy, and the author of the article automated the search for a trading instrument and showed his analysis of the results - "where to look".
Once again, we're playing a positional game with incomplete information.
at any analysis of market data, at least consider the market from the position of economic cycles, at least it will be wave theory, at least mentioned MO..... in any case we have an "infinite pack of cards", it is "infinite" because you can assume that you have found the "beginning" of the cycle / wave / data for the MO, but there is nothing else to rely on except the assumption, perhaps you should have taken the data even deeper in history?
There is no complete randomness in the market, there are different states yes, but they change, probably not abruptly over time, but rather flow from one state to another.
If a pack of cards is even infinite in volume, it doesn't mean that the cards in it are unique....
There is not complete randomness in the market, there are different states yes, but they change, probably not abruptly over time, but rather flow from one state to another.
If a pack of cards is even infinite in volume, it does not mean that the cards in it are unique....
totally agree!
but the problem of "infinite pack of cards" is not less? .... there is a pack of cards for 54 pcs, we play a regular game, when the number of moves increases, we can guess the number of cards we need (the game is positional, right? i.e. we see the moves of the participants, in our case bars / price).
well, and in relation to the market this "infinite pack of cards": there are patterns found in history, but we cannot say that these patterns "are still in the deck" and / or these patterns will occur in the near future.
market condition - context - yes there is, and I am trying to look for it, but here the task, most likely, cannot be automated, i.e. the task should be reduced to searching for several TS, which should be applied at the trader's discretion at a particular time, in my opinion it makes sense to look for this, and not "a beautiful chart on the history for 10 years", imho.
I totally agree!
but the problem of "infinite pack of cards" is not less? .... there is a pack of cards for 54 pcs, we play a normal game, when we increase the number of moves, we can guess the number of cards we need (the game is positional, right? i.e. we can see the moves of the participants, in our case bars / price).
well, and in relation to the market, this "infinite pack of cards": there are patterns found in history, but we cannot say that these patterns "are still in the deck" and / or these patterns will occur in the near future.
market condition - context - yes there is, and I am trying to look for it, but here the task, most likely, cannot be automated, i.e. the task should be reduced to searching for several TS, which should be applied at the trader's discretion at a particular time, in my opinion it makes sense to look for this, not "a beautiful chart on the history for 10 years", imho.
Is it possible to calculate Brownian motion? You can, but only it is costly. There is no pure randomness in molecular motion, but pressure changes create a trend. With molecules it's easier, their functions are not discontinuous, unless you take quantum mechanics, of course. Here the functions are discontinuous, and predetermined by many unaccounted factors. Is it possible to take into account all the factors in historical data? I don't know, I think that taking into account and training on multifactor data will probably increase the chance of machine learning, but how to formalise these factors to numerical or logical values.
New article Extract the profit down to the last pip:
Author: fxsaber