Discussion of article "Metamodels in machine learning and trading: Original timing of trading orders"
I haven't looked at the code, purely from the description: it looks like the metamodel will be trained on future data already on the 2nd iteration. I think we should train on past data only.
The metamodel is trained on the same data as the base model + a piece of history back to the past. You can change it to the future. No difference, it just needs a piece of data that the baseline has not seen.
I'm talking about the phrase
Сначала необходимо обнулить книгу плохих сделок, если в ней что-то осталось после предыдущего обучения. Затем в цикле задается необходимое количество итераций.
This book will accumulate information about the future. It's not like it's reset before every training.
I'm talking about the phrase
This book will accumulate information about the future. It's not like it's reset before every training.
What future? There are examples for the metamodel.
the more iterations the base model gets wrong, the more the examples get stronger and fall into "don't trade".
These examples are the errors of the base model, where it is most often wrong. They need to be amputated so that they don't pull the gradient on themselves.
similar to the 3rd class of "do not trade", except that they are thrown out of the classifier altogether, increasing its accuracy
A machine can be trusted to accurately execute the algorithm put into it. The person who put this algorithm into it cannot be trusted.
It is inherent in human beings to be wrong, to be deliberately misleading, or to deliberately mislead others. All works on machine learning in the absence of a normal market theory are like looking for a watch under a lantern (from an anecdote). Algorithms and foundations should be looked for in quantum mechanics. If there is a desire, I can explain.
Since the time of Pythagoras, a distinction has been made between common and secret knowledge. The topic is very large, it touches upon philosophy, politics, and economics. Concretising this statement in relation to the stock exchange, we pay attention to James Simons. The man doubled his capital every month using his knowledge of the mathematics of multidimensional spaces. It means that there is a real possibility to calculate the behaviour of this very multidimensional space, i.e. the stock exchange field. A number of authors, for example, N. Rudyk, "Behavioural Finance", Lars Tweed, "Psychology of Finance", J. Soros,"Alchemy of Finance", pay attention to the fact that the market is first of all psychology. The psychology of masses is not described by Gauss distribution, it is Tsipf-Pareto distribution. Probability theory is inapplicable to psychology because of an error in the fundamental foundations of this theory. Question: what is applicable to the psychology of the masses? I.Danilevsky wrote a good book: "Structures of collective unconscious. Quantum-like social reality". Its meaning: the collective unconscious is a quantum field, which can and should be studied with the mathematical apparatus of quantum mechanics.
Secret knowledge is secret because it is not widely available. But nevertheless, if there is a desire, it is possible to go beyond the well-travelled track and to search for one's own way.
Secret knowledge is secret because it is not widely available. But nevertheless, if one is willing, one can go beyond the beaten track and find one's own way.
The explanation must consist of verifiable facts, not an appeal to arcane knowledge and buzzwords
The theoretical apparatus of quantum mechanics was developed in the early twentieth century, practical use began 50 years later - masers, lasers, etc. If you don't like to read smart books, it doesn't mean that there are no such books and people who wrote them. For you, it seems that "the pot is empty is much more appreciated". Keep pouring from empty to empty.
The theoretical apparatus of quantum mechanics was developed in the early twentieth century, practical use began 50 years later - masers, lasers, etc. If you don't like to read smart books, it doesn't mean that there are no such books and people who wrote them. For you, it seems that "the pot is empty is much more appreciated". Keep pouring from empty to empty.
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
You agree to website policy and terms of use
New article Metamodels in machine learning and trading: Original timing of trading orders has been published:
Metamodels in machine learning: Auto creation of trading systems with little or no human intervention — The model decides when and how to trade on its own.
First, I need to make a small remark. Since the researcher deals with uncertainty while developing trading systems (including the ones applying machine learning), it is impossible to strictly formalize the object of search. It can be defined as some more or less stable dependencies in a multidimensional space that are difficult to interpret in human and even mathematical languages. It is difficult to conduct a detailed analysis of what we get from highly parameterized self-training systems. Such algorithms require a certain degree of trust from a trader based on the results of backtests, but they do not clarify the very essence and even the nature of the pattern found.
I want to write an algorithm that will be able to analyze and correct its own errors iteratively improving its results. To do this, I propose to take a bunch of two classifiers and train them sequentially as suggested in the following diagram. The detailed description of the idea is provided below.
Each of the classifiers is trained on its own dataset, which has its own size. The blue horizontal line represents the conditional history depth for the metamodel, and the orange ones stand for the base model. In other words, the depth of history for a metamodel is always greater than for the base one and is equal to the estimated (test) time interval, on which the combination of these models will be tested.
Author: Maxim Dmitrievsky