Discussion of article "Extract profit down to the last pip" - page 16
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
e.g. log(gain %/drawdown %)
As an optimisation criterion, these values are equivalent to the expression under the logarithm.
Composter above suggested a classical FV with time rationing.
Something is not well understood. We are discussing three entities for reinvest-TS.
A few words to MT5-Tester.
The article showed how MT5-Tester was used to find market patterns. But the description of the research work when writing a TS was completely missing.
As a rule, several experimental TSs are written, which in themselves are a kind of research. They may differ in some blocks from each other. Most often, these are not trading blocks themselves, but algorithms of forming trading signals. That is, the changes are contained in small but determining parts.
With due diligence, such TSs can accumulate a lot of them. And market research starts to be reduced to analysing how well ready algorithms can squeeze out profit from a given trading interval.
MT5-Tester is multi-core (and cloud-based) and has a good built-in implementation of genetic algorithm, which allows even a beginner to perform efficient and fast calculations on a huge (can be cyclopean in size) space of input parameters. And when the task of comparing TCs arises, it comes down to comparing the results of Optimisation of each TC.
Doing it manually is extremely tedious even in such a simple and powerful tool as MT5-Tester. It is especially hard when you need to do it often. And in algotrading it may be required at least once a week or more often, which depends only on the degree of laziness of the TS author. Although not only TS authors, but also ordinary users may face such a task. For example, if you want to compare several ready-made Expert Advisors from the Market.
MT5-Tester has inbuilt automation mechanisms, but they require a good qualification. That's why the order execution settings (whether they slide, etc.) and commissions.
To implement each item you do not need to run the Tester.
I.e. such a deep research of the TS can be carried out even for paid robots from the Market.
Let's assume that there is some market regularity. It can be characterised by the fact that you can write several TS with different logic and they will capture profit, but only with different efficiency.
I measure the efficiency of profit capturing, for example, by the profit of the best pass when optimising with a constant lot. Higher the profile - higher the efficiency. I.e. the higher the profit is, the more accurately the trades are overlaid on the price chart.
Classic vs NS
It is interesting that different TS are not only different sets of input parameters, but also logic. The logic itself can be described through NS with its own number of inputs in the universal internal logic.
When I optimise a classical TS, where the logic comes from the head, there is no question of adjusting the logic to the market. I.e. the logic is already created, and further you just optimise its few input parameters.
However, when dealing with the TS on the EA, there is a huge number of input parameters of the "universal TS". Accordingly, the probability of fitting is much more intuitively higher than with classical TCs.
Where is the flaw in my reasoning?
Trading costs
For scalping TSs, commissions play a big role. In this case, the price is exactly up to it. I.e. market regularities practically do not depend on this indicator. Accordingly, it is logical to use the profit without commission when comparing TSs. But the commission cannot be ignored in real trading. That is why TS with a higher expectation and, of course, a lower profit indicator, which was obtained during optimisation without commission, are launched into the battle. It turns out that TSs that describe market regularities worse than they could, go into battle. It is somewhat paradoxical.
And, accordingly, if a fighting TS describes the market regularity less effectively, its probability of losing is higher than that of the one that shows a wonderful profit without commission (the expectation matrix is higher than the commission).
Comparison of TSs
Let's assume that TS1 has the same number of explicit input parameters as TS2. But the best pass of TC1 gets 10% more profit than TC2. Let it be thousands of rollover trades, so that we don't get distracted by statistical significance.
Is 10% a lot or a little to say that TC1 describes the market pattern better than TC2? How to understand where it is about the statistical error of the profit, and where it is about a really more accurate algorithm for describing the market?
In general, I have written rather ramblingly. If anyone has seen the logic in this set of eloquence and has some thoughts on the topic, it would be interesting to hear them.
Comparison of TCs
Suppose TC1 has the same number of explicit input parameters as TC2. But at the same time the best pass of TS1 gets 10% more profit than TS2. Let it be thousands of rollover trades, so that we don't get distracted by statistical significance.
Is 10% a lot or a little to say that TC1 describes the market pattern better than TC2? How to understand where we are talking about the statistical error of the profit and where we are talking about a really more accurate algorithm for describing the market?
I think the standard basic method is to compare their "annualised Sharpe ratios". Or is it about something else?
All in all, it's pretty rambling. If anyone saw logic in this set of slurs and has thoughts on the topic, it would be interesting to listen.
Any variant of the received strategy can be a conditional fit to the data - it doesn't matter whether it was invented or obtained with the help of the MO.
If we analyse why the TS stopped earning, it may turn out that the market has changed significantly, i.e. the strategy has fallen just on the data about which it knew nothing, or this data was earlier and it was not possible to learn to earn on it. The problem of a small time window for learning.
Another peculiarity of MO is that well correlated predictors in the training time window can be scattered when applying the model, i.e. the correlation was false. I observe this sort of thing in trees when a portion of similarly structured leaves stop earning. I.e. part of the splits closer to the leaves can be a random value due to random correlation of predictors "filtered" for the best split, as a result similar leaves will be profitable and unprofitable although their structure is extremely similar. This is a problem of both non-stationarity and insufficient data, and, given the peculiarities of these circumstances, it is necessary to make adjustments to the MO process itself, I have ideas, but unfortunately I am not a programmer to implement them.
I measure the efficiency of profit capture, for example, by the profit of the best pass when optimising with a constant lot. Higher profit - higher efficiency. I.e. the higher the profit, the more accurately the trades are overlaid on the price chart.
The following conclusions depend on the criterion chosen here. Therefore, it seems to me that it is wrong to take just profit here.
Classic vs NS
It is interesting that different TS are not only different sets of input parameters, but also logic. The logic itself can be described through NS with its own number of inputs in universal internal logic.
When I optimise a classical TS, where the logic comes from the head, there is no question of adjusting the logic to the market. I.e. the logic is already created, and further you just optimise its few input parameters.
However, when dealing with the TS on the EA, there is a huge number of input parameters of the "universal TS". Accordingly, the probability of fitting is much higher intuitively than in classical TS.
Where is the flaw in my reasoning?
Everything is logical.
Probably the difference is just in the "small number of inputs" of the classical variant and competent and logical inputs for the neural network.
Trading costs
Commissions play a big role for scalping TS. In this case, the price is exactly up to it. I.e. market patterns are practically independent of this indicator.
I disagree. What is "price"? Is it the same everywhere, or is there a benchmark?
My point is that, perhaps, the size of a particular broker's commission may depend on his quotes (number and methods of LP aggregation, filters, etc.).
Not to mention the direct mark-up added to prices.
fxsaber:
Accordingly, it is logical to use commission-free profit when comparing TCs.
I fundamentally disagree.
A strategy is obliged to take costs into account, as is the criterion for choosing the best one.
Why do we need a "perfect description" of spherical quotes in a vacuum that cannot be traded?
I disagree. What is "price"? Is it the same everywhere, or is there a benchmark?
My point is that, perhaps, the size of a particular broker's commission may depend on its quotes (number and methods of LP aggregation, filters, etc.).
Not to mention the direct mark-up added to prices.
Price is the best trading conditions available without throwing anything on top.
Fundamentally disagree.
Strategy has a duty to consider costs, as does the criterion for selecting the best one.
Why do we need a "perfect description" of spherical quotes in a vacuum that cannot be traded?
The better the TS describes the price, the closer its logic corresponds to the market regularity. The price does not include the commission. Having determined the logic of such a TS, it makes sense to move further from it towards its combat application. I.e. first you need to find not a TS that will give profit with commissions. But the TS (logic) that will give the highest profit without commissions.
It seems the standard basic method is to compare their "annualised Sharpe ratios". Or are we talking about something else?
Well acquainted with sharpe ratios. I have no confidence in the direction of any optimisation criterion to talk about the quality of logic in describing market patterns.
A conditional fit to the data can be any variant of the obtained strategy - it does not matter at all whether you came up with it yourself or got it with the help of the MOE.
If I understand MO correctly, it is a "Universal TS" with a huge number of input parameters. Therefore, I do not quite share this thesis.
If I understand the MO correctly, it is a "Universal TC" with a huge number of input parameters. Therefore, I do not quite share this thesis.
Personally, I use those predictors for training, which I use in real trading, i.e. this is information that has been processed, and MO is only engaged in the actual search for patterns in this information, the same as if I sat down and started to make up my own patterns from these indicators, and then check them on the tester.