Discussion of article "Self-adapting algorithm (Part IV): Additional functionality and tests" - page 7

[Vladimir Karputov]

Amba404:

!!!

Choza mushrooms!!!

I've been writing and writing a comment for 2 hours. Sent it. And there it was, and then it was gone. It's gone.

i.m.wshockerbyl....

You didn't post anything, you just quoted a huge chunk of text.

Attached the text - so it doesn't get lost...

[Amba404]

Vladimir Karputov:

You didn't write anything, you just quoted a huge chunk of text.

I have attached the text so that it will not be lost.....

Thank you, kind man!

I didn't just quote it, I wrote my answers in the same block.

Now I will edit and repost....

[Amba404]

Maxim Romanov:

to: Maxim Romanov

You need to close new series as often as possible - Not only more often, but also with sufficient profit. From my experience: drawdown most of the time, the sum of profit on profitable positions is always less than the sum of minus on unprofitable ones, in case of unfavourable development losses increase exponentially, while the profit grows linearly. Accumulated profit (taking into account that 20% should remain, and let it be not 1 cent) is usually not enough to close the most minus trade (series), only part of it, that's why I called it "partial", although the article did not imply such a thing.

theoretical base of pricing - I will not say anything about shares, but for forex somehow I have not even considered the possibility of its creation, I can't cope with it. Correlation is measurable (does not depend on the theoretical model) and is not constant (function of time), it is not an easy task to predict it with sufficient probability. But to protect against the influence of correlation, for example, with the help of the same hedging of assets, seems to me more achievable.

Ideally, this algorithm and locks are not needed, it is necessary to calculate the points of highest probability of price reversal on the basis of market data. I agree, but here is how it turns out: either 100% correct entry (which is not realistic at all), or there will be rare deep minuses in some directions. If the sum of total profit is much greater than the sum of minus, then we will close losses and profits at once at the end of a week or a month, and we will not cry. Then the second in importance will be the algorithm of the general exit. If the minus is "not a little", then constant, prompt cutting of tails is our everything.

According to my observations, it makes sense to build up a position only against the price movement. I can suggest to consider the following situation. Two currency instruments with different sizes of the currently traded block, but with the same value. We bought both of them and expect an exit in two blocks. One went up, the other went down, with the same speed. After one block we buy more (building up the downside), but the upside is not. After two blocks we got +2 on the first one (and closed), we didn't close -3 on the second one (but we are just now buying 1 more block on it, and when it will end is unknown). Trading is more about faith than knowledge. We believe in the reliability of the system, Well, or not believe, but trust it, with a share of confidence, say 0.6. We don't know the reliability of each signal, we can't say at once that this one is 100 per cent sure and this one is wrong; we trust all 100 per cent of signals in general, but only 60 per cent of each. And if we are so sure that a minus series will reach the planned result, and we build it up, why do we refuse to trust an initially positive trade (not a series, because we did not build it up)?

- Minimum and maximum values of the length of the analysed blocks, as well as their ratio. I have observed that sometimes, with sufficient difference between min and max search size, you can get multidirectional signals at the beginning and end of the range of steps, H block =const.

- The block size is chosen minimum... with decreasing block size, the percentage of outweighing ... candle closing price does not coincide with the block closing price. We discussed this earlier, I will repeat it here briefly. The minimum closest trading distance that makes sense (IMHO) is a triple spread (one for profit and two for losses on entry-exit execution). If on a small block size the distortion errors from the construction increase and it affects the result (which is not necessarily the case), you need to increase the size. If the construction errors depend on the discretisation of the source data (by Close), you need to focus on the size of such discretisation. I use OHLC when plotting, and I am oriented on 5 * iATR(sy, PERIOD_M1, 1440). The fact that close M1 does not coincide with the close of a block is not critical (in my opinion), because according to the construction algorithm, the last block by time is always closed where close[1] M1 is, and the other blocks may well be closed between the two closes, since we initially simplify the price movement to a linear close chart. Regarding the overshoot percentage: I have not fully accepted this paradigm. I define a run frequency distribution plot on history. Then I assign the top of the obtained bell (or hat) and its vicinity, including N% of all observed cases, as a flat. The graph depends only on the block size and the length of the statistical sample, and also takes into account the general direction of movement in the sample (the centre of gravity of the hat will be shifted along the movement). I assign the percentage of flatness arbitrarily. But what run lengths will fall into the specified range to be considered trend or flat - the stat-sample will show, and I see no point in adjusting the boundary between trend and flat depending on the block size.

- There is no boundary between trend and flat- I agree, but see above.


- entry volume is an important issue. Totally agree, I still haven't formed a definite opinion on what is better of the options I've tried above. And I still do not exclude the option of trading with min. lot on all N instruments.

- when a signal is lost, I start over - does it mean closing the open series on which the signal was lost?

- The $10 fix takes the system back to expectation 0 because of randomness. I disagree. It's long to write, and all from "I tried it", no maths, I'll only add that it's not just the $10 fix, but some F(), including from the cost of the block. The cost of the block (and in the case of minlot - the minimum step of profit change) is a basic value for the system, and should be chosen based on the financial expectation (risks and profits) and the size of the deposit. F() itself is a subject of research and discussion.

- total profit I believe is also a certain F(), which depends more on the number of open positions, total profit, the same block value and the amount of confidence in the signal. I think it is important to close all positions regularly, to reduce over sitting and not to get into the market for all available margin, even if profit expectations are not fully realised.

- I was planning to try the partial closing mechanism. I wrote above what I really meant by "partial" closing. I am sorry if I misled you with my unclear statement :(

[Evgeniy Ilin]

Over-sitting on averaging with martin gives straight charts, but as already noted the maximum drawdown on equity is comparable to the profit for the same test period, you can certainly subdivide and find areas that will look good. But the system is dangerous though it is workable, you need a large depot in the region of 10000 quid to withstand over-sitting, I myself also now made on averaging owls, but it's all quite a slippery slope. If the system works and there is a good signal behind it, just a signal to enter and a signal to exit and that's all. But there will be no more such charts, it will be terrible waves and it is not a fact that they will be able to overcome the spread. As practice and tests have shown, the number of falling blocks or just a long price movement does not always give the necessary rollback. And to be honest, the deeper the sitting is, the deeper it can go. In order to include over-sitting you need to know the levels below which it will definitely not sag, and this is impossible in my experience. But you go on, maybe you can do it, you may get a lot of details. But I even made a step multiplier for averaging and the step was multiplied constantly, but there are areas where it just chisels down or up so that neither bottom nor ceiling is visible. It is possible to adjust the peresidku to pull out of course, but the drawdown is comparable to the profit. Try to switch off over-siding and see what the algorithm will give without it. Or the easiest way is to just hide if the over-sit is too high. But it is honest and you will not cheat yourself. The mechanism is certainly beautiful, but very harmful ))))

[Maxim Romanov]

Amba404:

to: Maxim Romanov

You need to close new series as often as possible - Not only more often, but also with sufficient profit. From my experience: drawdown most of the time, the sum of profit on profitable positions is always less than the sum of minus on unprofitable ones, in case of unfavourable development losses increase exponentially, while the profit grows linearly. Accumulated profit (taking into account that 20% should remain, and let it be not 1 cent) is usually not enough to close the most minus trade (series), only part of it, that's why I called it "partial", although the article did not imply such a thing.

theoretical base of pricing - I will not say anything about shares, but for forex somehow I have not even considered the possibility of its creation, I can't cope with it. Correlation is measurable (does not depend on the theoretical model) and is not constant (function of time), it is not an easy task to predict it with sufficient probability. But to protect against the influence of correlation, for example, with the help of the same hedging of assets, seems to me more achievable.

Ideally, this algorithm and locks are not needed, it is necessary to calculate the points of highest probability of price reversal on the basis of market data. I agree, but here is how it turns out: either 100% correct entry (which is not realistic at all), or there will be rare deep minuses in some directions. If the sum of total profit is much greater than the sum of minus, then we will close losses and profits at once at the end of a week or a month, and we will not cry. Then the second in importance will be the algorithm of the general exit. If the minus is "not a little", then constant, prompt cutting of tails is our everything.

According to my observations, it makes sense to build up a position only against the price movement. I can suggest to consider the following situation. Two currency instruments with different sizes of the currently traded block, but with the same value. We bought both of them and expect an exit in two blocks. One went up, the other went down, with the same speed. After one block we buy more (building up the downside), but the upside is not. After two blocks we got +2 on the first one (and closed), we didn't close -3 on the second one (but we are just now buying 1 more block on it, and when it will end is unknown). Trading is more about faith than knowledge. We believe in the reliability of the system, Well, or not believe, but trust it, with a share of confidence, say 0.6. We don't know the reliability of each signal, we can't say at once that this one is 100 per cent sure and this one is wrong; we trust all 100 per cent of signals in general, but only 60 per cent of each. And if we are so sure that a minus series will reach the planned result, and we build it up, why do we refuse to trust an initially positive trade (not a series, because we did not build it up)?

- Minimum and maximum values of the length of the analysed blocks, as well as their ratio. I have observed that sometimes, with sufficient difference between min and max search size, you can get multidirectional signals at the beginning and end of the range of steps, H block =const.

- The block size is chosen minimum... with decreasing block size, the percentage of outweighing ... candle closing price does not coincide with the block closing price. We discussed this earlier, I will repeat it here briefly. The minimum closest trading distance that makes sense (IMHO) is a triple spread (one for profit and two for losses on entry-exit execution). If on a small block size the distortion errors from the construction increase and it affects the result (which is not necessarily the case), you need to increase the size. If the construction errors depend on the discretisation of the source data (by Close), you need to focus on the size of such discretisation. I use OHLC when plotting, and I am oriented on 5 * iATR(sy, PERIOD_M1, 1440). The fact that close M1 does not coincide with the close of a block is not critical (in my opinion), because according to the construction algorithm, the last block by time is always closed where close[1] M1 is, and the other blocks may well be closed between the two closes, since we initially simplify the price movement to a linear close chart. Regarding the overshoot percentage: I have not fully accepted this paradigm. I define a run frequency distribution plot on history. Then I assign the top of the obtained bell (or hat) and its vicinity, including N% of all observed cases, as a flat. The graph depends only on the block size and the length of the statistical sample, and also takes into account the general direction of movement in the sample (the centre of gravity of the hat will be shifted along the movement). I assign the percentage of flatness arbitrarily. But what run lengths will fall into the specified range to be considered trend or flat - the stat-sample will show, and I see no point in adjusting the boundary between trend and flat depending on the block size.

- There is no boundary between trend and flat- I agree, but see above.


- entry volume is an important issue. Totally agree, I still haven't formed a definite opinion on what is better of the options I have tried. And I still do not exclude the option of trading with min. lot on all N instruments.

- when a signal is lost I start over - does it mean closing the open series on which the signal was lost?

- The $10 fix takes the system back to expectation 0 because of randomness. I disagree. It's long to write, and all from "I tried it", no maths, I'll only add that it's not just the $10 fix, but some F(), including from the cost of the block. The cost of the block (and in the case of minlot - the minimum step of profit change) is a basic value for the system, and should be chosen based on the financial expectation (risks and profits) and the size of the deposit. F() itself is a subject of research and discussion.

- total profit I believe is also a certain F(), which depends more on the number of open positions, total profit, the same block value and the amount of confidence in the signal. I think it is important to close all positions regularly, to reduce over sitting and not to get into the market for all available margin, even if profit expectations are not fully realised.

- I was planning to try the partial closing mechanism. I wrote above what I really meant by "partial" closing. I am sorry if I misled you with my unclear statement :(

Without theory, there is no point in measuring correlation and collecting statistics. They will just be numbers that don't understand how they change. A theoretical model is necessary. someone will describe that it is a martingale, someone will write that it is a perisivitel or a grid. Without a theory it can look like that. With a theoretical model it becomes quite different. For example, I can show how changing the lots on the fund can make a positive expectation. From the outside it will seem that this is something like martingale, because lots are constantly floating. But there is one difference: without theory it is just floating lots, but with theory I know what lots should be set and why. The model should not be for forex or funds, it should describe the differences and similarities between markets, including commodities and cryptocurrencies.

You don't need 100% probability of entry. There will always be unaccounted factors that will not allow you to make 100%. The more accurate the model, the higher the probability.

Again, why can't we just collect stat data from different scales and use it. Without a theoretical model it will just be numbers whose behaviour cannot be predicted. Error based correction is needed to avoid collecting stat data. We need to determine the nature of the market using statistical data, understand how well it fits the model, and take the parameters of performance not based on statistical data but on the model. If we are guided only by statistics, then.... Anyone can do it, there is nothing surprising here, any mathematician can do it. But then everything turns into chasing fluctuations of these parameters and we do not understand why and when it all changes. Collecting stat data and launching a robot to trade is no different from optimisation. That is why I am developing a theoretical model. The robot does not trade the market, but a model whose parameters change.

If the signal is lost, the series is not closed. I don't understand what lost means. The series starts when a signal appears and continues until the price reaches the calculated closing point. The estimated closing point is constantly being adjusted due to the appearance of new blocks. It cannot happen that positions have not closed and the chart has already corrected. To be more precise, it can happen in my case, but it is already the details of the algorithm implementation and the mechanism of compensating unprofitable positions is triggered here.

As for the total profit, if you adjust it dynamically from some parameters, you can come up with something meaningful, but I haven't done it yet.

As for partial closing, there is also something to work on, but for me this stage will be later, I haven't thought about it yet.

[[Deleted]]

hello

I will add some water )

the algorithm is not an algorithm - on different pairs different results
the wind blew, the logic broke
an attempt to create an algorithm.

but in general it's cool
please continue
waiting with interest for the final result

[Aleksey Mavrin]

Спартак Угланов:

hello

I'll add a little water.)

the algorithm is not an algorithm - different results on different pairs
the wind blew, the logic broke
an attempt to create an algorithm

but in general it's cool
please continue
waiting with interest for the final result

This is not a valid statement. Like "a car is not a car, it is not the same on the bog as on the tarmac". It's rubbish.)

[Marcel Fitzner]

Excellent article! Gives much inspiration. Ty for your hard work! 

[Maxim Romanov]

Marcel Fitzner:

glad you liked it

[Guilherme Mendonca]

Congratulations!!

Very good article. Thanks for share your knowledge.

It began with a simple idea and was getting more complex.