Algorithm Optimisation Championship. - page 108
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
Well. Since fitting is only a special case of optimization problems, then how can speed and efficiency not be necessary. What's the use of finding maxima if they change faster than the algorithm looks for them. Imagine that you have a mechanism for accurate minutiae prediction (let it be at least 60% of the time), but the calculation is done for a period of more than a minute, and what is the use of such predictions. And even if you are right in fact the forecasts are correct. It's the same here.
In my understanding, optimization is the improvement of the ratio of two or more values/parameters or something else (price/quality, for example). Or in terms of trading, it is bringing an already profitable (positive) TS (position) up to the best parameters (TS or mm), risk/return or something like that. Optimization in one way or another only serves to modify existing ones. If you do not have positive rules, what do you optimize? If you optimize a random series then whether quick or slow optimization is useless, if the output is also random. I think Andrey Zelinsky was referring to these meanings. I may be wrong. Searching for a maximum makes sense when you know that this maximum is meaningful. Otherwise why bother searching for it at all. And that's where speed matters. And if you know how to detect this meaningful maximum it is already a ts with speed. And optimisation is just fine-tuning. That is, it is not primary. And we are told here that optimization is the first and the more so its speed. Probably one can use optimization algorithms and add other algorithms to them to obtain ts with Mo, but it will not be optimization in its pure form. Presumably, Dick has such symbiosis, that is, a profitable algorithm that can be plugged into a common optimization algorithm, of course, then he is interested in options to find out whether these common optimization algorithms (essentially just a wrapper) are faster than his. The other participants are competing only for competition. Because they have no such symbiosis and naively think that competitions will give them something.
Example number 1. An Expert Advisor on a neural network (any network or any other similar technology) with self-learning. Such an Expert Advisor can work by itself controlling its final trading stats and start self-optimization when needed.
Example #2. Portfolio Trading. It requires going through hundreds of trading instruments to put together a portfolio with the required characteristics.
Example number 3. In general, TS using the direct signals from the indicators need to be often optimized, and it does not guarantee a profit in the future. But there is an approach where indicators are not a direct source of signals, but a kind of support, on which TS is based, and the profitability is obtained due to the statistical characteristics of the series. In this case it is necessary to find the optimal parameters to transform the statistical characteristics of series into signals.
Example number 4. Traders often argue that optimization of Expert Advisors, for example, for profit is just a tweaking and it is stupid to expect profitability from such an EA in the future. And this is a correct statement. But why are not they looking for ways of optimizing not TC parameters but optimization criteria? Optimization of optimization criteria is like the first derivative.
I can give infinite examples of practical use of optimization algorithms, traders come up with ideas every day and need to look for the best implementation of these ideas. Only short-sighted people can deny the urgent need for accurate optimization algorithms for traders, and I hope you are not one of those people.
Example number 1. An Expert Advisor on a neural network (any network or any other similar technology) with self-learning. Such an Expert Advisor can work by itself controlling its final trading stats and start self-optimization when needed.
Example #2. Portfolio Trading. It requires going through hundreds of trading instruments to put together a portfolio with the required characteristics.
Example number 3. In general, TS using the direct signals from the indicators need to be frequently optimized, and it does not guarantee a profit in the future. But there is an approach where indicators are not a direct source of signals, but a kind of support, on which TS is based, and the profitability is obtained due to the statistical characteristics of the series. In this case it is necessary to find the optimal parameters to transform the statistical characteristics of series into signals.
Example number 4. Traders often argue that optimization of Expert Advisors, for example, for profit is just a tweaking and it is stupid to expect profitability from such an EA in the future. And this is a correct statement. But why are not they looking for ways of optimizing not TC parameters but optimization criteria? Optimization of optimization criteria is like the first derivative.
I can give infinite examples of practical use of optimization algorithms, traders come up with ideas every day and need to look for the best implementation of these ideas. Only short-sighted people can deny the urgent need for accurate optimization algorithms for traders, and I hope you're not one of them.
You have given interesting examples.
However, adjusting the trading strategy parameter values to maximise the profitability of the trading strategy being tested over the selected historical period with the expectation of future profit using the same values in the current trade is not always a fool's errand.
It all depends on the trading strategy itself and on the parameters to be optimized. If they are not "stupid" per se, then such fitting can indeed be useful for current trading as well.
For example, if a trader calculates the volume jump value relative to its average value for the current three months. By the optimization results, the trader finds the jump value (in percents) when it is the most profitable to open a deal, because such deals turn out to be the most profitable.
I think you can find more examples of not "stupid" fitting.
Lately, interest in neural networks has waned, as traders find it very difficult to find a profitable strategy outside the training sample. The same is true for classical indicator-based TS. But the problem is not with the neural networks and not with indicators, the problem is with the wrong criteria, the wrong optimization objectives. Very often traders reject their TS ideas, as they fail outside the optimization area, but they do not even realize that the ideas are profitable, but the optimization criteria are wrongly selected. It turns out that even the simplest TS on two MA bars may be profitable, yes, we just need to look at them from a different angle, adding a "smart" stop, adding a trawl, providing a competent MM, optimize it all correctly and voila! - Working TS. You can even make soup from the axe, as we know the chief ingredient - the wit of the cook.
The optimization algorithm itself is only a tool, a microscope in the capable hands of a scientist, but if the scientist has no brains, no microscope, even the most accurate one, will help him.
What do you mean by"optimization criteria"? It seemed to me that there is only one optimization criterion - it is a set of optimization parameters (the smaller their number - the better, and ideally one - period), providing the maximum value of FS recovery factor = ratio of net profit to maximum drawdown.
It is possible. But if we look for no-brake entries, i.e. as places to put the biggest stop within a forecast, the speed of getting a forecast on a small timeframe (within the framework of a high price forecast) is also important. It's not pipsing and long time to make long positions on a large timeframe but the time to enter a position is important too. If we cannot increase the accuracy of forecast on m5, then why not to look at similar prediction on m1. And then accuracy just depends on velocity.
I've made a walkthrough according to the principles of one classical method of finding extrema of a function of many variables.
I took F(x1,x2,x3)=exp(x1+x2+x3)/(x1*x2*x2*x3*x3*x3) as function to check;
I got these results.
The specified search error is 0.01.
Initial parameters (first lookup) x1=x2=x3=0.5;
Search range 0-100
Number of times the function is called - 51
Minimum Fmin=3.76210
x1=1.1; x2=2.1; x3=3.1;
Can anyone check the minimum?
Not on its own. But without optimisation of domestic and not only domestic expenses, for example, all the more so.
And in general, if you do not need optimization - this branch is not for you.
In fact, if ... - this thread is not for you.
With these words from a "master" and a major "expert" on "optimization" systems -- any discussion ends