Principles of working with an optimiser and basic ways of avoiding fitting in. - page 10
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
So, I have already said that the tester allows you to shift the parameters relative to their extremum. What does it mean? Suppose, using the same tester, we have determined the values of some parameters, at which the Expert Advisor's profit is maximal at the time interval being tested. The market at that time interval is fully known and any of its measurable properties will determine its state precisely and correctly. After finding the optimum parameters giving the profit extremum with the help of the tester, we make a shift relative to this extremum and look at the change of the result. In the future, the opposite picture will occur: the market will move away from our extrema by some value and the extent to which this extremum is located in the wide range of profitability will define the stability of the system as a whole.
Now let's turn to a specific chart:
The group of values located in a completely different place and generating profit less than the extremum is of real interest! The tester, on the other hand, knows nothing about it and if it is given the task to search for the maximum of profitability, it will find an insignificant statistical surge. In the future, even the smallest change in market parameters will result in us finding ourselves at the foot of that very peak! To avoid this we should focus on stable ranges rather than on specific values of individual combinations. There is only one way to find a stable diapason: by looking at the parameters relative to a fixed market footprint, forming a 2D or 3D stability map.
Indirectly, we can determine that we may be at the extremes of a statistical spike and not in the stability zone by the smoothness of the equity trend. We are intuitively more attracted to a stable flat positive result (trend) than to a jagged curve with a final balance even higher than that of the parameters generating steady growth. This is explained by the fractal nature of the process in question. If the process as a whole is unstable or random, we will observe the same unstable characteristics at any of its segments, i.e. scales: violent drops giving way to unexpected rises.
Here is a good enough illustration of this thought. I made the tester find the most profitable combination of parameters in a knowingly senseless (almost random) Expert Advisor. The result is a great final balance and something resembling an upward movement. Note that the test for the same parameters in the non-optimization sample (left half of the chart) was also successful. This is because time shifts alone are no guarantee of stability. A statistical spike can last much longer than our search window and OOS in this case simply says so. But there are other methods of tester shifts that allow us to properly analyse the results, but more about that later.
...After finding the optimum parameters giving the profit extremum with the tester, we make a shift relative to this extremum and look at the change in the result. ...forming a 2D or 3D stability map.
Reminds us of mountains and plateaus of stable parameter values:
Prog: 3D1V8 - with a description and my specific example from the owl optimization report. For visualisation of planar set selection of external variables included in owl - excellent option, IMHO. I use it myself.
Prog: 3D1V8 - with description and my specific example from the owl optimisation report.
By the way, the MT5 tester has a built-in display of optimisation results in 3D, well in 2D too, i.e. you don't even need external software to look at cliffs and plateauses.
In the future the opposite will happen: the market will move away from our extremums by a certain amount, and how wide the range of profitability is in this extremum will determine the stability of the system as a whole.
In the language of statistics it means stability of dispersion, and its value is a drawdown. The variability of this variance is the stability of TS.
Why don't we use the tried and tested ideology and set out our thoughts in climbing terms?
By the way, 3D is three parameters of the TS as I understand it, and if there are 4, what can be seen?
Translated into statistical language, this means stability of the variance, and its magnitude is the drawdown. The variability of this variance is the stability of the TS.
Why don't we use the tried and tested ideology and set out our thoughts in climbing terms?
By the way, 3D is three parameters of TC as I understand it, and if 4, what can be seen?
3D is two parameters relative to one metric. Any parameter can be a metric such as profitability, profit factor, expectation, etc.
2D tester charts are also three dimensional space, just the metric does not have its scale, and higher values of the metric are coloured in richer colour.
4D - this is where the difficulty of perception comes in. Multi-parameter EAs form multi-dimensional spaces. And they are not easy to analyse visually. I believe that we should use the split method: if there are 4 parameters, we build four true 2D plots. The Y axis is the value of the metric, for example profitability, the X value is the optimization of the parameter. The graph I presented above is essentially two dimensional, I just unfolded the surface so that the third dimension (2 optimization parameter) is not visible in perspective (like a 2D drawing). We analyse 4 plots, look for stable groups of values, then put them all together and see the result. The method is not perfect particularly as the parameters affect each other and finding their extrema separately is not the same as finding a stable group of values for all parameters simultaneously. But there is no exponential growth of enumeration and any simplest tester can cope with this task. If you have any suggestions on how this difficulty could be solved gracefully, please do it in a studio.
As I am without specialised education in this field, I am using my own handy tricks. I can not read books in bird's tongue, due to the lack of a doctorate in mathematics.
Reminds me of mountains and plateaus of stable parameter values:
Prog: 3D1V8 - with a description and my specific example from the owl optimisation report. For visualising the selection of a planar set of external variables included in the owl is a great option, IMHO. I use it myself.
Thanks for the walkthrough. I've been looking for something like that myself. It's just that my WealthLab is terribly glitchy and makes graphs intermittently.
In the language of statistics it means stability of dispersion, and its value is a drawdown. The variability of this variance is the stability of the TS.
Are there any statistical methods that allow us to search analytically for these stable plateaus?
But still agree, even having these analytical methods, a tester is still needed, which at least forms the result space on which these methods will work further.
Are there any statistical methods that allow us to search analytically for these stable plateaus?
But still agree, even having these analytical methods, the tester is still needed, at least to form that space of results, on which these methods will work further.
You can't do without a tester, because it provides statistics that should be analysed
In the past, no question. The variance must be stable. Deviations from the mo variance - the variance of the variance will give a measure of stability.