Discussion of article "Exploring Seasonal Patterns of Financial Time Series with Boxplot" - page 24
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
No hourly stats? I couldn't see it.
I.e. the initial hypothesis: in certain hours, the increments, on average, grow, i.e. in these periods the maximal profit will be reached. If we get hourly statistics of profits through optimisation, we can confirm/refute the hypothesis (without adding new parameters to the TS).
And if the optimiser finds other hours, it will be a fitting. This is very likely to happen when the number of TC parameters increases, i.e. genetics will trade in the wrong place and will not find anything from the hypothesis (as an option).
Almost regular optimisation/trading mechanism was running, no collection of statistics by hours is not built into this WFO. This is another task, but it seems to be not very difficult - whoever finds free time first will be the best ;-). For now you can focus on the parameters inStartHour; inCountHours, found by WFO (in the last column) - somewhere there the system suggests trading from the evening. But in a good way, the EA should be built in forced closing via inStartHour + inCountHours, because now it often outlasts losses (especially given the large periods of mashka).
1. without initial assumptions (statistical research) there is no optimisation, there is nothing to optimise. You can optimise your fantasies, but then the result will also be in the form of pink ponies
2. Don't confuse the general statistical study with the statistics obtained during the optimisation. Otherwise the tail starts wagging the dog.
It is not a bad idea to give links to citations.
1. Optimisation is applied to the System. The existence of a System has nothing to do with Research or Statistics. A strategy is a system interacting with market parameters. Its optimisation is a search for the best values in the chosen time frame. The result of optimisation is the values of parameters (presumably) providing maximum deposit growth.
2. The GA itself already contains a statistical study. But it is hidden from us and automated in the algorithm itself. Strategy statistics is different. It is the result of optimisation, not a part of its mechanism.
Theessence of optimisation is the search for parameter values that give the best value of the target parameter.
Theessence of GA is a method of compressing the area of search for values that give the best value of the target parameter.
Theessence of statistical research - accumulation of data, identification of relationships of Values, Events, Processes and analysis of their repetitions.
Theessence of regularity is the connection of objects identified and confirmed statistically.
Statistical research is used in the search for parameter values, i.e. optimisation. The study is embedded in the genetic algorithm. In the tester this process is automated and we see only the result.
Statistical study is used in the search for parameter values, i.e. optimisation. The study is embedded in the genetic algorithm. In the tester this process is automated and we only see the result.
Statistical study is not used in the search for specific values of specific parameters. It is used to find the parameters themselves and their distributions.
Optimisation is used to maximise the target function, that is why it is optimisation. Statistical research does not always lead to optimisation, but only when it is necessary to optimise some process on the basis of this research.
Once again, I suggest you go here to close this question at last.
Good article, easy and pleasant to read.
It would take a long time to find such a regularity by optimisation head-on (although you can try to get the GA on it).
But after macro-analysis, when you can already see "ponds with fish", it will be faster and easier to find specific parameters by the optimiser and check them on the forward.
Median position in the dispute "saber-Dmitrievsky" ;)
And why have stirred up a fight - it is not clear.
Good article, easy and enjoyable to read.
It would take a long time to find such a pattern in a head-on optimisation (although, you can try to get a GA on it).
But after macro-analysis, when "ponds with fish" are already visible, it will be faster and easier to find specific parameters by the optimiser and check them on the forward.
Median position in the dispute "saber-Dmitrievsky" ;)
And why have stirred up a fight - it is not clear.
I have the same position myself.
Scuffle because of false concepts that clog the brain, as if GA can do everything and even statistics, and there is no need to look anywhere.
And so, on the material you can already stamp bots in packs and optite themselves for joy.
thanks
Fighting over false concepts that clog the brain, supposedly GA can do everything and even statistics and you don't have to look anywhere.
I think it's more about misunderstanding.
I think it's more of a misunderstanding.
If you add free terms (parameters) to the TS, the GA starts to screw up hard with the choice of optimal ones, I have checked it more than once. Each additional parameter is an additional dimension in Hilbert space. And then it is impossible to interpret it (and what should be changed at all to make it work on OOS). Then you will understand where to look for it.
But such a simple example as in the article, of course, works perfectly.
If you add free members (parameters) to the TS, then the GA starts to mess up hard with the choice of optimal ones, I have checked it many times. Each additional parameter is an additional dimension in Hilbert space. And then it is impossible to interpret it (and what should be changed at all to make it work on OOS). Then you will have to understand where to look for it.
But such a simple example as in the article, of course, works perfectly.
Well, you can also optimise in different ways.
If you stupidly switch on the search of everything over the whole range with a small step, then, of course, you will fail.
But you can also optimise analytically: clamp all parameters in an intuitively correct position, search roughly one logical block, select its range of values. Clamp this block in the middle of the optimal values, and then optimise the next one. When reaching the last one, fix the already strongly narrowed ranges for all of them and sample in smaller increments.
Or see the patterns by hours, having checked them separately (and on each of them having looked at different MA parameters), then leave only the promising hours (1-4) and option the rest in detail on them.
The same analytics, but in profile.
I think saber meant something like that, not a stupid bruteforcing of everything.
Same analytics, but in profile.
Yes, one argues that multiplication on paper and pencil is static analytics, others argue that if you press a couple of buttons on a calculator, the result will be the same.
and those who know how to multiply with a column believe that those who multiplied on a calculator do not understand the essence of what is happening.
well, and then, as usual, a massive holivar with appeals to universal kneeling before those who shout louder, otherwise you will continue to calculate everything on a calculator, and the topic of "integral over a field" is not yet solved and your calculator will be powerless there.
Well, there's a lot of different ways to get back at it.
If you simply switch on the search of the whole range with small steps, then, of course, it will be a bummer.
But you can also optimise analytically: clamp all parameters in an intuitively correct position, search roughly one logical block, select its range of values. Clamp this block in the middle of the optimal values, and then optimise the next one. When you reach the last one, clamp all the already strongly narrowed ranges and optimise in smaller increments.
Or to see the patterns by hours, having checked them separately (and on each of them having looked at different MA parameters), then to leave only promising hours (1-4) and to opt the rest in detail on them.
The same analytics, but in profile.
I think saber meant something like that, not a stupid bruteforcing of everything.
There's no way to know what he meant, he hasn't written anything about it
Usually, his "research" ends with "I found something, it works for some reason. No conclusions."
In machine learning there is a clear division: Exploratory analysis and model(algorithm) creation. Doing Exploratory through genetics is something that was invented only on this forum.
For example, you can look at OOS and see dataset shift (covariate shift etc.) in the data, relative to the opt. subsample, which is shown in the article. Saber suggests opting until blue in the face to find a set that works well on both subsamples, without understanding what changed on that plot. This is nothing but monkey labour.