Machine learning in trading: theory, models, practice and algo-trading - page 2059

[Maxim Dmitrievsky]

Rorschach:

It's hardly interesting in its pure form.

Well, I've looked at them. Either few deals, or something like yours. Tried adding conditions - then retraining

[Igor Makanu]

Rorschach:

We buy on a certain day of the week and a certain hour with different SLs and TPs. For 10 years there are systems in the plus, where the profit is several times greater than the maximum drawdown.

The same system has been running by the window for a year and a month. It is obvious that some combinations of parameters appear more often than others.

Question to experts: Dear experts, I want to "scientifically" prove the significance of the result. For this purpose the law of large numbers and 4*sko are used. Can it be used for the TS with unequal SL and TP, and how many trades are needed?

it depends on your goal

- If you want to prove your case to the forum, then I think 100+ trades per year + test for? dunno... 3-5-10 years?

- if from a purely scientific point of view, then.... Do you understand the process you are researching?

ZS: ... on your fingers, you did some statistics, for example you researched that during the last 10 years the number of carbureted cars was decreasing and the number of injector cars was increasing, got ratios of growth and decline, made a forecast for 10 years ahead..... and here Ilon Musk with his Teslas took and messed with your statistics and made a mess of it

[Rorschach]

Igor Makanu:

It depends on the goal

- If proving yourself right to the forum, I think 100+ deals a year + test for? dunno... 3-5-10 years?

- if from a purely scientific point of view, then.... Do you understand the process you are researching?

ZS: ... on your fingers, you did some statistics, for example you investigated that during the last 10 years the number of carbureted cars was decreasing and the number of injected cars was growing, got ratios of growth and decline, made a forecast for 10 years ahead..... and here Ilon Musk with his Teslas took and messed with your statistics and made a mess of it

You can take a randomness test to begin with.

I don't like all kinds of sharps.

[Rorschach]

Maxim Dmitrievsky:

Well, I've looked at these. Either there are not enough deals, or something like yours. I tried adding conditions - then retraining.

I can't get enough of these pictures.

Entering twice a week at a certain hour. The bill is not a penny.

[Rorschach]

Rorschach:

You can pass the randomness test to begin with.

It won't, because there will clearly be a correlation between the increments.

[[Deleted]]

Rorschach:

You can take the randomness test to begin with

I don't like all kinds of sharps.

https://ru.qaz.wiki/wiki/Randomness_tests

In this case,"take the randomness test" is, pardon me, complete nonsense, nonsense.

[Andrey Khatimlianskii]

Rorschach:

I can't get enough of these pictures.

Entrance twice a week at a certain hour. The accounts are not penny stocks.

And if something went wrong, averaging with a x10 lot, judging by the dangling snot?

[Aleksey Nikolayev]

Rorschach:

We buy on a certain day of the week and a certain hour with different SLs and TPs. For 10 years there are systems in the plus, where the profit is several times greater than the maximum drawdown.

The same system has been running by the window for a year and a month. It is obvious that some combinations of parameters appear more often than others.

Question to experts: Dear experts, I want to "scientifically" prove the significance of the result. For this purpose the law of large numbers and 4*sko are used. Can it be used for TS with unequal SL and TP, and how many deals are needed?

There is an implicit problem with estimating significance here - biased sampling estimation. The reason is that the best of all possible variants is selected. Here is a simple model example.

Let the return on each transaction have a standard normal distribution (zero mean and unit variance). The number of different series of deals is 120 = 24 hours * 5 working days. Each series has 150 trades - about three years. The code in R:

nw <- 150
nts <- 120
bst <- rep(0, nw)
for (i in 1:nts) {
  tst <- rnorm(nw)
  if (mean(tst) > mean(bst)) bst <- tst
}

mean(bst)
plot(cumsum(bst), type = 'l')

result:

best-pass average = 0.3418549

equity of the best pass:

Naturally, this is not a guarantee of no possible profit at real prices)

[Evgeny Dyuka]

All the charts in this thread start in the lower left corner and end in the upper right ))
I wish I lived that way...

[Rorschach]

Andrey Khatimlianskii:

And if something went wrong, averaging with a x10 lot, judging by the dangling snot?

About that. There are a lot of things to pick on, the author, the broker's statement, the tester's report.