Machine learning in trading: theory, models, practice and algo-trading - page 1819

 

It is possible to make, artificially, several sampling modes with different distributions and random transitions between modes. This will capture more interesting laws than simple random sampling

etc.

 
Maxim Dmitrievsky:

Yes it's all the same theme, randomly searching for predictor-target dependencies

trades are opened randomly, then into a neural network. A lot depends on competent sampling

For example, with this I found dependencies between TFs, as you suggested

Thank you, I will look at it.

If i have a trade with lags backward and forward for how many bars? or another way?

 
Valeriy Yastremskiy:

Deal with lags back and forth for how many bars? or something else?

A simple example:

with 0.5 probability we sell or buy on a new bar (or don't trade)

on the next bar with 0.5 probability we close or hold our position

if it is closed, go to point 1, if it is not closed wait for the next bar and look at the probability until the position is closed

This is the simplest example.
 
Maxim Dmitrievsky:

It is possible to make, artificially, several sampling modes with different distributions and random transitions between modes. This will capture more interesting laws than simple random sampling

etc.

What kind of patterns are we looking for between? Seasonal or temporal are logically inappropriate here. Random search for correlations doesn't really work either. Besides the aim we have is lead, then there is a sense, if there are some regularities, but lead is intermittent and quite random, it won't give anything. We will have a prediction or a lag.

Although if they are found, there will be something to optimize)).

 
Maxim Dmitrievsky:

A simple example:

with 0.5 probability we sell or buy on the new bar (or don't trade)

on the next bar with 0.5 probability we close or hold the position

if it is closed, we go to point 1; if it is not closed we wait for the next bar and look at the probability until we close our position.

This is the simplest example.

No, not the question that in the NS. If I look at the time and price of the deal or the next bar clauses too?

 
Valeriy Yastremskiy:

What kind of patterns are we looking for? Seasonal or temporal patterns are logically inappropriate here. Random search for correlations is not very good either. Besides, our goal is to get ahead of the curve, then it makes sense, even if there are some regularities, but getting ahead is intermittent and quite random, it won't give us anything. We will have one prediction or lag.

Although, if they are found, there will be something to optimize)))

between the predictors and the direction of the deal. The predictors may be anything, for example the price increments.

 
Valeriy Yastremskiy:

Nah, that's not the question in the NS. Only the time and price of the trade or the clows of the bars next to it too?

Increments, indicators, anything on entry. For the exit the direction of trades.

 
Maxim Dmitrievsky:

between the predictors and the direction of the deal. Predictors can be anything, for example price increments

Any of them is too much)))) We have only the averaging except for the increments))))). It would be logical to understand how many bars backward is optimal.

 
Valeriy Yastremskiy:

(Any of them is too much)))) We only have averaging apart from the increments))))). I think it would be logical to understand how many bars backwards are optimal.

Although, I still miss volumes.

 
Valeriy Yastremskiy:

(Any of them is too much)))) We only have averaging apart from the increments)))))). It would be logical to understand how many bars backwards are optimal.

This is another selection block, not related to sampling. The maximum backward bars are taken, and then the best combinations are selected. This is a problem of qualitative approximation, not sampling.

As a result, if properly sampled, the optimal number of lag predictors is selected that maximally describe the direction of trades.

Then it is monte carved and the best options are chosen. I don't think anyone has done it here, but I've had it for a long time.) Laws look for ra-two, if any... but there's no limit to perfection :)

want a more interesting sampling now