Machine learning in trading: theory, models, practice and algo-trading - page 786

 

About states - that's very right.

I once wrote an ephemeral example like this. There's like a lake, there's fish swimming in it, and the fins are catching the surface, so there are waves, and the wind is blowing. And we want to guess when the next wave is coming.
We may not know about the fish and their movements, about the wind and everything else, but we know that THERE IS, and the waves are just a consequence of THOSE movements. Who are THEY? I don't know, but we can analytically try to determine their position and trajectory, the effect on the waves.

In the end, ephemerally we have a process (waves) with hidden states (fish).
Or specifically we have a price with puppeteers.

 
Dr. Trader:

About the states - that's very correct.

I once wrote such an ephemeral example here. There is like a lake, there are fish swimming in it, they touch the surface with their fins, and that creates waves and wind blows. And we want to trade those same waves.
We may not know about the fish and their movements, about the wind and everything else, but we know that THERE IS, and the waves are just a consequence of THOSE movements. Who are THEY? I don't know, but we can analytically try to determine their position and trajectory, their influence on the waves.

In the end, ephemerally we have a process (waves) with hidden states (fish).
Or specifically we have a price with puppeteers.

Well, that's what the whole RL you hate is based on :)

There is an unknown environment and an agent who tries to do something in it and gets experience, looking for patterns, moving from state to state

 
RL is an overfit, the most brutal and merciless. It has none of the necessary properties for predicting non-stationary time series.
 
Dr. Trader:
RL is an overfit, the most brutal and merciless. It has none of the necessary properties for forecasting non-stationary time series.

Thank you!

I almost didn't get into it.

 
SanSanych Fomenko:

Thank you!

I almost didn't get into it.

You're relying on bullshit judgments without even getting into the subject.

Of course there's nothing presented on a silver platter.

And in your supervised methods you don't have a hard overfit, you'd think :)) overfit or 0 accuracy

 
Anatolii Zainchkovskii:

I think you can train on each bar, then look at the result of trained forward prediction, and if there is a pattern that at a certain time, the forecast works better, then use only this time range in the future.

If you work in a certain window, then you have to train at a certain time. The rest of the bars between windows are garbage. I was thinking, since the article about BW is unlikely to see the light of day, and there is a good introduction. I'll ask Rashid and if he gives me a chance, I'll post it here, at the Buy and check the predictive ability of the regression.

 
Aleksey Vyazmikin:

The important thing is not what will be in X bars, but what was in 10 bars, that is, if in 10 bars reached X points, then open.

You're confused. I'm sorry. I sincerely ask you to adequately formulate your thoughts and arguments against, if any. You're absolutely right, we will look back 10 bars to get a prediction 10 bars later. This is how all TS-NSs are usually built.

We make the forecast first. We get its value and then we decide what actions to take at this value or that one....

 
Mihail Marchukajtes:

If you work in a certain window, then you need to teach at that time. The rest of the bars between the windows is garbage. I was thinking, since the article about BO is unlikely to see the light of day, and it has a good introduction. I'll ask Rashid and if he gives me a chance I'll post it here, at BO and check the predictive ability of the regression.

Why do I need an authorization? Just post it in your blog. And here the link.
 
Mihail Marchukajtes:

You've got your pooch all mixed up. I'm sorry. I'm begging you to adequately formulate your thoughts and arguments against them if you have any. You're absolutely right, we will look back 10 bars to get a prediction 10 bars later. This is how all TS-NSs are usually built.

We make the forecast first. We get its value, and then we decide what action to take at this or that value....

Hmmm.... I guess I'm thinking how to make money, and the system will make predictions about the position of one bar relative to another... I do not understand why we do not want to work with take profit, while laying the logic of the system.

 
Maxim Dmitrievsky:

You rely on all sorts of delusional judgments without even getting into the subject

Of course there's nothing presented on a silver platter.

And in supervised methods you don't have a hard overfit, you'd think :))) overfit or 0 accuracy

In teacher models I know what an overfit is and how to deal with it.

In you I have never seen anything about overfit and the first judgement on the subject is Doc's, and he doesn't throw words to the wind and is very well versed in overtraining.


So let's have a concrete refutation of Doc's words, without emotion.

We should use the first file, preferably with cross validation, and then look outside this file. And it is desirable that the transactions were more than a hundred.