Machine learning in trading: theory, models, practice and algo-trading - page 1811
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
Discretization is nonsense, you can use regularization. The additional training of the model in the course of trading is also nonsense, it won't work.
That's too radical)))
That's pretty radical)))
You can't find a pattern through these ways, it's just zafit. How to stuff a slot with absorbent cotton tighter
Do-learning doesn't change the way you search, but it adds new data for learning. Why this is a bad thing.
Separation is more complicated, and there are no direct logics, so how can you be sure?
Do-learning doesn't change the way you search, but it adds new data for learning. Why is this a bad thing?
Separation is more complicated, and there are no direct logics, how can you be sure?
Because it's sliding window training, which is exactly the same as retraining. And it is impossible to control it with new data.
Control only by the fact of evaluating the last window, and then only when the new characteristics of the series have become significant and we got a lag. There are a lot of data and if we are serious about it then we should train on all data minimizing the lag. This is just a variety of series.
It is quite possible that the new data repeats another tool, which was not involved in the training.
Control only by the fact of the evaluation of the last window, and then only when the new characteristics of the row became significant and we got the lag. There are a lot of data today, and if we are serious, we should train on all data minimizing the lag. This is just a variety of series.
It is very likely that the new data repeats another tool, which has not been involved in the training.
I don't see any difference between training on all data and in the sliding window. If the characteristics of series change very smoothly, then it makes sense. But there is no such thing on the market.
It only makes sense in a bible of obtained characteristics, nothing more. It may not be a complete solution, but it works to find repetitions on different instruments at different times. And only as auxiliary data for the sliding window.
the meaning is only in the bibliography of the obtained characteristics, no more. It may not be a complete solution, but it does work to find repetitions on different instruments at different times. And only as auxiliary data for sliding window.
What characteristics? There will be an array of NS weights that are not interpreted
Yes, the incorrectness between desires and instruments. Indeed an array of weights with NS will not be sufficient. and it is hardly possible to get interpretations from them)
The characteristic of a series is the simplest possible mathematical model describing it with a sufficiently small error))))
Vladimir Perervenko:
...
It's a great illusion that you can train a model on a huge interval of past data and then use it for a long time without retraining.
...
How long is how long - I have a model that's been running at least half a year on the upside. I discovered this with about a month ago, when I was going through old file archives - took a model and it works, but now I don't know how it was trained - CatBoost model.
Another example - also built on leaves, most of which are collected on the 2014-2018 sample inclusive, a tree in February 2020, which is used as a filter, and this synergy works well in 2020.
However, I didn't risk putting money on it all - and that's my mistake.
Now haunted by the thought that just as I approach the criterion long and it all breaks down.