Machine learning in trading: theory, models, practice and algo-trading - page 2361

New comment
 
mytarmailS:

I try something like "one shot learning" but in my own way, or simply, I try to look for complex patterns...


I give one rebound and many "not rebounds" in the proportion of about 1 to 200, so it is like a training with one example, then I take probability from the model and look at the new data what happens to the price when the model shows higher probability...

It's almost the same as comparing the current price with my own pattern and look at the closeness measure, only here I look at the model probability...


Frankly speaking, sometimes it is very good, although there are not many deals but it is only one pattern and there may be many of them

Here is an example of one successfully found pattern, the first one is a kind of train, all the others are new data

Doesn't sound so bad to me.

How do you train? Are you using oversampling? Gradient descent classification can't handle 1 in 200 samples.

 
Aleksey Mavrin:

1) How do you train?

2) Do you use oversampling?

3) Gradient descent classification can't handle a 1 in 200 sample, can it?

1) Forest

2) no

3) You can do it with genetics

 
mytarmailS:

1) forrest

2) no

3) genetics can

Got it. Next you should probably ask for active learning, add the most "unsuccessful" examples from the OOS to the tutorial and finish learning.

 
Aleksey Mavrin:

Got it. Next, you're probably asking for active learning, adding the most "unsuccessful" examples from the OOS to the tutorial and learning more.

No..

Then the next step is to generate the correct rules that suit the market data.

I am attracted to symbolic regression (genetic programming) as a tool, but it is too resource-intensive, so I am still thinking...

 
mytarmailS:

No..

Next, generation of correct rules, which are suitable for market data.

I liked symbolic regression (genetic programming) as a tool, but it eats up a lot of resources.

Aren't you tired of eating a cactus? )

Imagine looking for patterns on SB

 
Maxim Dmitrievsky:

Imagine looking for patterns on the SB

I've got an eye for it.)

I want to soar on the wings of creativity)

 

If you train a model on 10 primitive signs and want to describe the whole market, then it's okay.

And if I describe one situation with a hundred signs and a whole model, then it's SB?

Are you sick?

 
mytarmailS:

Cheering?

Cheer for you. Break the back of forex, break it completely.

 
Aleksey Nikolayev:

Cheer for you. Break the back of forex, break it completely.

))

At least he has a good sense of humor.

 
mytarmailS:

If you train a model on 10 primitive signs and want to describe the whole market, then it's okay.

And if I describe one situation with a hundred signs and a whole model, then it's SB?

Are you sick?

All the signs are derived from the price.

all you do is create a bunch of rules/trees that MO can do better than you

take ROCKET and make a bunch of traits, especially since there's a miniRocket

well, either keep humping your horns in the hope that the whole world is sick and you're d'artagnan

MiniRocket: Fast(er) and Accurate Time Series Classification
MiniRocket: Fast(er) and Accurate Time Series Classification
  • Alexandra Amidon
  • towardsdatascience.com
Most state-of-the-art (SOTA) time series classification methods are limited by high computational complexity. This makes them slow to train on smaller datasets and effectively unusable on large datasets. Recently, ROCKET (RandOM Convolutional KErnel Transform) has achieved SOTA of accuracy in just a fraction of the time as other SOTA time...
1...235423552356235723582359236023612362236323642365236623672368...3701
New comment