Machine learning in trading: theory, models, practice and algo-trading - page 3397

New comment
 

I also plan to experiment not with positional matchmaking, as in the article, but with Propensity score. This will allow to calibrate probabilities at the same time.

Then I'm going to use that google liba to see what I can get out of it.

Maybe later I'll roll out the results.

 
If there are negative values for deposit, profit or equity in the bottom lines of the optimisation parameters, when obtaining results, this result will definitely shoot someday. All optimisations are to obtain data for this period only. Although it is not sad, but it is so.
 
Maxim Dmitrievsky #:

I also plan to experiment not with positional matchmaking, as in the article, but with Propensity score. This will allow to calibrate probabilities at the same time.

In theory, you can search for and match samples through it

For example, randomly mark one piece of the sample as 0 and another as 1. Teach the NS to separate to classify which sample belongs to which sample. This is also called Adversarial validation.

Ideally, the NS should not identify the sample, the error should be around 0.5. This means that the original sample is well randomised.

propensity
0.38        3
0.40        3
0.41        3
0.42       20
0.43       27
0.44       40
0.45       56
0.46      140
0.47      745
0.48     3213
0.49     8041
0.50    11718
0.51     5324
0.52     1187
0.53      749
0.54      209
0.55       95
0.56       54
0.57       29
0.58       12
0.59       14
0.60        8
0.61        6
0.63        1

Anything in the neighbourhood of 0.5 is good, this data can be used for training. The extreme values are outliers.

Then for each "probability" you can calculate the percentage of guessed cases.

So far, it's a bit mind-blowing to take this approach.

 

An interesting ongoing contest - for those who want to compare their success in predicting quotes with other participants.

Numerai
Numerai
  • numer.ai
The hardest data science tournament on the planet. Build the world's open hedge fund by modeling the stock market.
 
Aleksey Vyazmikin #:

Interesting ongoing competition - for those who want to compare their success in predicting quotes with other participants.

So this link has been here many times already
 
mytarmailS #:
This link has already been here many times

I didn't remember it - I guess it wasn't clear then what to do, but now I read the help and it became clearer. Anyway, it's a fact that this idea has been working for a long time. As I understand it, they pay there with some kind of crypto for good forecasts.

The disadvantage, of course, is that the code is open and must be transferred for participation.

 

The future is here. I'm running Google's LLM locally. Now I don't need a wife and friends.


 
Maxim Dmitrievsky #:

The future is here. I'm running Google's LLM locally. Now I don't need a wife and friends.

https://blog.google/technology/developers/gemma-open-models/

Gemma: Introducing new state-of-the-art open models
Gemma: Introducing new state-of-the-art open models
  • blog.google
Gemma is designed with our AI Principles at the forefront. As part of making Gemma pre-trained models safe and reliable, we used automated techniques to filter out certain personal information and other sensitive data from training sets. Additionally, we used extensive fine-tuning and reinforcement learning from human feedback (RLHF) to align...
 

A good summarisation of the whole thread


 
Aleksey Vyazmikin #:

A good summarisation of the whole thread


Listened with pleasure
1...339033913392339333943395339633973398339934003401340234033404...3515
New comment
Reason: