Machine learning in trading: theory, models, practice and algo-trading - page 3126

 

kozul inference for housewives.

study better for the brave and cowardly :)

 
You just need to exercise your maths skills within the limits of a senior kindergarten class and count the number of tails)
 
Maxim Dmitrievsky #:

kozul inferens for housewives

study better for the brave and cowardly :)

Just need to combine causal inference ideas with San Sanych's ideas about stability. Then the Grail is inevitable 🤑

But it's not for sure)

 
Maxim Dmitrievsky #:

kozul inferens for housewives

study better for the brave and cowardly :)

Maxim, you have already firmly learnt that the tail should wag the dog ))))

Keep on convincing others. I've had enough. Bye-bye.

 

I've noticed an interesting thing here.

Everyone knows about data drift. We are used to kicking only predictors, but I decided to see what happens to the strategy itself over time.

I took the data of one strategy that gives an entry signal on the crossing of 23.6% of the daily ATR(3).

So, I did the math in each month:

- Number of all signals

- Number of positive signals (1)

- Percentage of positive signals of all signals(TP)

I smoothed the resulting numerical series with a moving average with a value of 6.

So, what we got.

On the first chart we can see that the number of all signals of the basic strategy is growing over time.


Diagram 1.

On the second chart, we can see that the number of positive signals of the basic strategy grows over time, but at a slower rate.


Graph 2.

On the third chart, we see that the percentage of profitable signals out of all signals stagnates.


Diagram 3.

Probably, similar dynamics can be seen in the splits (Q cuts) of predictors....

 
Aleksey Vyazmikin #: Graph 1.

In the second graph, we see that the number of positive signals of the basic strategy is increasing over time, but at a lower rate.


What is the conclusion? That the market becomes more efficient over the years? Or is the model losing efficiency?
 
Aleksey Nikolayev #:

Just need to combine causal inference ideas with San Sanych's ideas about stability. Then the Grail is inevitable 🤑

But that's not accurate)

Especially when they don't want to state the essence of the process of determining stability
 
Uladzimir Izerski #:

Maxim, you have already firmly learnt that the tail should wag the dog ))))

Keep on convincing others. I'm done. Bye-bye.

I wish you'd done it right away, because he's suffering, poor man.
 
Forester #:
What's the conclusion? That the market gets more efficient over the years? Or the model loses its efficiency?

I think that given the strategy, we can tentatively conclude that the market has started to change trend more frequently intraday.

The question is whether there are factors in the history that have simply become more frequent now, and then maybe they can be predicted or even plotted as a linear increase in probability bias.

Or they are completely new events (combinations of predictor indicators) that have not been seen before.

One thing is obvious that we need a different way of building a model that takes into account the dynamics of the situation. Then we can try to explain the change of quantum segment probability due to appearance/reinforcement of other factors, and try to predict these other factors beforehand. In other words - it is necessary to understand what has changed and whether this change can be predicted, and further to take into account these changes in the final model.

 
Aleksey Vyazmikin #:

I think that given the strategy, we can tentatively conclude that the market has started to change trend intraday.

The question is whether there are factors in the history that are just now appearing more frequently, and then maybe they can be predicted, or if there is a linear growth in the probability bias at all.

Or they are completely new events (combinations of predictor indicators), which have not been seen before.

One thing is obvious that we need a different way of building a model that takes into account the dynamics of the situation. Then we can try to explain the change of quantum segment probability due to appearance/reinforcement of other factors, and try to predict these other factors beforehand. In other words - it is necessary to understand what has changed and whether this change can be predicted, and further to take into account these changes in the final model.

What are we trading?

Trends?

Patterns?

Statistics of increments through garches?

First we need to decide, and then draw conclusions.