Machine learning in trading: theory, models, practice and algo-trading - page 2512
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
How is that? I'm thinking the same thing about searching through the grid, so I'm interested in the methodology already implemented.
Sometimes, as a crutch, it can pull a strategy close to a negative mathematical expectation.
I mean the metric, sometimes I evaluate a model not by profit, but by the dynamics of correct class predictions. Essentially the same balance, but the change is fixed. The point is that the strategy can be affected not only by classification accuracy, but also by fluctuations in market volatility, and we need to look at the dynamics of classification accuracy without monetary expression.
I just collected all my target, traits. The design variables were the parameters of the signs, the target one. I generated a trio with 2 indicators + a target one and trained it with a catbust. I selected them according to the maximum precision of learning on the test sample. I filtered the selected trios according to the possibility of the target one to give adequate signals for trading.
As a result, I have found 5 trios of features+target one. But as I've already shown, 93% prediction accuracy for a target giving good signal for trade is not enough. By the way, I tried to train full-link neural networks of different configuration with datasets for the found trios, random forest, and got +- the same training accuracy on a test sample and the same test results.
Good idea, thanks, I'll try to finish it.
I trade minutes, a fast market, so I know that in a second the price can go by more than my expectation.
I don't know how you can work on the minutes... And increasing the averaging window to smooth out the noise gives a picture close to a higher timeframe.
Can you give me a hint, what's the point of trading on 1-minute timeframes? Maybe I don't understand something?
In short, all in vain, with MO the market can not cheat.
I found the signs and the target, whose class distribution is shown in the first figure.
The accuracy on the test and training katbust models trained using this dataset was 93%
The second figure shows the balance and equity graph of the target trade:
The third figure shows the graph of balance and equity trading on the signals of the trained catbust model:
So, ladies and gentlemen, disperse.
The work you are doing is very close to my heart.
Could you do the following:
By the way, gradient boosting behaves erratically for some reason. Most likely retrained because of trying to get perfect at the expense of boosting.
I just gathered all my target, target traits. The design variables were the parameters of the traits, the target. Formed a trio of 2 traits + target trait and trained them with catbust. I selected them according to the maximum precision of learning on the test sample. I filtered the selected trios according to the possibility of the target one to give adequate signals for trading.
As a result, I have found 5 trios of features+target one. But as I've already shown, 93% prediction accuracy for a target giving good signal for trade is obviously not enough. By the way, I tried to train full-link neural networks of different configuration by the found trios, random forest, and got +- the same training accuracy on the test sample and the same test results.
Are you sure it's right to train different targets in the same sample? After all, the signals should be comparable, for example if the trend reverses and enters the trend from the flat, different set of indicators for the predictors should be expected.
There are questions about the catbust:
- Are you using only two samples, and you are not using the test sample at once?
- How many trees are in the model?
- do you use a training stop on the test sample?
- what is the learning rate?
- is class 1 responsible for the direction of the input, or the signal with the direction already set?
Good idea, thanks, I will try to finish it.
You are welcome :) It is also useful to look at the profit and accuracy of the probability distribution.
I don't know how you can work on the minutes. And increasing the averaging window to smooth out the noise gives a picture close to the higher timeframe.
Can you give me a hint, what's the point of trading on 1-minute timeframes? Maybe I do not understand something?
I trade Si - everything is quite technical there, plus there are more signals for learning.
Basically I use information from a lot of TFs, a lot of horizontal levels in predictors, and the minutes just allow me to react faster to an event that affects the price in my mind.
By the way, the gradient boosting behaves unstable for some reason. Most likely it overtrains because of attempts to get the ideal at the expense of boosting.
Here just did an experiment on this topic, so far with one sample, but the point is that increasing the learning rate has a positive effect on the result, as there is more generalization, which, when there is no correct answer to all questions and when the sample is not representative, is more effective than fitting on history for each example.
In fact, I use information from a lot of TFs, a lot of horizontal levels in predictors,
How do you use horizontal levels? The distance from the current price to the nearest hundred, 500, 1000?
How do you use horizontal levels? The distance from the price to the nearest hundred, 500, 1000?
I try not to use points - I use percentages. If I have a grid for the day (say ATR) I put predictors and opening price of the current bar in it, so I know where the price is relative to the level, preferably to mark events of level crossing, how long ago it was...
They are all retrained, we have almost random input.
Covariance and correlation are still there... (seems to be random).
We may not know how many basic components to follow in practice, some thumb rules apply.
After all, it's only
What is predictive modeling: Predictive modeling is a probabilistic process that allows us to predict outcomes based on certain predictors. These predictors are basically the functions that come into play in determining the final outcome, that is, the outcome of the model.
And what predictors and labels we choose and what algorithm we use is a matter of personal taste... - if we initially consider the system broader than the final Bull/Bear/Hold-on...
that's why the description of simulation possibilities by
Maxim Dmitrievsky
is always worth its weight in gold !
1. First the search area is determined: empirically or on the basis of assumptions, statistical tests are made. Then the MO algorithm is selected,
2nd variant -
2. Searching strategy through any classifier, analyzing its internal structure (feature importance, shap values and different metrics).
- only "in hope"..."by some polynomial"... + picking up useful data transformations (regardless of MO) is indeed an Art!..! They CANNOT be the same when analyzing any data (as some "gurels" here) - at least in linear and quadratic programming both functions are different and results are different and their Interpretation!