Machine learning in trading: theory, models, practice and algo-trading - page 2793
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
Evaluate Walking Forward with a test.
It's an evaluation of the whole flock. And lousy sheep are culled piece by piece.
It's an evaluation of the whole flock. And lousy sheep are culled piece by piece.
At 500 bars to estimate is not a statistic at all, you can fit anything, by the law of large numbers
50 fics = 50 valking fovard tests with fics removed 1 at a time. It is long, but the result will be obtained by the model.
This way you can get the result only in case of complete independence of features, and it doesn't happen that way.
At 500 bars to estimate is not a statistic at all, you can fit anything, by the law of large numbers
To evaluate the predictive ability is quite enough. It is possible to select fiches that give the teacher's prediction error up to 20% using the sliding window technology.
You can only get results with this method if you have complete feature independence, and it doesn't work that way.
You feed the same data into your packets. You can't get anything either?
In preprocessing, as a step, I remove correlated chips. Out of 170, about 50 remain if correlation is not higher than 75% (!). When correlation is not higher than 50%, a few pieces remain. But I didn't set the goal to collect NOT correlated fiches.
In preprocessing, as a step, I remove correlated features. Out of 170, about 50 remain if correlation is not higher than 75% (!). If correlation is not higher than 50%, a few pieces remain. But I didn't set a goal to collect NOT correlated fiches.
Those 50 are the ones you can check with the model.
So they're correlated! The result depends on the order in which the features are discarded.
For evaluating predictive ability, it is sufficient. It is possible to select fiches that give teacher prediction error up to 20% using sliding window technology.
You can fit 500 bars, no doubt, if you predict for 1 trade. You need statistics, I don't believe that it will predict better than random on average. But the variant has the right to life.
regarding the selection of features and multicollinearity (I specifically asked a question to the developers of bousting) - it makes sense to select only in case of contests, to get cleaned models and in the struggle for fractions of %. In all other cases it makes almost no sense to do such preprocessing. They zero out the rubbish perfectly well by themselves.