Machine learning in trading: theory, models, practice and algo-trading - page 737
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
What's that got to do with it, you're just being told what you're doing wrong and that there's no fish there. You need to be more flexible and perceive the information, change the paradigm, shake up the picture of the world and so on.
Look... The models have already worked for a week... Haven't touched them with my finger and I plan to keep them running for another week or two... Let's see....
And yes, they are trained on no more than 40 examples...... Whether it's right or wrong, I don't know, but it makes a profit.... And whether it's obtained the right way or the wrong way doesn't matter... The important thing is the final result and the point.....
... And whether it is obtained in the right way or in the wrong way doesn't matter. The important thing is the end result and the point.....
You can't argue with that.
The equity curve has to be long enough, at least a thousand trades and if we look at Sharpe Ratio instead of No Returnee, then everyone will agree, and if 10 trades or even 100 with a profit less than the drawdown is not serious, the ones who "escalate the deposit".
I think even such a variant will not be immune to accidents and falsification. Ideally, to test the methods of machine learning, you need to offer traders a simple and accessible tool that allows them to generate models and check them on the forward.
The equity curve should again be long enough, at least a thousand trades and if you look at Sharpe Ratio, not no returnee, then everyone will agree, and if 10 trades or even 100, with a profit less than the drawdown is not serious, these are those who "escalate the deposit".
Where are these numbers coming from? 1000 trades. You have to understand, there is no grail. The market is too volatile, and when I see a value of 1000 trades I understand that this is the closest approximation to what does not exist (the Grail) and what's wrong with statistics of 100 trades? Here's an example... Yes I already posted it. The model was trained on 40 points and worked for 80. What do you mean this crank's stats are lousy???? Not enough profit? Aren't the requirements for the TS too high, Toxic????
Not to mention the fact that we learn about the quality of the model immediately after training and we do not need to allocate another mini OOS plot to see how it works. This becomes known after obtaining metrics trained models. That is, trained several models, calculated their metrics, chose the one with the best performance and immediately into battle, without delays.
The choice of model is very simple. If an output variable has an equal number of zeros and ones, the entropy of such a variable is usually 0.7 if rounded. That is, the uncertainty of the output is quite high. The model presented in the post above had an VI of 0.87. This is the sum of the VI of the two polynomials. And all this on the training plot. Before putting the model into operation. Having received this value, I immediately realized that it was the model, because all other models derived from the same training file were either 0.7 or close to it. By the way, such models at the same section of the OOS were about zero, and also models with the result of 0.6 and lower. These merged.
Hence I draw a conclusion. A working model is a model whose mutual information index (MI) with respect to the output is greater than the entropy of the output itself. The output uncertainty is 0.7 and the VI of the model result is 0.87. That is, the model knows more about the output than the uncertainty of the output itself. That's when your model trained on 40 values will work much longer than the training interval. Anyway I deduced exactly this fact.....
And if VI is more than unity or slightly lower within the range of 0.95, then this is a clear sign of overtraining. I manipulated the data and purposefully retrained the model. So all the VI was greater than unity. Here's a fact and a thought for you....
The choice of model is very simple. If an output variable has an equal number of zeros and ones, the entropy of such a variable is usually 0.7 if rounded. That is, the uncertainty of the output is quite high. The model presented in the post above had an VI of 0.87. This is the sum of the VI of the two polynomials. And all this on the training plot. Before putting the model into operation. Having got this value, I immediately understood that this was the model, because all other models derived from the same training file were either 0.7 or close to it. By the way, such models at the same section of the OOS were about zero, and also models with the result of 0.6 and lower. These merged.
Hence I draw a conclusion. A working model is a model whose mutual information index (MI) with respect to the output is greater than the entropy of the output itself. The output uncertainty is 0.7 and the VI of the model result is 0.87. That is, the model knows more about the output than the uncertainty of the output itself. That's when your model trained on 40 values will work much longer than the training interval. Anyway I deduced exactly this fact.....
And if VI is more than unity or slightly lower within the range of 0.95, then this is a clear sign of overtraining. I manipulated the data and purposefully retrained the model. So all the VI was greater than unity. Here's a fact and a thought for you....
Now it all comes down to the output (teacher). Is it profit? drawdown? profit factor? or something else?
Now it all comes down to the output (teacher). Is it profit? drawdown? profit factor? or something else?
I make several output variables by signal profit indicator. From -40 -20 0 20 20 40 60 80. This allows me to pick a data set that has the maximum number of important variables on the maximal sample. And what will be the output, even with a profit of -40 pips. I know it and I will try to get 40 points better at opening a deal than the close of the bar with the signal. This is a form of adaptive exit. At least for me so....
Again, the equity curve must be long enough, at least a thousand trades and if we look at Sharpe Ratio instead of No Return, then everyone will agree, and if 10 trades or even 100 with a profit less than the drawdown is not serious, those are the ones who "escalate the deposit".
He has a THINK about the relationship of target and predictors - mutual information and this thought can override all the shortcomings of statistics, because this thought has cut off all the noise among predictors, and this noise is the main evil.
Cleverest, is there sense to use this program https://basegroup.ru/deductor/download for initial understanding of neural networks (I am generally interested in finding patterns in series of numbers)? Just in this business I am very new, and I would like to have a software in Russian and with a visualization of the results of the grid (solution search).
Please look at it with a professional eye.