Machine learning in trading: theory, models, practice and algo-trading - page 1382
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
No, I'm just saying that they are returns, instead of -1 you can put log(), the same thing will happen, i.e. logreturns. This loss of information is very significant, since you have only 20 of them.
For my problem 20 is enough. In general case you may need more. It depends on the specifics.
Imho, you can't put log - it's a non-linear transformation. The usefulness, in the general case, is very doubtful.
In general case I pass input series through sigmoid or tanh, so that the main prices were in the "linear" part, and non-linearity was limited only by outliers.
Why do you need predictors?
Don't you like the raw chart anymore?
You are not very knowledgeable in MO and algotrading, I have to advise as many people as possible not to buy your"grails" in the market, in view of the author's incompetence.
For my problem 20 is enough. In the general case, you probably need more. It depends on the specifics.
Imho, you can not put log - it is a non-linear transformation. The usefulness, in the general case, is very doubtful.
Log is needed, because the price changes not as a cumulative sum of changes, but as product, like compound interest, though in case of currencies where changes are very small, it is not the point.
For my problem 20 is enough. In the general case, you probably need more. It depends on the specifics.
Imho, you can not put log - it is a non-linear transformation. The usefulness, in the general case, is very doubtful.
In the general case, I pass the input series through sigmoid or tanh, so that the main prices were in the "linear" section, and non-linearity was limited to outliers only.
like NS will pass them through sigmoid by itself? )) or it gives something additional
like the NS will pass them through the sigmoid itself? ))) or is it something extra that gives
But not on the input). It gives limitation of outliers in the input row, we don't overload neurons at the input of the NS.
But not on the input). It gives limitation of outliers in the input series, we don't overload the neurons on the NS input.
You can also run it through an arctangent... Fisher's transform...
like making the distribution more normal... but what's left of the data is not up to me ))
you can run it through an arctangent... the Fisher Transform
kind of makes the distribution more normal... but what will be left of the data is not up to me ))
You can do it any way you want, as long as the central part of the output is linear in the main range.
All the data will be in place, except for the outliers, which will be a bit suppressed. Absolutely normal method of signal processing.
Otherwise, some spike in the history will clobber you all the neurons, and will clobber you until it's out of your sight. Your NS will simply go into denial or unconsciousness. )
The log is needed because the price changes not as a cumulative sum of increments, but as a product, like compound interest, although in the case of currencies where the changes are very small, it is not the point.
I assume that in a more or less stable market, M(dC/C) =~const.
I apply not division (P[i] / P[0]), but subtraction (P[i] - P[0]), i.e. not relative price change, but absolute. I preliminarily remove outliers (1% by quantity from the largest and smallest values).
Does division give any advantages? I am currently using a forest that does not need normalization and scaling.
I apply not division (P[i] / P[0]), but subtraction (P[i] - P[0]), i.e. not relative price change, but absolute. I preliminarily remove outliers (1% by quantity from the largest and smallest values).
Does division give any advantages? I am currently using a forest that does not need normalization and scaling.
It is a common opinion that the difference of price logarithms (the same as the logarithm of their ratio) is closer to the normal distribution than just the difference. For the same reason, to derive the Black-Scholes formula, price is modeled by geometric Brownian motion.