Machine learning in trading: theory, models, practice and algo-trading - page 1042
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
As you get tired of beating the ice, think about the fact that both price/time scales are not inherently linear. This is if approached from the perspective of pure algo-trading (without understanding the market).
Time is what we measure periodic processes with. Time makes little sense to measure processes that have a random nature of occurrence.
The time scale at small (tick, "quantum") intervals is nonlinear and random, it seems that for events of tick dimension, time as a significant factor does not exist at all.
At large intervals, due to the superposition of daily, weekly, news release-related and other periodic heterogeneity, the time scale can be considered more close to linear and the significance of time increases.
You can't. Check the currencies for the Hearst index. It clearly shows the randomness of the market. And what can be done in a random market? Only martin. But on the other hand, inefficiencies of different time of existence appear in the market. And people make money on them. And this is no longer random. That's the direction in which inefficiencies should be searched for. I would like to automate this process. But I cannot feel what to start from. Neural networks are not suitable for this. They need ready-made patterns for learning.
Why not build on what obviously exists and works, what has been saving our plan for billions of years and what helps algotraders optimize and tweak their EAs - the inertia, the market memory?
In the neighboring topic, right in the first post it is written, that no tricks can destroy non-markovian pricing, although the rest of their discussion, around tick distributions and Integro-diffusions, IMHO is suitable perhaps only for research of filters for dataphids, but we are in the MO theme :)
And neural networks, IMHO, are the best for this task...
Yooooooooooo.... Where did you ever see a branch drop below the first page????? Yes.... you guys got it up and running... you guys got it up and running. In the meantime, I am doing well and the results are quite encouraging and all thanks to an accidental mistake :-)
It was like this.....For example...
And now it's like this.... Can anyone tell if the data has improved or not????
There are really a lot of questions. How to interpret the principal component graph????? Still a question These two sets of data are taken over the same time period. The target is the same, but saving the predictors was done in two different ways. So. Your exit statists, this task is just for you!!!!!
Which of the given data sets is better??? The first or the second. Your opinion gentlemen!!!!!
And then I'll give you my opinion...... okay???
I couldn't get through two pages of this thread.
I will just give my opinion. And so machine learning is a set of statistics on the tool, analysis and the algorithm itself by the result of the work done, but... There is one important note, no algorithm can guarantee that you will get the desired result if all the conditions are fulfilled, and that means no matter how much data you analyze and complicate the algorithms for decision-making, there will always be a probability of the expected outcome.
It is the probability that you trade, and as a consequence you need to look for the outcome of higher odds. The market itself is monotonous, based on my analysis (not exact, but it was not accurate for both longs and shorts), the market for the same time interval gives approximately the same (49%/51% or 51%/49%) number of profitable deals in both directions.
And so, the algorithm for making a decision should start from the highest estimate (qualitative) probability of the outcome with additional filters at your discretion.
Yooooooooooo.... Where did you ever see a branch drop below the first page????? Yes.... you guys got it up and running... you guys got it up and running. In the meantime, I am doing well and the results are quite encouraging and all thanks to an accidental mistake :-)
It was like this.....For example...
And now it's like this.... Can anyone tell if the data has improved or not????
There are really a lot of questions. How to interpret the principal component graph????? Still a question These two sets of data are taken over the same time period. The target is the same, but saving the predictors was done in two different ways. So. Your exit statists, this task is just for you!!!!!
Which of the given data sets is better??? The first or the second. Your opinion gentlemen!!!!!
And then I'll give you my opinion...... OK???
Michaelo got to the PCA... why, are your hands itchy? )
the graph of the principal components should be interpreted in an orthogonal basis :D
red ones are sort of orthogonal predictors, and what do the numbers mean?
Michaelo got to the PCA... why, are your hands itchy? )
The graph of the principal components should be interpreted in an orthogonal basis :D
So which one is better?
So which one is better???
well the second one, 55%
Well the second one, 55%.
There is no 55% on the principal component plot(the first plot). The 55% is the clustering graph, where in both cases the data represent two well-divisible areas. One is better than the other and let's go back to the first graph. Why is the bottom one better than the top one????
For this you need to know how to interpret them!!!!
There is no 55% on the principal component plot(the first plot). 55% is a clustering graph where in both cases the data represent two well-divisible areas. One is better than the other and let's go back to the first graph. Why is the bottom one better than the top one????
You have to know how to interpret them!!!!
well if the numbers are points then the variance on the 2 components is lower than on the 1st, no?
http://setosa.io/ev/principal-component-analysis/
You can read the graphs with your mouse and see everything becomes clear at once.
well, if the numbers are points, then the variance on the 2 components is lower than on the 1, no?
http://setosa.io/ev/principal-component-analysis/I agree!!!! But that's not all... It turns out that the second graph is better because there are such vectors that are as close to the zero axes as possible. In this example, it's not so obvious, but now we have such datasets where the vectors of components coincide with the zero axes and divide the field into even 4 squares. In the first case, the component axes are scattered along the dioganals between the zeros, while in the second picture there are such component vectors that are as close to the zero lines as possible. Knowing the name of the predictor we train the optimizer as long as the inputs are those predictors that form the component vector closest to the zero axis and it doesn't matter in what direction. Again, this is my IMHO!!! That's why I wanted to clarify how right I am!!!!