Machine learning in trading: theory, models, practice and algo-trading - page 1017
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
Why is everyone fixated on chips?
I've tried 1000 and 200 chips - there is very little difference in predictions. There is a much bigger difference in the data dump method. There is a very big difference between trying to predict a full 1 bar and 0.5 bar than reducing 1000 chips to 200.
Gianni:
Unfortunately secrecy is the main requirement here))) Itis about the protocol of exchange of obfuscated C++ models, that take raw data from exchange and produce forecasts, so I can take a model with description of its inputs and outputs, use it for a month or for how long without modifications (additional training, etc.) and draw conclusions (buy, rent, etc.)
Probably this is one of the best options - to present the model as an executable application, can be console, receiving data on stdin or csv files from the command line as this option can be easily, without any difficulties used and protected, through obfuscation by time, etc.
Numerai, yes.
Their way makes sense. I tried my models to predict hundreds of thousands of random instances. Then for "black box" prediction, I looked for the closest point by coordinates, and used its result as the prediction itself. This prototype worked, but we could improve it for real - find 3 closest points and triangulate the average result. But it's computationally expensive, even with opencl view it can take a couple of seconds to make a prediction.
I agree, the seampling model in a point cloud is a simple but effective method of "baking" a classification\regression model, by the way there are nearest neighbor models that look fast, n*log(n) instead of n^2, like trees, somewhere saw on the pluses on the net
Perhaps this is one of the best options - to represent the model in the form of an executable application, you can console, taking data on stdin or csv files from the command line because this option can be easily, without hassle to use and protected by obfuscation in time, etc.
Probably yes, console one is unlikely, more likely a dll, so you can load it into your application and run it in the tester, in general first check how the forecast correlates with the market, sometimes the model is weak, the spread does not win, but nevertheless it can be used in ensemble, if it does not roll over very often but positively correlates with the market (price increments) by 1-2% and not strongly with other ensemble models, but there you should look, many ways to cheat.
Another issue is that no one can control that the black box, which was for the "test" will turn out to be the same for dough, but this is in the interests of both sides in the idea...
In general, we can try to publish models like eurobucks for m1 from Dukas, but that would not HFT, 10-30 transactions a day or so, "seal" these models, such as crypto signature, and after a month or so we'll see who has them really wins the market and who has a fantasy.
The signal should start ... and then bother with sealing, printing and integration in MT (and others). Otherwise, you'll spend a lot of time on service functions instead of making a good model. But it may turn out to be in vain. If the signal will be worthy of attention - you can bother.
How are you doing with the P models, any progress?
I wanted to invest somewhere, but no one has signals :)
How are you doing with the P models, any progress?
I wanted to invest somewhere, but no one has signals :)
I have not found anything interesting yet. Then maybe I'll come back.
Probably yes, console one is unlikely, more likely a dll, so you can load it into your application and run it in the tester, in general first check how the forecast correlates with the market, sometimes the model is weak, the spread does not win, but nevertheless it can be used in ensemble, if it does not roll over very often but positively correlates with the market (price increments) by 1-2% and not strongly with other ensemble models, but there you should look, many ways to cheat.
Another issue is that no one can control that the black box, which was for the "test" will turn out to be the same for dough, but this is in the interests of both sides in the idea...
In general, you can try to publish models such as eurobucks for m1 with Dukas, but that would not HFT, 10-30 transactions a day or so, "seal" these models, such as crypto signature, and after a month or so we'll see who has them really wins the market and who has fancy.
Well, now we have a clearer picture.
In order that it would be convenient to load, control and show, so to speak. In order to make loading, controlling, and showing the "product face" more convenient, I don't think there is much wisdom involved, let us display the models in the form of ready-made Expert Advisors - MT4, MT5, Dukas CTrader, etc...
For the test, you can display them in compiled form, and if the "agreement of the parties" happens, then transfer the source code.
Only this should not turn into abstract piskomeration, we should check that the Expert Advisor is based on the MO model.
To do this, I think we will have to tighten requirements and set only those EAs, which have been trained on a specific task.
The very task may contain only objectives in the form of trading signals for the coincidence of which, by the way, we will be able to check the model.
The task can be set, at least for MT, in the form of a marked and saved to a file template from the chart where the signals are shown in the form of graphical objects or indicators.
Well, now we have a clearer picture.
In order to make it convenient to load, control and show, so to speak. "I think there's nothing to wiz, let's expose models in the form of ready-made Expert Advisors - MT4, MT5, Dukas CTrader, etc...
MT-tester cannot be customized, it's a black box too, you can't even test whether it works properly, and the speed is still orders of magnitude lower than what you can set up on your own, especially if simple options for optimization. Well, how should I say it ... in short, you will not be taken seriously with "advisor" for any public platform, and C++ dll is used equally in Rentech and Goldmansax.
For the test, you can expose them in compiled form, and if the "agreement of the parties" happens, then already hand over the sources.
Only this should not turn into abstract piskomeration, we should check that the EA is based on the MO model.
To do this, I think we will have to tighten requirements and set only those EAs, which are trained on a specific task.
The task itself, in order not to limit the choice of predictors, can only contain targets in the form of trading signals the coincidence with which, by the way, will allow us to check the model.
The task can be set, at least for MT, in the form of a marked and saved to a file template from the chart in which the signals are marked in the form of graphical objects or indicators.
The main thing is to enable automatic activation, connection, testing in the tester in the future when it comes and getting the result in the form of equity.
You're not part-owner of some bank together with Alyosha, are you?
Even Gref tries to keep quiet about it.