Machine learning for robots - page 4
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
google code is still available, don't know if it will work or not
hit me up
Ivan, what you have come up with is amazing, I myself started to think about it a few months ago, but only in my wildest dreams, as I'm a programmer and not a beginner.
There are three things in a trading system
1) indicator
2) all the signals
3) right signals
What do you need to filter? all three points or 2 and 3 or just 3?
And what if my system works on non-standard indicators or levels? will your algorithm be able to simulate my trading if i trade on levels?
Ivan, what you have come up with is amazing, I myself started to think about it a few months ago, but only in my wildest dreams, as I'm a programmer and not a beginner.
There are three things in a trading system
1) indicator
2) all the signals
3) right signals
What do you need to filter? all three points or 2 and 3 or just 3?
And what if my system works on non-standard indicators or levels? will your algorithm be able to simulate my trading if i trade on levels?
The idea here is to try the MO according to point 3, i.e. by correct signals, which in turn can be filtered out of all signals obtained by any method, from any sources.
Filtration is done primitively, by the difference (potential profit in pips) on a certain number of bars, the inputs may be OHLC itself or values of any indicators calculated using a specified formula.
You can generate several different variants of models, but whether they will be able to simulate trade, that's what the experiment is for...
but whether they can simulate the trade, that's what the experiment is for...
They can't. At most, it will be the same story-fitting.
They won't. At most, it will still be the same kind of fitting to history.
And optimization, which everybody is always doing, isn't fitting to history, and in your analysis are you using data from the future or a special approximation?)
Here's an idea, try MO by point 3, i.e. by correct signals, which in turn can be filtered out of all signals obtained by any method, from any sources.
Filtration is done primitively, by the difference (potential profit in pips) on a certain number of bars, the inputs may be OHLC itself or values of any indicators calculated using a specified formula.
You can generate several different variants of models, but whether they will be able to simulate trade, that's what the experiment is for...
OK, I see what you mean, but I have some questions.)
1) Let me see, if I take your MACD as an example, it means that you generate inputs by MACD and then transmit them to the net and the net generates a MACD phantom in its head and trade by it.
2) If I understand correctly, why should you train the net on "alien" systems with disadvantages if you can simply take the inputs of the zigzag and it will be an ideal trade, the net output will be an ideal system
3) What about levels? I have an indicator that builds levels, but levels are not a function, while your network works with functional data - just indicators, am I right?
3) And what about levels, I have an indicator that builds levels, but levels are not a function, and your network works with functional data those are purely indicators, am I right? Again, if correct, can I get around this?
The level is a function of an array of price data.
OK, I see what you mean, but I have some questions.)
1) Let me see, if I take your MACD as an example, it means that you generate inputs by MACD and then transmit them to the net and the net generates a MACD phantom in its head and trade by it.
2) If I understand correctly, why should you train the net on "alien" systems with disadvantages if you can simply take the inputs of the zigzag and it will be an ideal trade, the net output will be an ideal system
3) What about levels? I have an indicator that builds levels, but levels are not a function, and your network works with functional data - just indicators, am I right?
1) Exactly, the network learns to trade as a filtered indicator from wrong inputs, in the example of MACD.
2. The reason is that not all profitable entries are technically justified, for example if the movement is based on news, then what is the point of teaching the net to predict it based on price patterns.
3. if your indicator builds levels based only on price data, then the network in "its head" probably can simulate them, so the methodology may be the same as for other indicators - form signals by levels, filter the wrong ones and train, of course, we must first decide on the input training sequence - size, offset, recalculation formula.
1. that's right, the network learns to trade as a filtered indicator from erroneous inputs, in the MACD example.
2. Just not all profitable entries are technically justified, for example if the movement is based on news, what is the point of teaching the network to predict it based on price patterns.
3) If your indicator makes levels based only on price data, the network in "its head" probably can simulate them, so the methodology may be the same as for other indicators - form signals by levels, filter out false signals and teach, of course, we must first decide on the input training sequence - size, offset, re-calculation formula.
Ok, I will write a code, try to generate signals and send it to you. If you have tried to teach the network by zigzag, tell me the results.
... But have you tried zig-zagging the network?
Of course I tried, and not only me, for example in the thread on MO, there are those who did it, repeating the mantra of rubbish in the input and apparently forgetting that the rubbish in the formal output when trained with a teacher is not much better, with the selection and shuffling vector of features does not save from overfitting.