adviseExitLong adviseExitShort

The thresholds for exiting an already open position may be different for those for reversing the position. and rules for such should be established separately. Essentially you may not always want to be Long or Short but sometimes may want to be Neutral.
Is there a way to train for this behavior in the current version?

The result of exiting a position obviously depends on when you've entered it. What exactly do you want to be trained here?

Deciding whether to enter a trade is only part of the strategy. The other part is deciding whether to stay in a trade. One can use the Trade params, Stop, Trail, etc. but those are set at the beginning, when the trade is ordered. They do not take any signals into account.

As it stands now, if you don't set some kind of exit conditions for the "next trade" after adviseLong/Short, your training targets are based on the performance of never-ending positions. However, some strategies don't use exit conditions and instead they use signals that come in the future. i.e. Workshop4_2.

This is what I had in mind but I'm sure you could think of a better solution that is simpler and fits with Zorro's structure.

- adviseLong is a boolean function over the input signals which tries to be true when it would be wise to enter a long position.
- adviseExitLong is also a boolean function over the input signals. It tries to be true when it would be wise to exit any existing long positions.
- The two functions are not independent. adviseExitLong affects the trade performance, in other words the training targets, of adviseLong.

That is the problem that I meant: When shall adviseExitLong be true and when false? I'm at a loss here what this function should return.

How do you define the goodness of an exit without having an open trade?

I don't have a clear proposal. I'm just stating a problem.

That is the problem of learning algorithms: You need some classification criteria. In the examples, timed exits are normally used. So the criteria is just the profit at a certain time after entry. But you can also do it the other way around. If you have a fixed entry condition, you can use the same advise functions for generating exit rules. I only see no practical method for generating entry and exit rules separately from each other.

First, how to use adviseLong for generating exit rules?

You have a clear point here. Still, I would like to play with the idea a bit. You provide exactly 2 advise methods, for Long and Short. Can I have more?

You have two per algo, so by adding more algos you can have an unlimited number of advise methods for the same asset. "Long" and "Short" are just names, both functions behave exactly identical as to my knowledge. So you could use adviseLong for entry and adviseShort for exit, for example.

A way for generating an exit rule would be to write some code for classifying the exit by peeking into the future. In the training phase you can then pass the current classification value to the Prediction parameter at any bar.

adviseLong, adviseShort - no shit man I bet that even jcl is not so crazy to use the current implementation of those

Any profitable systems out there using machine learning and not getting creamed when trading in real? Would appreciate to see a proof that finally this stuff bears some fruits. But so far I have seen only terrible results from it - adviseLong, adviseShort, account gone

Let's finally bring into Zorro some neural network capabilities or even better SVM engine (or not because finding the right SVM kernel is a bit too tough problem to solve). Then we could feed them with all the other stuff which we already got and finally Zorro will take from the rich and give to the poor

Don't forget to check my results here: http://www.myfxbook.com/members/4Xconsult/show-case-1/1164183

If I get creamed some day you can laugh about me but at least I am standing here and sharing what I've got.

Jcl, why don't you share some adiseLong, adviseShort kinda strategy results?

adviseLong's PATTERN mode appears in the workshops. Except for that, in the rest of the manual it's clearly presented as a toy. Probably, I'm the only one who cares about PERCEPTRON and DTREE. That's more of a measure of my level desperation than it is anything else.

And, I already placed a feature request for better model types. http://www.opserver.de/ubb7/ubbthreads.php?ubb=showflat&Number=442414#Post442414