Machine learning in trading: theory, models, practice and algo-trading - page 1471

[Грааль]

elibrarius:

That's about what I'm trying to do. I mark each bar 1 if the TP and 0 if the SL is triggered. I do not mark the right boundary, I just look 10,000 bars ahead - for sure a TP or SL will trigger.

And what (what TS) is initially traded? What is the TP/SL with respect to?

I mean, we tried to predict not the characteristics of TAC, but various derivatives of strategies based on the hypothesis of their higher persistence than the basic characteristics of TAC, but no meaningful result has been obtained. The first thing that comes to mind is forecasting of equity dynamics by TS but equity is not as good as forecasting of price direction or trend/flit change.

[Forester]

Grail:

And initially what (which TS) is being traded? Regarding what the TP/SL.

Regarding each bar. If we open a deal on it.
I think the forest/NS will cover all possible strategies. In idea, each leaf in the tree is a separate strategy with its own parameters.

Grail:
The first thing that comes to mind with this topic is predicting equity dynamics by TS, but equity also sucks as well as price direction or trend/flat change.

Equity is already a secondary thing, without a good prediction of successful trades/price movement, it will be bad too. I have about 50% too.

[Грааль]

elibrarius:
Regarding each bar. If you open a trade on it.

Where to open, how to make a decision?

So TPLs are also secondary, they make sense only within a particular strategy, for example, you can take a machine with certain parameters and teach the forest to exhibit TPLs, depending on, for example, the last 50 values of the selected machine and ATP, etc. change strategy TPLs "fly off"

PS In general, there is an opinion among algotraders that TP/SL is only for manual improvisation, especially for TP, the MO says that the bull market is in longs, bear market - in shorts, change - flip, TP/SL is when there is no possibility to follow the market and the position is opened when we do not know what the market situation, and the robot is constantly aware of it, so there is no need for TP/SL.

[Maxim Dmitrievsky]

Another example of meta markup. Well, they do everything by the book, the copyists in short

There's one funny guy who decided to create an open sorts hedge fund by the book.

https://www.quantopian.com/posts/meta-labeling-advances-in-financial-machine-learning-ch-3-pg-50

[Forester]

The Grail:

Where to open, how to make a decision?

So TPLs are also secondary, they make sense only within a particular strategy, for example, you can take the machines with certain parameters and teach the forest to set TPLs, depending on, for example, the last 50 values of the selected machine and ATP, etc. change strategy TPLs "fly off".

PS In general, there is an opinion among algotraders that TP / SSL is only for manual improvisation, especially for TP, the MO says that the bull market is in longs, bear market - in shorts, change - flip, TP / SSL is when there is no possibility to follow the market and open positions when you do not know what the market situation, and the robot always knows, so TP / SSL is of no use to him.

I'm training two separate models.

1 to buy with parameters TP/SL
2 to sell with parameters TP/SL

Models with 3 or more classes (where there will be buy/wait/sell) I think will be less effective.

I am still trying to solve a simple problem - to calculate the percentage of winning trades with fixed TP/SL at trade opening. Tracking the market and open trades is much more difficult and again secondary, initially we need to find the right entries. And the tracking and normal trailing can be added later.

[Forester]

And in general, models like simple solutions. I coach on 2017, and it was on the rise. There are models who say that the whole year on every bar should have been bought)) With profits in the tens of thousands of pips for the year. But there are the same huge drawdowns up to 2-3 months.

But the problem is that we don't know when the global trend will end.
In general, I believe that we should look at the global movement in large timeframes. I tried to teach scalping on M1, but it was 50% +-5%.

[Maxim Dmitrievsky]

They build the first model, which global market regimes clusters/projects

then the second model is adjusted to the specific modes. It's a lot of fuss, but it's the only normal option

[Forester]

Maxim Dmitrievsky:

I'm interested in ARIMA chips and stuff like that. I transform the quotire in a clever way so it would be more stationary but would not lose levels, over autoregression, in blue - the predicted values over the actual ones, the right one - the net forecast. So far this last rise was predicted by the pussy, now it says sell. I may use it to look on top of MO and see what happens.

Is it like a deletion of a trend? What if we subtract the MA with a larger period from the quote? And get a scalping strategy?

Well, not MA, but maybe something more sophisticated...

[Forester]

Maxim Dmitrievsky:

Such a purely scalper trade turned out. Although on H4 it will still fall. Well, small TF will redraw a little on impulses always

I do not know how to algorithmize it properly. I don't know if I can algorithmize it properly.

I'm confused by the fact that your forecast is a long time ahead. It would be good to guess 5 bars ahead or 50 pts. And the farther the forecast, the less likely its success. The Hydrometeorological Center still screws up, even though it has accumulated a lot of statistics and new tools in the form of MO.

[Maxim Dmitrievsky]

I spent half a fucking day rewriting my overgrown rl lib for oversampling, but finally got it done... the first time (oversampling is automatic), while before I had to choose from a list of models

I got nothing with the meta markup, it's probably just another book nonsense, or my bad hands/thoughts

But, of course, such beautiful curves type of training in a month, and then a few years to get unrestrained profits, as here recently demonstrated, while not succeed