Machine learning in trading: theory, models, practice and algo-trading - page 237
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
so what? all bad, apparently? defeating the market in the form of buy and hold has failed.
the question is not...
The question is why a model that is trained on randomness performs better on new market data (OOS) than a model that was originally trained on market data
p.s. no one tried to create a super system
the question is not...
The question is why a model that is trained on randomness performs better on new market data (OOS) than a model that was originally trained on market data
p.s. nobody tried to create a super system
It is retrained. The retrained model has no characteristics at all.
questions...
1) why the one that is trained on random is not retrained?
2) why the one that is trained on the randome doesn't lose directed deposit?
3) why the one that is trained on the real data is losing directed those with a trend?
questions...
1) why the one that is trained on random is not retrained?
2) why the one that is trained on the randome does not lose directed deposit?
3) why the one that is trained on the real data is losing directed those trend?
It seems to me (not sure)
questions...
1) why the one that is trained on random is not retrained?
2) why the one that is trained on the randome does not lose directed deposit?
3) why the one which is trained on the real data is losing directed by the trend?
The thing is that when we train on noise we get a neutral system. It also essentially works randomly with market data. And it is more profitable to act randomly in the market than to think that it knows where the market will go(a trained grid thinks it does).
Nothing surprising therefore, the market always tries to move against its statistics (against the masses), and learning is rote statistics at worst and capturing patterns at best. But neither statistics nor patterns work in the future, because the market always tries to move against its statistics, against the masses. The circle is closed. Learning makes no sense, anything that will be learned (in a good way, without overtraining) will be useless on the OOS.
Hence the average (not good and not bad) results on market data system trained on random. These thoughts were voiced by me somewhere in 2009, when I suggested to generate a synthetic series essentially from random data but with parametric characteristics and study how the TS behaves on such data, to apply then to the real market data. This is a "pessimistic approach" to the market.
The "optimistic" approach consists in the "flowing patterns" I mentioned the same year. The meaning is the same - the market is constantly changing, but the difference is in tracking these changes, tracking the market derivative and trading against the changes (or according to the changes).
The two approaches "pessimistic" and "optimistic" do not contradict each other, just looking at the market from different angles (facet/profile).
And note, I didn't say a word about the market being random. If the market were random, we would not have seen similar effects with models trained on random data. And they won't let Uncles be a random market (economics by her leg).
I tried to select working patterns "my way" from the random, the model was trained very long, I selected the first best pattern for buy, and then the studio crashed and had to cover it through the completion of tasks, I managed to save the model, but the target was not saved and essentially the search for patterns became impossible, I need to re-train the whole model, such a nuisance...
The only pattern I managed to save was pretty good(or so I'd like to think :)) )
It looks very nice, don't you agree!!!
kol. profitable 77% is good
Take/stop ratio more than 1k2 is also +-not bad
i got a lot of feedbacks from the oos and i dont know how to use them, they are coming out of the market.
So, I have to teach the model again and test, test, think and test... In the meantime, good luck to all)
I tried to select working patterns "my way" from the random, the model was trained very long, I selected the first best buy pattern and then the studio got stuck and had to close it through the completion of tasks, I managed to save the model, but the target was not saved and essentially the search for patterns became impossible, I need to re-train the whole model, such a nuisance...
The only pattern I managed to save was pretty good(or so I'd like to think :)) )
It looks very nice, don't you agree!!!
kol. profitable 77% is good
Take/stop ratio more than 1k2 is also +-not bad
i have a feeling that the market is flat and i dont know what to do with it.
So, I have to teach the model again and test, test, think and test... In the meantime, good luck to all)
I looked at the new dataset for interest.
You're a little ahead of the curve. I could hardly find you))) catch up...
https://numer.ai/
Something went wrong.
I took the same model with the same parameters as last time, the result after evaluation was worse. I had to pick up the parameters of the model again and crossvalidate. I decided to bypass this procedure and train another model that showed much better results on training data. Now I filled the forecast of this second model and the logloss on numerai was worse. It is not good. I'll go back to the first model and crossvalidate.
I checked the same pattern, but not for half a year (oos)(like the first time), but for 5 years (oos)
The indicators have fallen dramatically, but I can't say that the pattern does not work, stops\takes are the same, nothing has changed or changed and moreover notice a steady downward trend in the market, the pattern is long, so we can only trade long
If we optimize the stops and takeoffs we may get a better picture, but it's a fine adjustment.
I'm a bit confused by all this, why does it work at all!?
=================================================================
And another question for the room: Is it interested in anyone what I put up, and as it were not observed any interest ...
Maybe no one needs it and I'm just littering the thread?
=================================================================
And another question for the room: Is it interesting to anyone, what I put out, and as it is not observed any interest ....
Maybe no one needs it, and I'm just littering the branch?
Of course it's interesting.
Now realtime. And don't forget to monitor here.