Machine learning in trading: theory, models, practice and algo-trading - page 3548

New comment
 
Maxim Dmitrievsky #:
The particles would, li, zh, b, l, l, g are always written separately.

It's time to write only with interjections and unprintable words :-)

 

How can the trading of the majority of TCs in the market be expressed through the MO ?

What should be the target?

What areyour thoughts?

 
Maxim Kuznetsov #:

it's time to write only in interjections and unprintable words :-)

That's when you have weak nerves. Sometimes such people come here, but they get even more upset when you start to communicate with them in their language. And in general, it is inherent in women.
 

When you get good TCs on average, you start to think about which one to choose, by what criteria?

Train/val/test starts to seem like an inferior selection criterion, you want something more. Plus you want to train on all the data now.

That's where CV comes in handy.

I looked at the trained model first:

>>> models[-1][1].get_best_score()
{'learn': {'Accuracy': 0.8869165023011177, 'Logloss': 0.3232375323689657}, 'validation': {'Accuracy': 0.8467432950191571, 'Logloss': 0.3847962120305579}}

What is good here? Relatively close values of accuracy and logloss per train/test.

Then I did CV on 5 folds:

>>> pd.DataFrame(models[-1][-1]).drop(['iterations', 'test-Logloss-std'], axis=1)
     test-Accuracy-mean  test-Accuracy-std  train-Accuracy-mean  train-Accuracy-std  test-Logloss-mean  train-Logloss-mean  train-Logloss-std

0            0.820566           0.008358             0.863584            0.003167           0.425334            0.368001           0.003848

What's good here? Average accuracy per test is not much lower, logloss is slightly lower.

Is it now possible to get rid of the test subsample when selecting a TS, and train on all data and rely on CV results? I think so.

 
I think simulations and confidence intervals work better than CVs
 
Maxim Kuznetsov #:

The "schedule" (the attraction of turning moments to a certain time) is there.

That is, the reversals are not so punctual anymore) Then, of course, a simple calculation of correlation between neighbouring bars will not show such a "schedule".

If it is simply a question of increasing the number of corrections when volatility increases, it is banal and non-triggable.

For tradability, there should be something like an increase in the probability for small corrections to become big ones, if they started at certain time intervals. Here there is a problem of dependence of the "schedule" on the time interval of the history taken for its calculation. And, as usual in such cases, it becomes unclear whether the market really changes with time, or whether it is another "fooling by chance".

Maxim Kuznetsov #:
And by the way, it was YOU who tried to calculate the schedule in artificial retracements on EURUSD and did not see it - don't blame your ogrichi on me.
At first glance - some nonsense. At second glance - also.
 
mytarmailS confidence intervals work better than CV.

I don't know, I hadn't thought about it.

 
mytarmailS #:

Do you do a mash and some kind of indium in the chips? Have you tried just increments?

I'll make a sampler using this method today, I'll see.

So far, the best one I've had so far is using splines. It'll be cool if this one's better.

 

The principle itself is simple: build a channel around the spline for a certain value and forbid to open trades inside it, when marking 1 and 0. You can use quantiles.

Everything below - to buy, everything above - to sell. For the strategy of returning to the average. For a momentum strategy, it's the other way round. The deals that do not meet the condition, you throw out of the dataset. Look at the improvement of metrics.

You should not make the channel very wide, because then there will be few examples left. It is enough to spread the deals a little in the space of attributes.
 
Maxim Dmitrievsky #:

1 In fiches do you do mashka and some turkey?

2 Have you tried just increments?

3 I'll make a sampler using this method today, I'll see.

So far the best one I've had so far is using splines. It will be cool if this one is better.

1 yes

2 no

3 you are doing something ABSOLUTELY different from what I was talking about.

1...354135423543354435453546354735483549355035513552355335543555...3625
New comment