Machine learning in trading: theory, models, practice and algo-trading - page 66

 
Alexey Burnakov:
Sort of. But it is not necessary to send all bars to the machine. It is possible to take all strong signals (they will be few!) and the same number of weak ones. To balance. And reduce the volume of training.
Pilya so how do you know that this is a strong signal and it should be submitted to the predictor for discussion???? By what criteria? A ball with a width of N points or a volume greater than that. Based on what criteria? That's exactly what I'm asking.....
 
Mihail Marchukajtes:
If you have analyzed the chart, you have got the points after which the rate went up, built a model and now the rate is ticking, at what point do you apply the system? At each bar you will ask the predictor, "Tell me, friend, is this the bar that will cause an upward movement of 100 points? Predictor: "No, my lord." Okay, we wait for the next bar. "Tell me, Predictor, is this the bar?" Predictor: "Yes my lord." You "Woah, woah, woah" and in the end it turns out that we have to analyze each bar, with 10 entries the optimal number of entries is 100, when the predictor can saw it to zero. It turns out that only 4 days you can cram into it so that the generalization ability was at an appropriate level. I do 100 records at 5 minutes over 3 weeks, analyzing the market over the oodles of bars at the same level of generalization. That's the difference....

Pfft. I don't understand your bravado.

With 10 inputs, the optimal number of entries is 100

Since when...?

I have 10 years of history in my teaching. Sliced about 25,000 examples. The model is built on a maximum of 10 predictors. Lately I've been making do with 5. Why do I even need a model where the number of predictors to the number of observations is such that it describes everything perfectly. I do not understand.

 
Yury Reshetov:

If the classifier, which we use as a "lie detector" of the wands, reported that the wands "do not lie", then we open a deal according to the wands' readings. If the classifier reported that the wands are "lying", then we can open a trade in the opposite direction to the wands' readings.

This was for the binary classifiers.

The ternary classifier says with "-" that it cannot say with adequate probability if the testers are lying or not, thus urging to sit on the fence and smoke bamboo until the next signal - testers crossing.

By the way, about the Ternary classifier it is quite appropriate. I look forward to when it can be saved to a file, then I will uncover the secret of data preparation. How about this compromise????
 
Mihail Marchukajtes:
Fuck so how do you know it's a strong signal and it needs to be submitted to the predictor for discussion???? According to what criteria? A ball with a width of N points, or with a volume greater than that. Based on what criteria? That's exactly what I'm asking.....
In the course of trading, a trained machine will process every bar when there is no trade in the market. If that's what you mean.
 
Alexey Burnakov:

Pfft. I don't understand your bravado.

What the...?

I have 10 years of history in my teaching. About 25,000 examples have been sliced. The model is built on a maximum of 10 predictors. Lately I've been making do with 5. Why do I even need a model where the number of predictors to the number of observations is such that it describes everything perfectly. I do not understand.

Well, here is purely my observation, it's in case if inputs are either 0 or 1, then 100 records will be unique. Anyway, you can send me a file, I'll train it by my method and send you model, and you can see how it will work in future... So... just curious!!!!
 
Alexey Burnakov:
In the course of trading, a trained machine will process every bar when there is no trade in the market. If that's what you mean.
Well yes I am saying that it will process every bar.... that's right..... now we got it.... :-)
 
Mihail Marchukajtes:
Well, this is purely my observation, if the inputs are either 0 or 1, then 100 records will be unique. Anyway, you can send me the file, I'll train it by my method and send you the model, and you can see how it will work in the future... So... just curious!!!!
Go ahead. I'll get it ready.
 
Mihail Marchukajtes:
Well, yes, that's what I'm saying, every bar will be processed.... that's right..... now we got it.... :-)
and no need to get nervous ))))
 
Alexey Burnakov:

Pfft. I don't understand your bravado.

What the...?

I have 10 years of history in my teaching. About 25,000 examples have been sliced. The model is built on a maximum of 10 predictors. Lately I've been making do with 5. Why do I need a model where the number of predictors to the number of observations is such that it describes everything perfectly. I do not understand.

The very idea of it is interesting.

Got hooked on the swabs. It really doesn't matter. The point is that his model predicts a change of 100 pips (in his example) in an undefined number of bars! At any rate, until the next crossing of the bars. Got it - we need to correct the position. I.e. some indicator that marks the kotir by time, and this is the most difficult.

For sure it cannot check the model but I like the idea and the rest is generally a matter of technique.

 
SanSanych Fomenko:

The very idea of it is interesting.

He's attached to the mashups. It doesn't really matter. The point is that his model predicts a change of 100 pips (in his example) through an undefined number of bars! At any rate, until the next crossing of the bars. Got it - we need to correct the position. I.e. some indicator that marks the kotier by time, and this is the most difficult.

It certainly cannot check the model, but I like the idea.

It is possible to use any TS in classification which leads to buy or sell signals. Any, as long as there is an event that happened..... For example Sequenta is intended for oversold and overbought zones, it is a counter-trend and it often turns counter-trend. Mashki is a trending strategy, by the way, I wanted to try a trending strategy for the classifier....