Machine learning in trading: theory, models, practice and algo-trading - page 1691

 
Maxim Dmitrievsky:

As the great and unsurpassed Alexander wrote (and continues to write on Smradlab) we need to work with market time, and we will all be happy.

This is the immutable truth, which must be learned like the "Our Father" and repeated while sitting at the monitor.

Understanding comes later, first - unshakable faith

I remember a movie about Shurik who pierced through time and space).

 
Igor Makanu:

about game theory, but as if on the fingers, as applied to MO

What do we usually do?

Here's a flip card game, here's the NS, and we immediately limit the NS by the rules, but not by the rules of the game, but exactly by our vision - like to go only from the smallest card and toss one card at a time - my father taught me so... so we started training and got some kind of trained card game AI

What's right, in terms of combinatorics: the general rules of the game and the target - the number of wins, and how there NS will go with the smallest card or toss exclusively trumps... why should we interfere? - the goal is the maximum number of wins


that's about the primitive way we were taught ... I forgot the name of the software that has beaten all Atari games - stupid search and even stupid without knowledge of the rules of the game - by analyzing pixels on the screen, it seems - I could be wrong here cursory and have long read

Well, two or three people will sit down against NS... They'll strip him to the bone. NS does not need max wins in as many games as possible. It's a max return on investment. NS has put something on the line, and can go all-in in certain situations. But NS has to know who he's playing against, right down to the last detail. It's more about poker than fooling around. There's equity there, too. The win goes to whoever exploits the opponents. If the level is high.

As they say on the internet: You can play against Phil Ivey and beat him, but you'll lose on the distance. Why? He knows howto play against you, based on what you've already shown in past games. And you don't know how to play against him. He'snot playing cards, he's playing you.

To play and win, you have to know your risk. Where you can risk everything and against whom and when, and where you're better off not sitting at the table at all!

When you sit down at the table, you know when to leave.
 

In Dence there is a parameter called Dropout.
Description:
Dropout consists in randomly setting a fraction rate of input units to 0 at each update during training time, which helps prevent overfitting.

The meaning is not very clear, but it seems to be some kind of fight against overfitting. So, if you increase it, the quality of training increases. But if I increase it more than 0.5, TensorFlow will start swearing:
WARNING:tensorflow:Large dropout rate: Please ensure that this is intended.

And it is after 0.5 that the quality starts.

Does anyone understand why this is so and what it is?

 
Evgeny Dyuka:

In Dence there is such a parameter Dropout.
Description:
Dropout consists in randomly setting a fraction rate of input units to 0 at each update during training time, which helps prevent overfitting.

The meaning is not very clear, but it seems to be some kind of fight against overfitting. So, if you increase it, the quality of training increases. But if I increase it more than 0.5, TensorFlow will start swearing:
WARNING:tensorflow:Large dropout rate: Please ensure that this is intended.

Namely after 0.5 the quality starts.

Anybody understand why and what it is?

I am far from being an expert on neurons, butDropout is a kind of regularization, so, yes, it is to combat overtraining, when training a neuron some of the neurons are zeroed (killed), it is done so that the neuron can better generalize and not concentrate much information in one neuron. Perhaps 0.5 is the maximum threshold for the number of neurons to zero out

 
Igor Makanu:

I doubt you can confirm this "axiom" - I know it's written on every "fence"

I don't read fences for a long time now. MM is a kind of "cosmetics", the real profit/risk ratio it can not bring to the plus from the minus, but it can create such an illusion on the backtest.

Of course to disassemble all the possible ways to cheat a trader is not real, in a number of people who have done it even in the coolest organizations. But on the example of martin it is easy to show, although probably not for everyone.

Let's take for example 2TS with the same series of random entries, the same TP/CL, but the first one has a constant lot, and the second, if the previous trade is lossy - double.

Here is one of the random examples.

As can be seen, the illusion of profitability has been created from the first ASR < 0, while the second has ASR > 3 - a miracle!

 
Kesha Rutov:

I don't read fences for a long time. The MM is a kind of "cosmetics", the real profit/risk ratio it cannot bring to the plus from the minus, but it can create such an illusion on the backtest.

Of course to disassemble all the possible ways to cheat a trader is not real, in a number of people who have done it even in the coolest organizations. But on the example of martin it is easy to show, although probably not for everyone.

Let's take for example 2TS with the same series of random entries, the same TP/CL, but the first one has a constant lot, and the second, if the previous trade is losing - double.

Here is one of the random examples.

As can be seen, the first ASR < 0 and the second ASR > 3 miracles!

everything is clear, I've been there, and more than once

 
Kesha Rutov:

I don't read fences for a long time. The MM is a kind of "cosmetics", the real profit/risk ratio it cannot bring to the plus from the minus, but it can create such an illusion on the backtest.

Of course to disassemble all the possible ways to cheat a trader is not real, in a number of people who have done it even in the coolest organizations. But on the example of martin it is easy to show, although probably not for everyone.

Let's take for example 2TS with the same series of random entries, the same TP/CL, but the first one has a constant lot, and the second, if the previous trade is losing - double.

Here is one of the random examples.

As can be seen, the illusion of profitability has been created from the first ASR <0 and the second ASR >3 Miracles!

miracles at the very end

hence the avalanche.

the avalanche is not made of poop, but of the focus on the inability to perform tehanalysis, i.e. the emphasis on the randomness of price movements
 
mytarmailS:

I am far from an expert on neurons, butDropout is a type of regularization, so yes, it is to combat overtraining, when you train a neuron, some neurons are zeroed (killed) , this is done so that the neuron is better to generalize and not concentrate much information in one neuron. probably 0.5 is the maximum threshold for the number of neurons zeroed

Ok, got it. It turns out that if it works well with dropout greater than 0.5, then there is a lot of unnecessary information in the vector.
 
Igor Makanu:

Yes, it's all clear, so to say passed and more than once

So what are you arguing about?

If you try to get a job in a bank or hedge fund and say that you optimize strategies through the backtester, and even with all the guts (mm, execution, etc.) it will be a black mark, they will not take you even to the provincial brokerage company.

 
Kesha Rutov:

So what are you arguing about?

This is the alphabet, if you try to get a job at a bank or hedge fund and say that you optimize strategies through the backtester, and even with all the guts (mm, execution, etc.) it will be a black mark, not even in the provincial DC.

I'm not arguing, I'm the instigator

Well, what's the point of arguing? You are not going to share your honestly earned profit with me, am I not promising to compensate for your losses?

))))


As for the banks, they have different objectives, but I can tell you that banks lose money too, and regularly ;)

As for the brokers, the real dealers are engaged in training, it really works. - i think that the objectives are different too.

SZS: i remember the fight with heretics, now it's time to find out who is right and who is to be burned at the stake ))))


UPD: it's time for great stories....


Here is a graph, EA itself generates TS, TS is not ideal, but imho, this graph can be worked with in the future