Machine learning in trading: theory, models, practice and algo-trading - page 175

 
Mihail Marchukajtes:
Real-time volume as well as delta (the number of buyers of sellers) I take on a subscription with clusterdelta, it costs 300 rubles per month, not that expensive. I also get daily volumes at CME for free. In the daily billet section, for the pound billet number 27 for the euro 39. I write them in a file, the indicator reads the file and displays them on the chart.
Really not expensive.
 
Mihail Marchukajtes:

Here you have praised Reshetov, without understanding his work, but he did MEGA cool thing, Tell me why????

He solved one of the important problems in the construction of NS. It's choosing those predicates that give the most generalization. I have an unloading file has about 40 predicates, and a third of the main, the rest of the lag from these data, but the optimizer builds models using only 4-5 predicates. And they are always different. Models that used more than 5 predicates as practice has shown work not very well, a lot of values was "I don't know", like this model is seldom but accurate, and works longer, and considering that I optimize the model every day, I do not need it, Well, the number of records I do for the last 5 days (in accordance with the volume and OI) as a result, I have a table of 45 columns and 30 rows and it is enough to earn (as shown in the screenshot) and it does not matter how the network divides the good and bad, what is important that she did it steadily. Not infrequently I have to flip the TS, because after training it starts to STABILIZE losing, just flip it and voila, we start to earn consistently, so that's how it is....

I will not say anything. No, I'll tell you: I did not sling mud at Reshetov's work. I just don't use it.

I'm really tired of forum squabbles over nothing.
 
J.B:

And have to wonder:) It worked fine at that time, I don't dare to talk about it now, because I'm forced to, but I "heard from someone who heard" about quants from the fund, where they give 10k vectors to CNN inputs, though I don't know their recent returnees, but in 2011 they had 12% on half a yard, which is cool, although they dropped by 8% in the middle, but still...

No comment about "profitable models" on 10 features)) All who really cuts the market favorable "guru" convincingly argue that the model should be as simple as possible, that not to be overtrained, that on the mashki and BB, you can build a "profitable model" and so on. Many thanks to them))

Bullshit. The innuendo is more than strange. If I didn't build models for myself with more than careful attention I wouldn't be writing here. Everything is based on practice.

I'm just not being very subtle and I've been hinting for a long time that the sort of specialists working in funds is quite diverse. A man who does not understand the difference in loss functions for different tasks, what can he do? Throw beautiful words in the forum. In my opinion, that's all.

Not only are your arguments sometimes wrong, but they are also often unsupported.

Can you at least say here publicly (not prove) what kinds of NS with what properties allow to process such amount of features painlessly? Asking you to provide some reserch is, I understand it useless in general.

Maybe you have the wrong topic?
 
Alexey Burnakov:
so many features

It seems to me that he meant just a series of 10,000 prices, i.e. there is only one feature (price, or its increments). I understand that for convolutional nets, it is normal to feed a long series of prices, and then it will find patterns, indicators, and whatever else it needs during training.

But usually if you train a neural network on 10000 predictors it will lead to nothing, I agree.

 
Alexey Burnakov:
1) Bullshit. Hints are more than strange. I wouldn't have written here if I hadn't been building models for myself with more than careful attention. Everything is based on practice.

I'm just not very subtle and have been hinting for a long time that the sort of specialists working in foundations is quite diverse. A person who does not understand the difference in the loss-functions for different tasks, what can he do? Throw beautiful words in the forum. In my opinion, that's all.

Your reasoning is not only wrong, it is often unsupported.

2) Can you here publicly say (not prove) at least what kinds of NS with what properties allow you to process such a number of features painlessly? Asking you to provide some reserch I understand it useless in general.

3) Maybe you're wrong topic?

1) Sure, sure... You're right, I was just confusing people, well... it benefits me and all the various hedge fund employees, so don't be offended.

Okay, well, since you're so clever In fact all hedge-funds just look for "patterns" on currency pairs one by one and compare models by R^2 and mutual entropy and then "take big losses" as it says in your profile, we all are like that, you're right, we are "different" only because we keep it secret and gossip with each other but we are ashamed of it.

2) CNN? Besides, did I say that "everything" is one neuron of the same type? Aren't there any dimension compression technologies (PCA, autoencoders, etc., etc.), feature selection, etc.? Do you have any idea how many data streams come from nasdaq alone if you subscribe to everything? And if it's not just nasdaq?

3) You're right, the attempt of misinformation failed, I'll go to merge the aggregated depot of clients with big jahs and wonder what's wrong, if by R^2 everything fits)))))

PS: I saw your performance on kaggle, it means you sometimes look in the right places, the winton contest is over, but there is a dataset and you can simulate and subdmit predicates and see if you are at least in the top 10, then we'll talk, but now keep exposing me, our conversation makes my colleagues laugh, and positive emotions received naturally are very important for a trader)))

 
Yeah... the branch is drowning in negativity... sadly...
 
J.B.:
What do you need to have in your luggage to get your team interested? Or a team like yours?
 

Colleagues, let's not quarrel! I'm certainly not a role model, but I only look through this thread from this forum.

Machine learning though the child of statistics, but with exponential speed filled with heuristics without sufficient grounds and the more so a strict theory, that is "who works is right", it is now mostly alchemy and therefore argue in dogmas and nuances taken out of context is not reasonable, much wiser to have the power to delve into other people's models and maybe pull something valuable from them for yourself.

 
J.B:

1) Sure, sure... you're right, I was just confusing people, well... it benefits me and all the various hedge fund employees, so don't be offended.

Okay, well, since you're so clever In fact all hedge-funds just look for "patterns" on currency pairs one by one and compare models by R^2 and mutual entropy and then "take big losses" as it says in your profile, we all are like that, you're right, we are "different" only because we keep it secret and gossip with each other but we are ashamed of it.

2) CNN? Besides, did I say that "everything" is one neuron of the same type? Aren't there any dimension compression technologies (PCA, autoencoders, etc., etc.), feature selection, etc.? Do you have any idea how many data streams come from nasdaq alone if you subscribe to everything? And if it's not just nasdaq?

3) You're right, the attempt of misinformation failed, I'll go to merge the aggregated depot of clients with big jahs and wonder what's wrong, if by R^2 everything fits)))))

PS: I've seen your performance on kaggle, it means you look in the right places sometimes, the winton contest is over, but there is a dataset and you can simulate, subsimmit predicates and see if you are at least in the top 10, then we'll talk, but now keep exposing me, our conversation makes my colleagues laugh, and positive emotions received naturally are especially important for traders)))

1) no comments. Bullshit.

2) Well, that's how we're talking about dimensional reduction in the first place. This is the steps BEFORE training. And about the properties of the network itself I have not heard.

3) What a load of bullshit. Something to say about L1, L2 norms for lossfunction?

On Kaggle, exactly the regression was done with the L1 weighted fitness function. And the top spots were taken by people who were constructing the fits on past price data.

And what can you add about the quality metrics obtained in the top places in that competition? Or is this just more bullshit? The looser is usually the one who roars...
 
Combinator:
What do you need to have in your luggage to get your team interested? Or a team like yours?
Answered in person.