neural network and inputs - page 14
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
How do you choose and serve what?
choose a bigger one and serve the key to the flat... on a plate... on a plate...
;)
to choose a bigger one and serve the key to the flat... on a plate... on a plate...
;)
Thanks... I knew it.... :-)
Respect!!!
No kidding -- from the heart.
Thanks... I knew it.... :-)
Respect!!!
No kidding - from the heart.
Thank you.
As a matter of fact, the grids are not very choosy about nutrition, the diet should more or less correspond to the task at hand. That is, since we are looking for unknown patterns, we need a series in which there are many, moreover, many correlated with the result (in our case - prediction).
Hence the second moment - what exactly to predict. I.e. what to feed to the output and what to teach the grid. The question is not about that in the sub-genre, but it should be, because otherwise the mesh may successfully learn and set forecasting records on the OutOfSample and, moreover, remain impotent because in your childhood you were subjected to abusive treatment.
It's just something to think about, I won't help you with specifics, because I'm too young myself...
;)
How about this.
We feed the current bar description vector to the input neuron. For example, opening price increments, opening price increments for other symbols, volumes, current time, interest rates for symbols, news calendar, lunar and solar cycles, and so on. It turns out that as many neurons as there are in the input layer, this is the history we should consider.
Then comes the network itself. The grid is a two-dimensional matrix of neurons (not necessarily for the convenience of describing the idea). To reduce number of calculations in neurons we limit the number of synapses, say, for example, 5-10, but every synapse can connect to any axon within its range limited by a certain circle, (you can put a more powerful processor and make it to any axon, any neuron). As a result, we should get a network of who knows what configuration, with a large number of feedbacks, which will consider the bazaar in aggregate. Then again, according to our imagination, we turn on the hangover function and periodically nail a couple of neurons. Birth function, periodically create a couple of three neurons. As a result, we see that the grid adjusts itself, its internal architecture and number of neurons, although we have no idea how to make the minimum number of neurons sufficient.
Output signal. First thing that arrives in ventral space, build a ts with perfect inputs and outputs, e.g. channel on polynomial of n-degree. Bounce on the channel boundaries and train the network by these signals.
Or a target function of perfect equity, a straight line to the top right corner of the monitor, and then let the network look for the correct inputs and outputs on its own.
Any TS, with or without neural networks, uses patterns that are sought on past data. So in essence, there is no guarantee of earning on these found patterns in the future. Or do you have some method of determining that you can make money in the future on patterns found on past data?
I am not the only one. If you write in detail, it would be a lot of text, if you are interested you can read it on Spider, in Neo's thread about patterns, for example, or in Felix White's threads. Avals appears on this forum, read his posts.
There are some universal regularities (or rather laws) that exist not only in the market, but everywhere. For example, inertia is enough not to remain without bread and butter. It worked in the past, it will work in the future. And if will stop working, the last thing we will be interested in is the breakdown of MTS and the loss of the deposit:). (Think about what will happen if the laws of physics change even slightly.
And guessing whether it will work / will not work - not our method, it goes to the gypsies.
I am not the only one. If you write in detail, it would be a lot of text, if you are interested you can read it on Spider, in Neo's thread about patterns, for example, or in Felix White's threads. Avals appears on this forum, read his posts.
There are some universal regularities (or rather laws) that exist not only in the market, but everywhere. For example, inertia is enough not to remain without bread and butter. It worked in the past, it will work in the future. And if will stop working, the last thing we will be interested in is the breakdown of MTS and the loss of the deposit:). (Think about what will happen if the laws of physics change even slightly.
And guessing will work / will not work - not our method, it's the gypsies.
Aha-ha-ha!
Also gravity - the price, like, goes up and then comes down
Yeah, yeah, yeah!
And gravity - the price, like, goes up and then comes down.
It's hard to guess whether a footballer will hit the ball or not, but when the ball is already in the air, there's no need to guess, it's clear that it will fall to the ground. (It may, however, get stuck on a tree - there's a stop-loss on the market for such cases).
The main thing is not to go ahead with predictions-guessing before you hit it, it's a hopeless case. You might take a swing, but then change your mind.
In general, imho, the system should be built in such a way that it trades some of the same patterns.
This fully applies to the TS both with and without the nerves. Such system, in its essence, will trade on OOS with activity decaying with time,
i.e., gradually, the number of transactions coming to the unit time over time will tend to 0 (efficiency remains at a constant level, decreases only the profitability, it is only necessary to retrain over time, to refresh the base of actual patterns), because "learned" patterns occur less and less often....
So there should be no hint of interpolation/approximation of the data handled by TC.
PS MetaDriver witnessed the concept of such my TS about 2 years ago.... but i gradually got bogged down in my endless experiments, that i lost the right way without noticing... of course there is a possibility to dig in archive sources, but it will be so difficult to do, that it will be easier to write all over again...(
Thank you.
As a matter of fact, the grids are not very choosy about nutrition, the diet should more or less correspond to the task at hand. I.e., since we are looking for unknown patterns, we need series in which there are many of them, moreover, many correlated with the result (in our case - prediction).
Hence the second moment - what exactly to predict. I.e. what to feed to the output and what to teach the grid. The question is not about that in the sub-genre, but it should be, because otherwise the mesh may successfully learn and set forecasting records on the OutOfSample and, moreover, remain impotent because in your childhood you were subjected to abusive treatment.
It's just something to think about, I won't help you with specifics, because I'm too young myself...
;)
:-)
Thank you.
I got it. Sir. On it.