How do you work with neural networks? - page 6

New comment
 
LeoV:


To be honest, I don't really understand your point.

The thought is simple:

Suppose we have two TCs:

1. On the rebound from the channel, i.e. contrend. It makes profit within the channel, and loses on a breakdown.

2. Breaking through the channel - trend. It incomes on the long trends, loses on the sideways.

If the chart is inverted around the horizontal axis, then both TCs will behave the same way, i.e. the trend one will make profit under a down trend even if the normal chart it was adjusted for was dominated by an up trend.

Actually, the grid should behave symmetrically, i.e. if it has been trained for an up trend, it should also "earn" in a down trend. If it cannot do so, and, for example, "earn" only in the upward trend, then it is worse than the above-mentioned primitive TS, because it will start losing both in the down trend and sideways.

I.e. when training the network it needs to download not only a sample of examples from the analyzed BP, but all examples in the inverted form.

For example:

Suppose there are some two training examples in the training sample. (inputs and outputs from -1 to +1):

0,35 -0,21 0,8 -0,51 -0,71

0,71 0,1 -0,21 -0,96 0,12


Where: From the first to the penultimate value - inputs, e.g. normalized values of (RSI - 50.0) / 50.0 or another oscillator. The last value is what we want the output to be. Accordingly, in order to obtain symmetrical training, each such example should be inverted, i.e. the training sample should be doubled:


0,71 0,1 -0,21 -0,96 0,12

-0,35 0,21 -0,8 0,51 0,71

-0,71 -0,1 0,21 0,96 -0,12

0,35 -0,21 0,8 -0,51 -0,71

 
Reshetov:
Not sorry. See GRNN-GA.

So? Has the neural network been replaced by a genetic algorithm?

What exactly does this link say? What is it about?

Can I get a quote -- for those who don't understand?

 
http://forecast-man.com/ here's another NS library on fuzzy logic, would love to hear comments but probably not many people have heard of it...
 
LeoV:

The solution of this issue is simple - you need to train the net on such a time interval, in which all types of movement are present. Both sideways, upside and downside. Of course you must realize that if the net is only trained on the upside, it will fail on the downside ))))


All correctly said... But what about the trend change? I didn't have much success... one good, then not so good...

ps.but if there are normal inputs for the teacher (i.e. - there is something to find in the inputs) - then it does not matter - a grid (any structure), maps koh, ha, k-method, etc..... everything will work...

 
Vizard:



ps.but if there are normal inputs for the teacher (i.e. - there is something to find in the inputs) - then it does not matter - a grid (any structure), maps koh, ha, k-method, etc..... everything will work...

It should be taken as an axiom that there are no normal inputs in BP. The cat has to know how to cook. See Predicting financial time series
 
Swetten:

So? Has the neural network been replaced by a genetic algorithm?

What exactly does this link say? What is it about?

Can I get a quote -- for those who don't understand?

The point is that by neural networks we mean various algorithms, including regression, genetic, etc. Reference to GRNN-GA is given as an example, because in fact it can be called a neural network only because it uses weighting coefficients - a search engine of relevant examples by query in the database.
 
Reshetov:
The point is that by neural networks we mean various algorithms, including regression, genetic, etc. Reference to GRNN-GA is given as an example, because in fact it can even be called a neural network only because weighting coefficients are used - a search engine of relevant examples by query in the database.

Let me remind you of the original statement: NS and GA are completely different things, unrelated to each other in any way.

One does not change into the other in any way.

Are you going to argue?

 

arguing about methods, algorithms (grids, ga, etc...) to search for signals by condition (teacher) is endless... as well as about the preparation of the vr... the point is different... as long as the general dynamics is maintained, all is well... if the dynamics changes, it's a drain....

 
Vizard: You said it right... just what about changing the dynamics?

It's not about dynamics. The dynamics can be anything - as long as the patterns the network found during the training period work in the future. That's the rub -.....)))
 
Reshetov:

The thought is simple:

Suppose we have two TCs:

1. On rebound from the channel, i.e. contrend. Earns in the channel, and loses on a breakdown.

2. Breaking through the channel - trend. It makes profit within long trends, it loses losses within sideways.



The only problem here is how to determine the next market phase and how long it will last.......

12345678
New comment